Sharon Goldwater's publications

2023

ALDi: Quantifying the Arabic Level of Dialectness of Text. Amr Keleg, Sharon Goldwater and Walid Magdy. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.
[ pdf | bib | abstract ]

Transcribed speech and user-generated text in Arabic typically contain a mixture of Modern Standard Arabic (MSA), the standardized language taught in schools, and Dialectal Arabic (DA), used in daily communications. To handle this variation, previous work in Arabic NLP has focused on Dialect Identification (DI) on the sentence or the token level. However, DI treats the task as binary, whereas we argue that Arabic speakers perceive a spectrum of dialectness, which we operationalize at the sentence level as the Arabic Level of Dialectness (ALDi), a continuous linguistic variable. We introduce the AOC-ALDi dataset (derived from the AOC dataset), containing 127,835 sentences (17% from news articles and 83% from user comments on those articles) which are manually labeled with their level of dialectness. We provide a detailed analysis of AOC-ALDi and show that a model trained on it can effectively identify levels of dialectness on a range of other corpora (including dialects and genres not included in AOC-ALDi), providing a more nuanced picture than traditional DI systems. Through case studies, we illustrate how ALDi can reveal Arabic speakers’ stylistic choices in different situations, a useful property for sociolinguistic analyses.
```
@inproceedings{keleg.goldwater.ea:aldi,
  author = {Keleg, Amr and Goldwater, Sharon and Magdy, Walid},
  title = {ALDi: Quantifying the Arabic Level of Dialectness of Text},
  booktitle = {Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
  year = {2023}
}
```
Self-supervised Predictive Coding Models Encode Speaker and Phonetic Information in Orthogonal Subspaces. Oli Danyi Liu, Hao Tang and Sharon Goldwater. In Proceedings of Interspeech. 2023.
[ pdf | bib | abstract ]

Self-supervised speech representations are known to encode both speaker and phonetic information, but how they are distributed in the high-dimensional space remains largely unexplored. We hypothesize that they are encoded in orthogonal subspaces, a property that lends itself to simple disentanglement. Applying principal component analysis to representations of two predictive coding models, we identify two subspaces that capture speaker and phonetic variances, and confirm that they are nearly orthogonal. Based on this property, we propose a new speaker normalization method which collapses the subspace that encodes speaker information, without requiring transcriptions. Probing experiments show that our method effectively eliminates speaker information and outperforms a previous baseline in phone discrimination tasks. Moreover, the approach generalizes and can be used to remove information of unseen speakers.
```
@inproceedings{liu.tang.ea:self-supervised,
  title = {Self-supervised Predictive Coding Models Encode Speaker and Phonetic Information in Orthogonal Subspaces},
  author = {Liu, Oli Danyi and Tang, Hao and Goldwater, Sharon},
  booktitle = {Proceedings of Interspeech},
  year = {2023}
}
```
Infant Phonetic Learning as Perceptual Space Learning: A Crosslinguistic Evaluation of Computational Models. Yevgen Matusevych, Thomas Schatz, Herman Kamper, Naomi Feldman and Sharon Goldwater. Cognitive Science 47 (7), pp. e13314. 2023.
[ pdf | bib | abstract | online journal ]

Abstract In the first year of life, infants’ speech perception becomes attuned to the sounds of their native language. This process of early phonetic learning has traditionally been framed as phonetic category acquisition. However, recent studies have hypothesized that the attunement may instead reflect a perceptual space learning process that does not involve categories. In this article, we explore the idea of perceptual space learning by implementing five different perceptual space learning models and testing them on three phonetic contrasts that have been tested in the infant speech perception literature. We reproduce and extend previous results showing that a perceptual space learning model that uses only distributional information about the acoustics of short time slices of speech can account for at least some crosslinguistic differences in infant perception. Moreover, we find that a second perceptual space learning model, which benefits from word-level guidance. performs equally well in capturing crosslinguistic differences in infant speech perception. These results provide support for the general idea of perceptual space learning as a theory of early phonetic learning but suggest that more fine-grained data are needed to distinguish between different formal accounts. Finally, we provide testable empirical predictions of the two most promising models and show that these are not identical, making it possible to independently evaluate each model in experiments with infants in future research.
```
@article{matusevych.schatz.ea:infant,
  title = {Infant Phonetic Learning as Perceptual Space Learning: A Crosslinguistic Evaluation of Computational Models},
  author = {Matusevych, Yevgen and Schatz, Thomas and Kamper, Herman and Feldman, Naomi and Goldwater, Sharon},
  journal = {Cognitive Science},
  volume = {47},
  number = {7},
  pages = {e13314},
  year = {2023}
}
```
Prosodic features improve sentence segmentation and parsing in English. Elizabeth Nielsen, Mark Steedman and Sharon Goldwater. In Proceedings of Interspeech. 2023.
[ pdf | bib | abstract ]

Parsing spoken dialogue presents challenges that parsing text does not, including a lack of clear sentence boundaries. We know from previous work that prosody helps in parsing single sentences, but we want to show the effect of prosody on parsing speech that isn’t segmented into sentences. In experiments on the English Switchboard corpus, we find prosody helps our model both with parsing and with accurately identifying sentence boundaries. However, we find that the best-performing parser is not necessarily the parser that produces the best sentence segmentation performance. We suggest that the best parses instead come from modelling sentence boundaries jointly with other syntactic boundaries.
```
@inproceedings{nielsen.steedman.ea:prosodic,
  title = {Prosodic features improve sentence segmentation and parsing in English},
  author = {Nielsen, Elizabeth and Steedman, Mark and Goldwater, Sharon},
  booktitle = {Proceedings of Interspeech},
  year = {2023}
}
```
Acoustic Word Embeddings for Untranscribed Target Languages with Continued Pretraining and Learned Pooling. Ramon Sanabria, Hao Tang and Sharon Goldwater. In Proceedings of Interspeech. 2023.
[ pdf | bib | abstract ]

Acoustic word embeddings are typically created by training a pooling function using pairs of word-like units. For unsupervised systems, these are mined using k-nearest neighbor (KNN) search, which is slow. Recently, mean-pooled representations from a pre-trained self-supervised English model were suggested as a promising alternative, but their performance on target languages was not fully competitive. Here, we explore improvements to both approaches: we use continued pre-training to adapt the self-supervised model to the target language, and we use a multilingual phone recognizer (MPR) to mine phone n-gram pairs for training the pooling function. Evaluating on four languages, we show that both methods outperform a recent approach on word discrimination. Moreover, the MPR method is orders of magnitude faster than KNN, and is highly data efficient. We also show a small improvement from performing learned pooling on top of the continued pre-trained representations.
```
@inproceedings{sanabria.tang.ea:acoustic,
  author = {Sanabria, Ramon and Tang, Hao and Goldwater, Sharon},
  title = {Acoustic Word Embeddings for Untranscribed Target Languages with Continued Pretraining and Learned Pooling},
  booktitle = {Proceedings of Interspeech},
  year = {2023}
}
```
Analyzing Acoustic Word Embeddings from Pre-trained Self-supervised Speech Models. Ramon Sanabria, Hao Tang and Sharon Goldwater. In Proceedings of the 48th IEEE International Conference on Acoustics, Speech and Signal Processing. 2023.
[ pdf | bib | abstract ]

Given the strong results of self-supervised models on various tasks, there have been surprisingly few studies exploring self-supervised representations for acoustic word embeddings (AWE), fixed-dimensional vectors representing variable-length spoken word segments. In this work, we study several pre-trained models and pooling methods for constructing AWEs with self-supervised representations. Owing to the contextualized nature of self-supervised representations, we hypothesize that simple pooling methods, such as averaging, might already be useful for constructing AWEs. When evaluating on a standard word discrimination task, we find that HuBERT representations with mean-pooling rival the state of the art on English AWEs. More surprisingly, despite being trained only on English, HuBERT representations evaluated on Xitsonga, Mandarin, and French consistently outperform the multilingual model XLSR-53 (as well as Wav2Vec 2.0 trained on English).
```
@inproceedings{sanabria.tang.ea:analyzing,
  author = {Sanabria, Ramon and Tang, Hao and Goldwater, Sharon},
  title = {Analyzing Acoustic Word Embeddings from Pre-trained Self-supervised Speech Models},
  booktitle = {Proceedings of the 48th IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023}
}
```

2022

Regularization or lexical probability-matching? How German speakers generalize plural morphology. Kate McCurdy, Sharon Goldwater and Adam Lopez. In Proceedings of the 44th Annual Conference of the Cognitive Science Society. 2022.
[ pdf | bib | abstract ]

Artificial language learning research has shown that, under some conditions, adult speakers tend to probability-match to inconsistent variation in their input, while in others, they regularize by reducing that variation. We demonstrate that this framework can characterize speaker behavior in a natural-language morphological inflection task: the lexicon can be used to estimate variation in speaker productions. In the task of German plural inflection, we find that speakers probability- match a lexical distribution conditioned on phonology, and largely disregard an alternative possible strategy of conditional regularization based on grammatical gender.
```
@inproceedings{mccurdy.goldwater.ea:regularization,
  title = {Regularization or lexical probability-matching? How German speakers generalize plural morphology},
  author = {McCurdy, Kate and Goldwater, Sharon and Lopez, Adam},
  booktitle = {Proceedings of the 44th Annual Conference of the Cognitive Science Society},
  year = {2022}
}
```

2021

Do infants really learn phonetic categories?. Naomi H. Feldman, Sharon Goldwater, Emmanuel Dupoux and Thomas Schatz. Open Mind: Discoveries in Cognitive Science. 2021.
[ pdf | bib | abstract | online journal ]

Early changes in infants’ ability to perceive native and nonnative speech sound contrasts are typically attributed to their developing knowledge of phonetic categories. We critically examine this hypothesis and argue that there is little direct evidence of category knowledge in infancy. We then propose an alternative account in which infants’ perception changes because they are learning a perceptual space that is appropriate to represent speech, without yet caring up that space into phonetic categories. If correct this new account has substantial implications for understanding early language development.
```
@article{feldman.goldwater.ea:do,
  author = {Feldman, Naomi H. and Goldwater, Sharon and Dupoux, Emmanuel and Schatz, Thomas},
  title = {Do infants really learn phonetic categories?},
  journal = {Open Mind: Discoveries in Cognitive Science},
  pages = {113–131},
  year = {2021}
}
```
Multilingual and unsupervised subword modeling for zero-resource languages. Enno Hermann, Herman Kamper and Sharon Goldwater. Computer Speech and Language 65. 2021.
[ preprint pdf | bib | abstract | online journal ]

Unsupervised subword modeling aims to learn low-level representations of speech audio in "zero-resource" settings: that is, without using transcriptions or other resources from the target language (such as text corpora or pronunciation dictionaries). A good representation should capture phonetic content and abstract away from other types of variability, such as speaker differences and channel noise. Previous work in this area has primarily focused on learning from target language data only, and has been evaluated only intrinsically. Here we directly compare multiple methods, including some that use only target language speech data and some that use transcribed speech from other (non-target) languages, and we evaluate using two intrinsic measures as well as on a downstream unsupervised word segmentation and clustering task. We find that combining two existing target-language-only methods yields better features than either method alone. Nevertheless, even better results are obtained by extracting target language bottleneck features using a model trained on other languages. Cross-lingual training using just one other language is enough to provide this benefit, but multilingual training helps even more. In addition to these results, which hold across both intrinsic measures and the extrinsic task, we discuss the qualitative differences between the different types of learned features.
```
@article{hermann.kamper.ea:multilingual,
  author = {Hermann, Enno and Kamper, Herman and Goldwater, Sharon},
  title = {Multilingual and unsupervised subword modeling for zero-resource languages},
  journal = {Computer Speech and Language},
  year = {2021},
  month = {jan},
  volume = {65}
}
```
Improved acoustic word embeddings for zero-resource languages using multilingual transfer. Herman Kamper, Yevgen Matusevych and Sharon Goldwater. IEEE Transactions on Audio, Speech and Language Processing 29. 2021.
[ preprint pdf | bib | abstract | online journal ]

Acoustic word embeddings are fixed-dimensional representations of variable-length speech segments. In settings where unlabelled speech is the only available resource, such embeddings can be used in "zero-resource" speech search, indexing and discovery systems. Here we propose to train a single supervised embedding model on labelled data from multiple well-resourced languages and then apply it to unseen zero-resource languages. For this transfer learning approach, we consider two multilingual recurrent neural network models: a discriminative classifier trained on the joint vocabularies of all training languages, and a correspondence autoencoder trained to reconstruct word pairs. We test these using a word discrimination task on six target zero-resource languages. When trained on seven well-resourced languages, both models perform similarly and outperform unsupervised models trained on the zero-resource languages. With just a single training language, the second model works better, but performance depends more on the particular training–testing language pair.
```
@article{kamper.matusevych.ea:improved,
  author = {Kamper, Herman and Matusevych, Yevgen and Goldwater, Sharon},
  title = {Improved acoustic word embeddings for zero-resource languages using multilingual transfer},
  journal = {IEEE Transactions on Audio, Speech and Language Processing},
  volume = {29},
  year = {2021}
}
```
A phonetic model of non-native spoken word processing. Yevgen Matusevych, Herman Kamper, Thomas Schatz, Naomi Feldman and Sharon Goldwater. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics. 2021. (Received an Honourable Mention for the Best Long Paper award.)
[ pdf | bib | abstract ]

Non-native speakers show difficulties with spoken word processing. Many studies attribute these difficulties to imprecise phonological encoding of words in the lexical memory. We test an alternative hypothesis: that some of these difficulties can arise from the non-native speakers’ phonetic perception. We train a computational model of phonetic learning, which has no access to phonology, on either one or two languages. We first show that the model exhibits predictable behaviors on phone-level and word-level discrimination tasks. We then test the model on a spoken word processing task, showing that phonology may not be necessary to explain some of the word processing effects observed in non-native speakers. We run an additional analysis of the model’s lexical representation space, showing that the two training languages are not fully separated in that space, similarly to the languages of a bilingual human speaker.
```
@inproceedings{matusevych.kamper.ea:phonetic,
  title = {A phonetic model of non-native spoken word processing},
  author = {Matusevych, Yevgen and Kamper, Herman and Schatz, Thomas and Feldman, Naomi and Goldwater, Sharon},
  booktitle = {Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2021},
  annote = {Received an Honourable Mention for the Best Long Paper award.}
}
```
Black or White but never neutral: How readers perceive identity from yellow or skin-toned emoji. Alexander Robertson, Walid Magdy and Sharon Goldwater. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW2), pp. 1–23. 2021.
[ preprint pdf | bib | abstract | online journal ]

Research in sociology and linguistics shows that people use language not only to express their own identity but to understand the identity of others. Recent work established a connection between expression of identity and emoji usage on social media, through use of emoji skin tone modifiers. Motivated by that finding, this work asks if, as with language, readers are sensitive to such acts of self-expression and use them to understand the identity of authors. In behavioral experiments (n=488), where text and emoji content of social media posts were carefully controlled before being presented to participants, we find in the affirmative – emoji are a salient signal of author identity. That signal is distinct from, and complementary to, the one encoded in language. Participant groups (based on self-identified ethnicity) showed no differences in how they perceive this signal, except in the case of the default yellow emoji. While both groups associate this with a White identity, the effect was stronger in White participants. Our finding that emoji can index social variables will have experimental applications for researchers but also implications for designers: supposedly “neutral“ defaults may be more representative of some users than others.
```
@article{robertson.magdy.ea:black,
  title = {Black or White but never neutral: How readers perceive identity from yellow or skin-toned emoji},
  author = {Robertson, Alexander and Magdy, Walid and Goldwater, Sharon},
  journal = {Proceedings of the ACM on Human-Computer Interaction},
  year = {2021},
  month = {oct},
  volume = {5 (CSCW2)},
  pages = {1–23},
  publisher = {ACM}
}
```
Identity Signals in Emoji Do not Influence Perception of Factual Truth on Twitter. Alexander Robertson, Walid Magdy and Sharon Goldwater. In Proceedings of the Fourth International Workshop on Emoji Understanding and Applications in Social Media. 2021.
[ pdf | bib | abstract ]

Prior work has shown that Twitter users use skin-toned emoji as an act of self-representation to express their racial/ethnic identity. We test whether this signal of identity can influence readers’ perceptions about the content of a post containing that signal. In a large scale (n=944) pre-registered controlled experiment, we manipulate the presence of skin-toned emoji and profile photos in a task where readers rate obscure trivia facts (presented as tweets) as true or false. Using a Bayesian statistical analysis, we find that neither emoji nor profile photo has an effect on how readers rate these facts. This result will be of some comfort to anyone concerned about the manipulation of online users through the crafting of fake profiles.
```
@inproceedings{robertson.magdy.ea:identity,
  title = {Identity Signals in Emoji Do not Influence Perception of Factual Truth on Twitter},
  author = {Robertson, Alexander and Magdy, Walid and Goldwater, Sharon},
  booktitle = {Proceedings of the Fourth International Workshop on Emoji Understanding and Applications in Social Media},
  year = {2021}
}
```
On the Difficulty of Segmenting Words with Attention. Ramon Sanabria, Hao Tang and Sharon Goldwater. In Proceedings of the Workshop on Insights from Negative Results in NLP. 2021.
[ pdf | bib | abstract ]

Word segmentation, the problem of finding word boundaries in speech, is of interest for a range of tasks. Previous papers have suggested that for sequence-to-sequence models trained on tasks such as speech translation or speech recognition, attention can be used to locate and segment the words. We show, however, that even on monolingual data this approach is brittle. In our experiments with different input types, data sizes, and segmentation algorithms, only models trained to predict phones from words succeed in the task. Models trained to predict words from either phones or speech (i.e., the opposite direction needed to generalize to new data), yield much worse results, suggesting that attention-based segmentation is only useful in limited scenarios.
```
@inproceedings{sanabria.tang.ea:on,
  author = {Sanabria, Ramon and Tang, Hao and Goldwater, Sharon},
  title = {On the Difficulty of Segmenting Words with Attention},
  booktitle = {Proceedings of the Workshop on Insights from Negative Results in NLP},
  year = {2021}
}
```
Early phonetic learning without phonetic categories: Insights from large-scale simulations on realistic input. Thomas Schatz, Naomi H Feldman, Sharon Goldwater, Xuan-Nga Cao and Emmanuel Dupoux. Proceedings of the National Academy of Sciences 118 (7). 2021.
[ pdf | bib | abstract | online journal ]

Before they even speak, infants become attuned to the sounds of the language(s) they hear, processing native phonetic contrasts more easily than non-native ones [1-3]. For example, between 6-8 months and 10-12 months, infants learning American English get better at distinguishing English [\textipa\*r] and [l], as in ‘rock’ vs ‘lock’, relative to infants learning Japanese [4]. Influential accounts of this \emphearly phonetic learning phenomenon initially proposed that infants group sounds into native vowel- and consonant-like phonetic categories—like [\textipa\*r] and [l] in English—through a statistical clustering mechanism dubbed ‘distributional learning’ [5-8]. The feasibility of this mechanism for learning phonetic categories has been challenged, however [9-16]. Here we demonstrate that a distributional learning algorithm operating on naturalistic speech can predict early phonetic learning as observed in Japanese and American English infants, suggesting that infants might learn through distributional learning after all. We further show, however, that contrary to the original distributional learning proposal, our model learns units too brief and too fine-grained acoustically to correspond to phonetic categories. This challenges the influential idea that \emphwhat infants learn are phonetic categories. More broadly, our work introduces a novel \emphmechanism-driven approach to the study of early phonetic learning, together with a quantitative modeling framework that can handle realistic input. This allows, for the first time, accounts of early phonetic learning to be linked to concrete, systematic predictions regarding infants’ attunement.
```
@article{schatz.feldman.ea:early,
  title = {Early phonetic learning without phonetic categories: Insights from large-scale simulations on realistic input},
  author = {Schatz, Thomas and Feldman, Naomi H and Goldwater, Sharon and Cao, Xuan-Nga and Dupoux, Emmanuel},
  journal = {Proceedings of the National Academy of Sciences},
  volume = {118},
  number = {7},
  year = {2021},
  publisher = {National Acad Sciences}
}
```

2020

Cross-lingual topic prediction for speech using translations. Sameer Bansal, Herman Kamper, Adam Lopez and Sharon Goldwater. In Proceedings of the 45th IEEE International Conference on Acoustics, Speech and Signal Processing. 2020.
[ pdf | bib | abstract ]

Given a large amount of unannotated speech in a low-resource language, can we classify the speech utterances by topic? We consider this question in the setting where a small amount of speech in the low-resource language is paired with text translations in a high-resource language. We develop an effective cross-lingual topic classifier by training on just 20 hours of translated speech, using a recent model for direct speech-to-text translation. While the translations are poor, they are still good enough to correctly classify the topic of 1-minute speech segments over 70% of the time – a 20% improvement over a majority-class baseline. Such a system could be useful for humanitarian applications like crisis response, where incoming speech in a foreign low-resource language must be quickly assessed for further action.
```
@inproceedings{bansal.kamper.ea:cross-lingual,
  author = {Bansal, Sameer and Kamper, Herman and Lopez, Adam and Goldwater, Sharon},
  title = {Cross-lingual topic prediction for speech using translations},
  booktitle = {Proceedings of the 45th IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020}
}
```
Multilingual acoustic word embedding models for processing zero-resource languages. Herman Kamper, Yevgen Matusevych and Sharon Goldwater. In Proceedings of the 45th IEEE International Conference on Acoustics, Speech and Signal Processing. 2020.
[ pdf | bib | abstract ]

Acoustic word embeddings are fixed-dimensional representations of variable-length speech segments. In settings where unlabelled speech is the only available resource, such embeddings can be used in "zero-resource" speech search, indexing and discovery systems. Here we propose to train a single supervised embedding model on labelled data from multiple well-resourced languages and then apply it to unseen zero-resource languages. For this transfer learning approach, we consider two multilingual recurrent neural network models: a discriminative classifier trained on the joint vocabularies of all training languages, and a correspondence autoencoder trained to reconstruct word pairs. We test these using a word discrimination task on six target zero-resource languages. When trained on seven well-resourced languages, both models perform similarly and outperform unsupervised models trained on the zero-resource languages. With just a single training language, the second model works better, but performance depends more on the particular training–testing language pair.
```
@inproceedings{kamper.matusevych.ea:multilingual,
  author = {Kamper, Herman and Matusevych, Yevgen and Goldwater, Sharon},
  title = {Multilingual acoustic word embedding models for processing zero-resource languages},
  booktitle = {Proceedings of the 45th IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020}
}
```
Input matters in the modeling of early phonetic learning. Ruolan Li, Thomas Schatz, Yevgen Matusevych, Sharon Goldwater and Naomi Feldman. In Proceedings of the 42nd Annual Conference of the Cognitive Science Society. 2020.
[ pdf | bib | abstract ]

In acquiring language, differences in input can greatly affect learning outcomes, but which aspects of language learning are most sensitive to input variations, and which are robust, remains debated. A recent modeling study successfully reproduced a phenomenon empirically observed in early phonetic learning—learning about the sounds of the native language in the first year of life—despite using input that differed in quantity and speaker composition from what a typical infant would hear. In this paper, we carry out a direct test of that model’s robustness to input variations. We find that, despite what the original result suggested, the learning outcomes are sensitive to properties of the input and that more plausible input leads to a better fit with empirical observations. This has implications for understanding early phonetic learning in infants and underscores the importance of using realistic input in models of language acquisition.
```
@inproceedings{li.schatz.ea:input,
  title = {Input matters in the modeling of early phonetic learning},
  author = {Li, Ruolan and Schatz, Thomas and Matusevych, Yevgen and Goldwater, Sharon and Feldman, Naomi},
  booktitle = {Proceedings of the 42nd Annual Conference of the Cognitive Science Society},
  year = {2020}
}
```
Analyzing autoencoder-based acoustic word embeddings. Yevgen Matusevych, Herman Kamper and Sharon Goldwater. In Workshop on Bridging AI and Cognitive Science at ICLR. 2020.
[ pdf | bib | abstract ]

Recent studies have introduced methods for learning acoustic word embeddings (AWEs)—fixed-size vector representations of words which encode their acoustic features. Despite the widespread use of AWEs in speech processing research, they have only been evaluated quantitatively in their ability to discriminate between whole word tokens. To better understand the applications of AWEs in various downstream tasks and in cognitive modeling, we need to analyze the representation spaces of AWEs. Here we analyze basic properties of AWE spaces learned by a sequence-to-sequence encoder-decoder model in six typologically diverse languages. We first show that these AWEs preserve some information about words’ absolute duration and speaker. At the same time, the representation space of these AWEs is organized such that the distance between words’ embeddings increases with those words’ phonetic dissimilarity. Finally, the AWEs exhibit a word onset bias, similar to patterns reported in various studies on human speech processing and lexical access. We argue this is a promising result and encourage further evaluation of AWEs as a potentially useful tool in cognitive science, which could provide a link between speech processing and lexical memory.
```
@inproceedings{matusevych.kamper.ea:analyzing,
  author = {Matusevych, Yevgen and Kamper, Herman and Goldwater, Sharon},
  title = {Analyzing autoencoder-based acoustic word embeddings},
  booktitle = {Workshop on Bridging AI and Cognitive Science at ICLR},
  year = {2020}
}
```
Evaluating computational models of infant phonetic learning across languages. Yevgen Matusevych, Thomas Schatz, Herman Kamper, Naomi Feldman and Sharon Goldwater. In Proceedings of the 42nd Annual Conference of the Cognitive Science Society. 2020.
[ pdf | bib | abstract ]

In the first year of life, infants’ speech perception becomes attuned to the sounds of their native language. Many accounts of this early phonetic learning exist, but computational models predicting the attunement patterns observed in infants from the speech input they hear have been lacking. A recent study presented the first such model, drawing on algorithms proposed for unsupervised learning from naturalistic speech, and tested it on a single phone contrast. Here we study five such algorithms, selected for their potential cognitive relevance. We simulate phonetic learning with each algorithm and perform tests on three phone contrasts from different languages, comparing the results to infants’ discrimination patterns. The five models display varying degrees of agreement with empirical observations, showing that our approach can help decide between candidate mechanisms for early phonetic learning, and providing insight into which aspects of the models are critical for capturing infants’ perceptual development.
```
@inproceedings{matusevych.schatz.ea:evaluating,
  title = {Evaluating computational models of infant phonetic learning across languages},
  author = {Matusevych, Yevgen and Schatz, Thomas and Kamper, Herman and Feldman, Naomi and Goldwater, Sharon},
  booktitle = {Proceedings of the 42nd Annual Conference of the Cognitive Science Society},
  year = {2020}
}
```
Inflecting when there’s no majority: Limitations of encoder-decoder neural networks as cognitive models for German plurals. Kate McCurdy, Sharon Goldwater and Adam Lopez. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.
[ pdf | bib | abstract ]

Can artificial neural networks learn to represent inflectional morphology and generalize to new words as human speakers do? Kirov and Cotterell (2018) argue that the answer is yes: modern Encoder-Decoder (ED) architectures learn human-like behavior when inflecting English verbs, such as extending the regular past tense form /-(e)d/ to novel words. However, their work does not address the criticism raised by Marcus et al. (1995): that neural models may learn to extend not the \em regular, but the \em most frequent class — and thus fail on tasks like German number inflection, where infrequent suffixes like \s can still be productively generalized. To investigate this question, we first collect a new dataset from German speakers (production and ratings of plural forms for novel nouns) that is designed to avoid sources of information unavailable to the ED model. The speaker data show high variability, and two suffixes evince ‘regular’ behavior, appearing more often with phonologically atypical inputs. Encoder-decoder models do generalize the most frequently produced plural class, but do not show human-like variability or ‘regular’ extension of these other plural markers. We conclude that modern neural models may still struggle with minority-class generalization.
```
@inproceedings{mccurdy.goldwater.ea:inflecting,
  title = {Inflecting when there’s no majority: Limitations of encoder-decoder neural networks as cognitive models for German plurals},
  author = {McCurdy, Kate and Goldwater, Sharon and Lopez, Adam},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year = {2020}
}
```
Conditioning, but on Which Distribution? Grammatical Gender in German Plural Inflection. Kate McCurdy, Adam Lopez and Sharon Goldwater. In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, pp. 59–65. 2020.
[ pdf | bib | abstract ]

Grammatical gender is a consistent and informative cue to the plural class of German nouns. We find that neural encoder-decoder models learn to rely on this cue to predict plural class, but adult speakers are relatively insensitive to it. This suggests that the neural models are not an effective cognitive model of German plural formation.
```
@inproceedings{mccurdy.lopez.ea:conditioning,
  title = {Conditioning, but on Which Distribution? Grammatical Gender in German Plural Inflection},
  author = {McCurdy, Kate and Lopez, Adam and Goldwater, Sharon},
  booktitle = {Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics},
  year = {2020},
  publisher = {Association for Computational Linguistics},
  pages = {59–65}
}
```
The role of context in neural pitch accent detection in English. Elizabeth Nielsen, Mark Steedman and Sharon Goldwater. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pp. 7994–8000. 2020.
[ pdf | bib | abstract ]

Prosody is a rich information source in natural language, serving as a marker for phenomena such as contrast. In order to make this information available to downstream tasks, we need a way to detect prosodic events in speech. We propose a new model for pitch accent detection, inspired by the work of Stehwien et al. (2018), who presented a CNN-based model for this task. Our model makes greater use of context by using full utterances as input and adding an LSTM layer. We find that these innovations lead to an improvement from 87.5% to 88.7% accuracy on pitch accent detection on American English speech in the Boston University Radio News Corpus, a state-of-the-art result. We also find that a simple baseline that just predicts a pitch accent on every content word yields 82.2% accuracy, and we suggest that this is the appropriate baseline for this task. Finally, we conduct ablation tests that show pitch is the most important acoustic feature for this task and this corpus.
```
@inproceedings{nielsen.steedman.ea:role,
  title = {The role of context in neural pitch accent detection in English},
  author = {Nielsen, Elizabeth and Steedman, Mark and Goldwater, Sharon},
  booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  publisher = {Association for Computational Linguistics},
  pages = {7994–8000}
}
```
Emoji Skin Tone Modifiers: Analyzing Variation in Usage on Social Media. Alexander Robertson, Walid Magdy and Sharon Goldwater. ACM Transactions on Social Computing 3 (2), pp. 1–25. 2020.
[ pdf | bib | abstract | online journal ]

Emoji are widely used in computer-mediated communication to express concepts and emotions. Skin tone modifiers were added in 2015 with the hope of better representing user diversity, and, indeed, recent work has shown that these modifiers are especially popular among darker-skinned users, who are a minority on Twitter. Previous work also showed that the vast majority of tone-modified emoji have a tone similar to the user’s own skin tone, suggesting that self-representation is a major factor in tone use. In this article, we first show that the basic finding (users mainly choose a tone that is similar to their own skin tone) generalizes to different sub-populations of users, including users from majority-Black regions. We then extend the analysis of tone use to quantify and examine cases where users modulate their tones—that is, for a particular emoji, they choose either to use a different tone than their usual one or no tone at all (after having previously used one). We show that even though these uses constitute only a small proportion of emoji usage, many instances are readily classifiable as ways of representing other people. The evidence we present is therefore crucial in working toward a broader understanding of the connection between emoji and identity expression online. We also offer explanations for why the darkest emoji skin tones are not used, by examining aspects of their design, which make them less suited to self-representational usage. This highlights the need for careful consideration of both design and human diversity when creating emoji. Moreover, despite early fears in the media, we find little evidence of negative usage, even when tones are used in a non-self-representational manner. In sum, our findings lend even more support to the highly positive role that emoji and skin tone modifiers play in identity and expression in computer-mediated communication.
```
@article{robertson.magdy.ea:emoji,
  title = {Emoji Skin Tone Modifiers: Analyzing Variation in Usage on Social Media},
  author = {Robertson, Alexander and Magdy, Walid and Goldwater, Sharon},
  journal = {ACM Transactions on Social Computing},
  volume = {3},
  number = {2},
  pages = {1–25},
  year = {2020},
  publisher = {ACM}
}
```
Analyzing ASR pretraining for low-resource speech-to-text translation. Mihaela C. Stoain, Sameer Bansal and Sharon Goldwater. In Proceedings of the 45th IEEE International Conference on Acoustics, Speech and Signal Processing. 2020.
[ pdf | bib | abstract ]

Previous work has shown that for low-resource source languages, automatic speech-to-text translation (AST) can be improved by pretraining an end-to-end model on automatic speech recognition (ASR) data from a high-resource language. However, it is not clear what factors –e.g., language relatedness or size of the pretraining data– yield the biggest improvements, or whether pretraining can be effectively combined with other methods such as data augmentation. Here, we experiment with pretraining on datasets of varying sizes, including languages related and unrelated to the AST source language. We find that the best predictor of final AST performance is the word error rate of the pretrained ASR model, and that differences in ASR/AST performance correlate with how phonetic information is encoded in the later RNN layers of our model. We also show that pretraining and data augmentation yield complementary benefits for AST.
```
@inproceedings{stoain.bansal.ea:analyzing,
  author = {Stoain, Mihaela C. and Bansal, Sameer and Goldwater, Sharon},
  title = {Analyzing ASR pretraining for low-resource speech-to-text translation},
  booktitle = {Proceedings of the 45th IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020}
}
```

2019

Pre-training on high-resource speech recognition improves low-resource speech-to-text translation. Sameer Bansal, Herman Kamper, Karen Livescu, Adam Lopez and Sharon Goldwater. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2019.
[ pdf | bib | abstract ]

We present a simple approach to improve direct speech-to-text translation (ST) when the source language is low-resource: we pre-train the model on a high-resource automatic speech recognition (ASR) task, and then fine-tune its parameters for ST. We demonstrate that our approach is effective by pre-training on 300 hours of English ASR data to improve Spanish-English ST from 10.8 to 20.2 BLEU when only 20 hours of Spanish-English ST training data are available. Through an ablation study, we find that the pre-trained encoder (acoustic model) accounts for most of the improvement, despite the fact that the shared language in these tasks is the target language text, not the source language audio. Applying this insight, we show that pre-training on ASR helps ST even when the ASR language differs from both source and target ST languages: pre-training on French ASR also improves Spanish-English ST. Finally, we show that the approach improves performance on a true low-resource task: pre-training on a combination of English ASR and French ASR improves Mboshi-French ST, where only 4 hours of data are available, from 3.5 to 7.1 BLEU.
```
@inproceedings{bansal.kamper.ea:pre-training,
  author = {Bansal, Sameer and Kamper, Herman and Livescu, Karen and Lopez, Adam and Goldwater, Sharon},
  title = {Pre-training on high-resource speech recognition improves low-resource speech-to-text translation},
  booktitle = {Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2019}
}
```
Data Augmentation for Context-Sensitive Neural Lemmatization Using Inflection Tables and Raw Text. Toms Bergmanis and Sharon Goldwater. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2019. (This is an updated version that fixes a cross-reference and a consequential typo: we used macro-averaged type accuracy, not micro-averaged as stated in the official published version.)
[ pdf | bib | abstract ]

Lemmatization aims to reduce the sparse data problem by relating the inflected forms of a word to its dictionary form. Using context can help, both for unseen and ambiguous words. Yet most context-sensitive approaches require full lemma-annotated sentences for training, which may be scarce or unavailable in low-resource languages. In addition (as shown here), in a low-resource setting, a lemmatizer can learn more from n labeled examples of distinct words (types) than from n (contiguous) labeled tokens, since the latter contain far fewer distinct types. To combine the efficiency of type-based learning with the benefits of context, we propose a way to train a context-sensitive lemmatizer with little or no labeled corpus data, using inflection tables from the UniMorph project and raw text examples from Wikipedia that provide sentence contexts for the unambiguous UniMorph examples. Despite these being unambiguous examples, the model successfully generalizes from them, leading to improved results (both overall, and especially on unseen words) in comparison to a baseline that does not use context.
```
@inproceedings{bergmanis.goldwater:data,
  author = {Bergmanis, Toms and Goldwater, Sharon},
  title = {Data Augmentation for Context-Sensitive Neural Lemmatization Using Inflection Tables and Raw Text},
  annote = {This is an updated version that fixes a cross-reference and a consequential typo: we used macro-averaged type accuracy, not micro-averaged as stated in the official published version.},
  booktitle = {Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2019}
}
```
Are we there yet? Encoder-decoder neural networks as cognitive models of English past tense inflection. Maria Corkery, Yevgen Matusevych and Sharon Goldwater. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.
[ pdf | bib | abstract ]

The cognitive mechanisms needed to account for the English past tense have long been a subject of debate in linguistics and cognitive science. Neural network models were proposed early on, but were shown to have clear flaws. Recently, however, Kirov and Cotterell (2018) showed that modern encoder-decoder (ED) models overcome many of these flaws. They also presented evidence that ED models demonstrate humanlike performance in a nonce-word task. Here, we look more closely at the behaviour of their model in this task. We find that (1) the model exhibits instability across multiple simulations in terms of its correlation with human data, and (2) even when results are aggregated across simulations (treating each simulation as an individual human participant), the fit to the human data is not strong—worse than an older rule-based model. These findings hold up through several alternative training regimes and evaluation measures. Although other neural architectures might do better, we conclude that there is still insufficient evidence to claim that neural nets are a good cognitive model for this task.
```
@inproceedings{corkery.matusevych.ea:are,
  author = {Corkery, Maria and Matusevych, Yevgen and Goldwater, Sharon},
  title = {Are we there yet? Encoder-decoder neural networks as cognitive models of English past tense inflection},
  booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
  year = {2019}
}
```

2018

Low-Resource Speech-to-Text Translation. Sameer Bansal, Herman Kamper, Karen Livescu, Adam Lopez and Sharon Goldwater. In Proceedings of Interspeech. 2018.
[ pdf | bib | abstract ]

Speech-to-text translation has many potential applications for low-resource languages, but the typical approach of cascading speech recognition with machine translation is often impossible, since the transcripts needed to train a speech recognizer are usu- ally not available for low-resource languages. Recent work has found that neural encoder-decoder models can learn to directly translate foreign speech in high-resource scenarios, without the need for intermediate transcription. We investigate whether this approach also works in settings where both data and computation are limited. To make the approach efficient, we make several architectural changes, including a change from character-level to word-level decoding. We find that this choice yields crucial speed improvements that allow us to train with fewer computa- tional resources, yet still performs well on frequent words. We explore models trained on between 20 and 160 hours of data, and find that although models trained on less data have considerably lower BLEU scores, they can still predict words with relatively high precision and recall—around 50% for a model trained on 50 hours of data, versus around 60% for the full 160 hour model. Thus, they may still be useful for some low-resource scenarios.
```
@inproceedings{bansal.kamper.ea:low-resource,
  author = {Bansal, Sameer and Kamper, Herman and Livescu, Karen and Lopez, Adam and Goldwater, Sharon},
  title = {Low-Resource Speech-to-Text Translation},
  booktitle = {Proceedings of Interspeech},
  year = {2018}
}
```
Context Sensitive Neural Lemmatization with Lematus. Toms Bergmanis and Sharon Goldwater. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2018.
[ pdf | bib | abstract ]

The main motivation for developing context-sensitive lemmatizers is to improve performance on unseen and ambiguous words. Yet previous systems have not carefully evaluated whether the use of context actually helps in these cases. We introduce Lematus, a lemmatizer based on a standard encoder-decoder architecture, which incorporates character-level sentence context. We evaluate its performance across 20 languages in both a full data setting and a lower-resource setting with 10k training examples in each language. We show that in the full data setting only, the use of context significantly improves performance against a context-free version of our model. We also compare to three previous context-sensitive lemmatization systems, which all use pre-extracted edit trees as well as hand-selected features and/or additional sources of information such as tagged training data. Without using any of these, our context-sensitive model shows a statistically significant improvement over the best competitor system (Lemming) on unseen forms and performs on par on ambiguous words.
```
@inproceedings{bergmanis.goldwater:context,
  author = {Bergmanis, Toms and Goldwater, Sharon},
  title = {Context Sensitive Neural Lemmatization with Lematus},
  booktitle = {Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2018}
}
```
Multilingual bottleneck features for subword modeling in zero-resource languages. Enno Hermann and Sharon Goldwater. In Proceedings of Interspeech. 2018.
[ pdf | bib | abstract ]

How can we effectively develop speech technology for lan- guages where no transcribed data is available? Many existing approaches use no annotated resources at all, yet it makes sense to leverage information from large annotated corpora in other languages, for example in the form of multilingual bottleneck features (BNFs) obtained from a supervised speech recognition system. In this work, we evaluate the benefits of BNFs for sub- word modeling (feature extraction) in six unseen languages on a word discrimination task. First we establish a strong unsu- pervised baseline by combining two existing methods: vocal tract length normalisation (VTLN) and the correspondence au- toencoder (cAE). We then show that BNFs trained on a single language already beat this baseline; including up to 10 languages results in additional improvements which cannot be matched by just adding more data from a single language. Finally, we show that the cAE can improve further on the BNFs if high-quality same-word pairs are available.
```
@inproceedings{hermann.goldwater:multilingual,
  author = {Hermann, Enno and Goldwater, Sharon},
  title = {Multilingual bottleneck features for subword modeling in zero-resource languages},
  booktitle = {Proceedings of Interspeech},
  year = {2018}
}
```
Evaluating historical text normalization systems: How well do they generalize?. Alexander Robertson and Sharon Goldwater. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2018.
[ pdf | bib | abstract ]

We highlight several issues in the evaluation of historical text normalization systems that make it hard to tell how well these systems would actually work in practice—i.e., for new datasets or languages; in comparison to more naive systems; or as a preprocessing step for downstream NLP tools. We illustrate these issues and exemplify our proposed evaluation practices by comparing two neural models against a naive baseline system. We show that the neural models generalize well to unseen words in tests on five languages; nevertheless, they provide no clear benefit over the naive baseline for downstream POS tagging of an English historical collection. We conclude that future work should include more rigorous evaluation, including both intrinsic and extrinsic measures where possible.
```
@inproceedings{robertson.goldwater:evaluating,
  author = {Robertson, Alexander and Goldwater, Sharon},
  title = {Evaluating historical text normalization systems: How well do they generalize?},
  booktitle = {Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2018}
}
```
Self-Representation on Twitter Using Emoji Skin Color Modifiers. Alexander Robertson, Walid Magdy and Sharon Goldwater. In Proceedings of the 12th International AAAI Conference on Web and Social Media (ICWSM). 2018.
[ pdf | bib | abstract ]

Since 2015, it has been possible to modify certain emoji with a skin tone. The five different skin tones were introduced with the aim of representing more human diversity, but some commentators feared they might be used as a way to negatively represent other users/groups. This paper presents a quantitative analysis of the use of skin tone modifiers on emoji on Twitter, showing that users with darker-skinned profile photos employ them more often than users with lighter-skinned profile photos, and the vast majority of skin tone usage matches the color of a user’s profile photo–i.e., tones represent the self, rather than the other. In the few cases where users do use opposite-toned emoji, we find no evidence of negative racial sentiment. Thus, the introduction of skin tones seems to have met the goal of better representing human diversity.
```
@inproceedings{robertson.magdy.ea:self-representation,
  author = {Robertson, Alexander and Magdy, Walid and Goldwater, Sharon},
  title = {Self-Representation on Twitter Using Emoji Skin Color Modifiers},
  booktitle = {Proceedings of the 12th International AAAI Conference on Web and Social Media (ICWSM)},
  year = {2018}
}
```
Inducing a lexicon of sociolinguistic variables from code-mixed text. Philippa Shoemark, James Kirby and Sharon Goldwater. In Workshop on Noisy User-Generated Text at EMNLP. 2018. (Received Best Paper award. Includes Supplementary Information.)
[ pdf | bib | abstract ]

Sociolinguistics is often concerned with how variants of a linguistic item (e.g., nothing vs. nothin’) are used by different groups or in different situations. We introduce the task of inducing lexical variables from code-mixed text: that is, identifying equivalence pairs such as (football, fitba) along with their linguis- tic code (football→British, fitba→Scottish). We adapt a framework for identifying gender- biased word pairs to this new task, and present results on three different pairs of English di- alects, using tweets as the code-mixed text. Our system achieves precision of over 70% for two of these three datasets, and produces useful results even without extensive parame- ter tuning. Our success in adapting this frame- work from gender to language variety suggests that it could be used to discover other types of analogous pairs as well.
```
@inproceedings{shoemark.kirby.ea:inducing,
  author = {Shoemark, Philippa and Kirby, James and Goldwater, Sharon},
  title = {Inducing a lexicon of sociolinguistic variables from code-mixed text},
  booktitle = {Workshop on Noisy User-Generated Text at EMNLP},
  year = {2018},
  address = {Brussels, Belgium},
  annote = {Received Best Paper award. Includes Supplementary Information.}
}
```

2017

Bootstrapping Language Acquisition. Omri Abend, Tom Kwiatkowski, Nathaniel J. Smith, Sharon Goldwater and Mark Steedman. Cognition 164, pp. 116–143. 2017.
[ preprint pdf | bib | abstract | online journal ]

The semantic bootstrapping hypothesis proposes that children acquire their native language through exposure to sentences of the language paired with structured representations of their meaning, whose component substructures can be associated with words and syntactic structures used to express these concepts. The child’s task is then to learn a language-specific grammar and lexicon based on (probably contextually ambiguous, possibly somewhat noisy) pairs of sentences and their meaning representations (logical forms). Starting from these assumptions, we develop a Bayesian probabilistic account of semantically bootstrapped first-language acquisition in the child, based on techniques from computational parsing and interpretation of unrestricted text. Our learner jointly models (a) word learning: the mapping between components of the given sentential meaning and lexical words (or phrases) of the language, and (b) syntax learning: the projection of lexical elements onto sentences by universal construction-free syntactic rules. Using an incremental learning algorithm, we apply the model to a dataset of real syntactically complex child-directed utterances and (pseudo) logical forms, the latter including contextually plausible but irrelevant distractors. Taking the Eve section of the CHILDES corpus as input, the model simulates several well-documented phenomena from the developmental literature. In particular, the model exhibits syntactic bootstrapping effects (in which previously learned constructions facilitate the learning of novel words), sudden jumps in learning without explicit parameter setting, acceleration of word-learning (the “vocabulary spurt”), an initial bias favoring the learning of nouns over verbs, and one-shot learning of words and their meanings. The learner thus demonstrates how statistical learning over structured representations can provide a unified account for these seemingly disparate phenomena.
```
@article{abend.kwiatkowski.ea:bootstrapping,
  author = {Abend, Omri and Kwiatkowski, Tom and Smith, Nathaniel J. and Goldwater, Sharon and Steedman, Mark},
  title = {Bootstrapping Language Acquisition},
  journal = {Cognition},
  volume = {164},
  month = {jul},
  pages = {116–143},
  year = {2017}
}
```
Spoken Term Discovery for Language Documentation using Translations. Antonios Anastasopoulos, Sameer Bansal, Sharon Goldwater, Adam Lopez and David Chiang. In Workshop on Speech-Centric Natural Language Processing at EMNLP. 2017.
[ pdf | bib | abstract ]

Vast amounts of speech data collected for language documentation and research remain untranscribed and unsearchable, but often a small amount of speech may have text translations available. We present a method for partially labeling additional speech with translations in this scenario. We modify an unsupervised speech-to-translation alignment model and obtain prototype speech segments that match the translation words, which are in turn used to discover terms in the unlabelled data. We evaluate our method on a Spanish-English speech translation corpus and on two corpora of endangered languages, Arapaho and Ainu, demonstrating its appropriateness and applicability in an actual very-low-resource scenario.
```
@inproceedings{anastasopoulos.bansal.ea:spoken,
  author = {Anastasopoulos, Antonios and Bansal, Sameer and Goldwater, Sharon and Lopez, Adam and Chiang, David},
  title = {Spoken Term Discovery for Language Documentation using Translations},
  booktitle = {Workshop on Speech-Centric Natural Language Processing at EMNLP},
  year = {2017}
}
```
Towards speech-to-text translation without speech recognition. Sameer Bansal, Herman Kamper, Adam Lopez and Sharon Goldwater. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics. 2017.
[ pdf | bib | abstract ]

We explore the problem of translating speech to text in low-resource scenarios where neither automatic speech recognition (ASR) nor machine translation (MT) are available, but we have training data in the form of audio paired with text translations. We present the first system for this problem applied to a realistic multi-speaker dataset, the CALLHOME Spanish-English speech translation corpus. Our approach uses un- supervised term discovery (UTD) to cluster repeated patterns in the audio, creating a pseudotext, which we pair with translations to create a parallel text and train a simple bag-of-words MT model. We identify the challenges faced by the system, finding that the difficulty of cross-speaker UTD results in low recall, but that our system is still able to correctly translate some content words in test data.
```
@inproceedings{bansal.kamper.ea:towards,
  author = {Bansal, Sameer and Kamper, Herman and Lopez, Adam and Goldwater, Sharon},
  title = {Towards speech-to-text translation without speech recognition},
  booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2017}
}
```
Weakly supervised spoken term discovery using cross-lingual side information. Sameer Bansal, Herman Kamper, Sharon Goldwater and Adam Lopez. In Proceedings of the 42nd IEEE International Conference on Acoustics, Speech and Signal Processing. 2017. (Copyright 2017 IEEE.)
[ pdf | bib | abstract ]

Recent work on unsupervised term discovery (UTD) aims to iden- tify and cluster repeated word-like units from audio alone. These systems are promising for some very low-resource languages where transcribed audio is unavailable, or where no written form of the language exists. However, in some cases it may still be feasible (e.g., through crowdsourcing) to obtain (possibly noisy) text translations of the audio. If so, this information could be used as a source of side information to improve UTD. Here, we present a simple method for rescoring the output of a UTD system using text translations, and test it on a corpus of Spanish audio with English translations. We show that it greatly improves the average precision of the results over a wide range of system configurations and data preprocessing methods.
```
@inproceedings{bansal.kamper.ea:weakly,
  author = {Bansal, Sameer and Kamper, Herman and Goldwater, Sharon and Lopez, Adam},
  title = {Weakly supervised spoken term discovery using cross-lingual side information},
  booktitle = {Proceedings of the 42nd IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2017},
  annote = {Copyright 2017 IEEE.}
}
```
From segmentation to analyses: a probabilistic model for unsupervised morphology induction. Toms Bergmanis and Sharon Goldwater. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics. 2017.
[ pdf | bib | abstract ]

A major motivation for unsupervised morphological analysis is to reduce the sparse data problem in under-resourced languages. Most previous work focuses on segmenting surface forms into their constituent morphs (e.g., taking: tak +ing), but surface form segmentation does not solve the sparse data problem as the analyses of ’take’ and ’taking’ are not connected to each other. We extend the MorphoChains system (Narasimhan et al., 2015) to provide morphological analyses that can abstract over spelling differences in functionally similar morphs. These analyses are not required to use all the orthographic material of a word (stopping: stop +ing), nor are they limited to only that material (acidified: acid +ify +ed). On average across six typologically varied languages our system has a similar or better F-score on EMMA (a measure of underlying morpheme accuracy) than three strong baselines; moreover, the total number of distinct morphemes identified by our system is on average 12.8% lower than for Morfessor (Virpioja et al., 2013), a state-of-the-art surface segmentation system.
```
@inproceedings{bergmanis.goldwater:from,
  author = {Bergmanis, Toms and Goldwater, Sharon},
  title = {From segmentation to analyses: a probabilistic model for unsupervised morphology induction},
  booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2017},
  address = {Valencia, Spain}
}
```
Training Data Augmentation for Low-Resource Morphological Inflection. Toms Bergmanis, Katharina Kann, Hinrich Schütze and Sharon Goldwater. In Proceedings of the CoNLL-SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection. 2017.
[ pdf | bib | abstract ]

This work describes the UoE-LMU submission for the CoNLL-SIGMORPHON 2017 Shared Task on Universal Morphological Reinflection, Subtask 1: given a lemma and target morphological tags, generate the target inflected form. We evaluate several ways to improve performance in the 1000-example setting: three methods to augment the training data with identical input-output pairs (i.e., autoencoding), a heuristic approach to identify likely pairs of inflectional variants from an unlabeled corpus, and a method for cross-lingual knowledge transfer. We find that autoencoding random strings works surprisingly well, outperformed only slightly by autoencoding words from an unlabelled corpus. The random string method also works well in the 10,000-example setting despite not being tuned for it. Among 18 submissions our system takes 1st and 6th place in the 10k and 1k settings, respectively.
```
@inproceedings{bergmanis.kann.ea:training,
  author = {Bergmanis, Toms and Kann, Katharina and Schütze, Hinrich and Goldwater, Sharon},
  title = {Training Data Augmentation for Low-Resource Morphological Inflection},
  booktitle = {Proceedings of the CoNLL-SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection},
  year = {2017},
  address = {Vancouver, Canada}
}
```
A segmental framework for fully-unsupervised large-vocabulary speech recognition. Herman Kamper, Aren Jansen and Sharon Goldwater. Computer Speech and Language 46, pp. 154–174. 2017. (Received 2021 Best Research Paper award (for a paper published in the journal during the previous five years).)
[ preprint pdf | bib | abstract | online journal ]

Zero-resource speech technology is a growing research area that aims to develop methods for speech processing in the absence of transcriptions, lexicons, or language modelling text. Early systems focused on identifying isolated recurring terms in a corpus, while more recent full-coverage systems attempt to completely segment and cluster the audio into word-like units—effectively performing unsupervised speech recognition. To our knowledge, this article presents the first such system evaluated on large-vocabulary multi-speaker data. The system uses a Bayesian modelling framework with segmental word representations: each word segment is represented as a fixed-dimensional acoustic embedding obtained by mapping the sequence of feature frames to a single embedding vector. We compare our system on English and Xitsonga datasets to state-of-the-art baselines, using a variety of measures including word error rate (obtained by mapping the unsupervised output to ground truth transcriptions). We show that by imposing a consistent top-down segmentation while also using bottom-up knowledge from detected syllable boundaries, both single-speaker and multi-speaker versions of our system outperform a purely bottom-up single-speaker syllable-based approach. We also show that the discovered clusters can be made less speaker- and gender-specific by using an unsupervised autoencoder-like feature extractor to learn better frame-level features (prior to embedding). Our system’s discovered clusters are still less pure than those of two multi-speaker term discovery systems, but provide far greater coverage.
```
@article{kamper.jansen.ea:segmental,
  author = {Kamper, Herman and Jansen, Aren and Goldwater, Sharon},
  title = {A segmental framework for fully-unsupervised large-vocabulary speech recognition},
  journal = {Computer Speech and Language},
  volume = {46},
  pages = {154–174},
  month = {nov},
  year = {2017},
  annote = {Received 2021 Best Research Paper award (for a paper published in the journal during the previous five years).}
}
```
An embedded segmental K-means model for unsupervised segmentation and clustering of speech. Herman Kamper, Karen Livescu and Sharon Goldwater. In Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). 2017. (Nominated for Best Paper award. Copyright 2017 IEEE.)
[ pdf | bib | abstract ]

Unsupervised segmentation and clustering of unlabelled speech are core problems in zero-resource speech processing. Most competitive approaches lie at methodological extremes: some follow a Bayesian approach, defining probabilistic models with convergence guarantees, while others opt for more efficient heuristic techniques. Here we introduce an approximation to a segmental Bayesian model that falls in between, with a clear objective function but using hard clustering and segmentation rather than full Bayesian inference. Like its Bayesian counterpart, this embedded segmental k-means model (ES-KMeans) represents arbitrary-length word segments as fixed-dimensional acoustic word embeddings. On English and Xitsonga data, ES-KMeans outperforms a leading heuristic method in word segmentation, giving similar scores to the Bayesian model while being five times faster with fewer hyperparameters. However, there is a trade-off in cluster purity, with the Bayesian model’s purer clusters yielding about 10% better unsupervised word error rates.
```
@inproceedings{kamper.livescu.ea:embedded,
  author = {Kamper, Herman and Livescu, Karen and Goldwater, Sharon},
  title = {An embedded segmental K-means model for unsupervised segmentation and clustering of speech},
  booktitle = {Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  year = {2017},
  annote = {Nominated for Best Paper award. Copyright 2017 IEEE.}
}
```
Topic and audience effects on distinctively Scottish vocabulary usage in Twitter data. Philippa Shoemark, James Kirby and Sharon Goldwater. In Workshop on Stylistic Variation at EMNLP 2017. 2017.
[ pdf | bib | abstract ]

Sociolinguistic research suggests that speakers modulate their language style in response to their audience. Similar effects have recently been claimed to occur in the informal written context of Twitter, with users choosing less region-specific and non-standard vocabulary when addressing larger audiences. However, these studies have not carefully controlled for the possible confound of topic: that is, tweets addressed to a broad audience might also tend towards topics that engender a more formal style. In addition, it is not clear to what extent previous results generalize to different samples of users. Using mixed-effects models, we show that audience and topic have independent effects on the rate of distinctively Scottish usage in two demographically distinct Twitter user samples. However, not all effects are consistent between the two groups, underscoring the importance of replicating studies on distinct user samples before drawing strong conclusions from social media data.
```
@inproceedings{shoemark.kirby.ea:topic,
  author = {Shoemark, Philippa and Kirby, James and Goldwater, Sharon},
  title = {Topic and audience effects on distinctively Scottish vocabulary usage in Twitter data},
  booktitle = {Workshop on Stylistic Variation at EMNLP 2017},
  year = {2017},
  address = {Copenhagen, Denmark}
}
```
Aye or naw, whit dae ye hink? Scottish independence and linguistic identity on social media. Philippa Shoemark, Debnil Sur, Luke Shrimpton, Iain Murray and Sharon Goldwater. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics. 2017.
[ pdf | bib | abstract ]

Political surveys have indicated a relationship between a sense of Scottish identity and voting decisions in the 2014 Scottish Independence Referendum. Identity is often reflected in language use, suggesting the intuitive hypothesis that individuals who support Scottish independence are more likely to use distinctively Scottish words than those who oppose it. In the first large-scale study of sociolinguistic variation on social media in the UK, we identify distinctively Scottish terms in a data-driven way, and find that these terms are indeed used at a higher rate by users of pro-independence hashtags than by users of anti-independence hashtags. However, we also find that in general people are \emphless likely to use distinctively Scottish words in tweets with referendum-related hashtags than in their general Twitter activity. We attribute this difference to style-shifting relative to audience, aligning with previous work showing that Twitter users tend to use fewer local variants when addressing a broader audience.
```
@inproceedings{shoemark.sur.ea:aye,
  author = {Shoemark, Philippa and Sur, Debnil and Shrimpton, Luke and Murray, Iain and Goldwater, Sharon},
  title = {Aye or naw, whit dae ye hink? Scottish independence and linguistic identity on social media},
  booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2017},
  address = {Valencia, Spain}
}
```

2016

Unsupervised Word Segmentation and Lexicon Discovery using Acoustic Word Embeddings. Herman Kamper, Aren Jansen and Sharon Goldwater. IEEE Transactions on Audio, Speech and Language Processing 24 (4), pp. 669–679. 2016. (Copyright 2016 IEEE.)
[ preprint pdf | bib | abstract | online journal ]

In settings where only unlabelled speech data is available, speech technology needs to be developed without transcriptions, pronunciation dictionaries, or language modelling text. A similar problem is faced when modelling infant language acquisition. In these cases, categorical linguistic structure needs to be discovered directly from speech audio. We present a novel unsupervised Bayesian model that segments unlabelled speech and clusters the segments into hypothesized word groupings. The result is a complete unsupervised tokenization of the input speech in terms of discovered word types. In our approach, a potential word segment (of arbitrary length) is embedded in a fixed-dimensional acoustic vector space. The model, implemented as a Gibbs sampler, then builds a whole-word acoustic model in this space while jointly performing segmentation.We report word error rates in a small-vocabulary connected digit recognition task by mapping the unsupervised decoded output to ground truth transcriptions. The model achieves around 20% error rate, outperforming a previous HMM-based system by about 10% absolute. Moreover, in contrast to the baseline, our model does not require a pre-specified vocabulary size.
```
@article{kamper.jansen.ea:unsupervised*1,
  author = {Kamper, Herman and Jansen, Aren and Goldwater, Sharon},
  title = {Unsupervised Word Segmentation and Lexicon Discovery using Acoustic Word Embeddings},
  journal = {IEEE Transactions on Audio, Speech and Language Processing},
  volume = {24},
  number = {4},
  pages = {669–679},
  year = {2016},
  annote = {Copyright 2016 IEEE.}
}
```

Statistical Learning, Inductive Bias, and Bayesian Inference in Language Acquisition. Lisa Pearl and Sharon Goldwater. In Lidz, Jeffrey and Snyder, William and Pater, Joe, editors. Oxford Handbook of Developmental Linguistics, chapter 28, pp. 664–695. Oxford University Press. 2016.
[ preprint pdf | bib ]

@incollection{pearl.goldwater:statistical,
  author = {Pearl, Lisa and Goldwater, Sharon},
  title = {Statistical Learning, Inductive Bias, and Bayesian Inference in Language Acquisition},
  booktitle = {Oxford Handbook of Developmental Linguistics},
  chapter = {28},
  editor = {Lidz, Jeffrey and Snyder, William and Pater, Joe},
  publisher = {Oxford University Press},
  pages = {664–695},
  year = {2016}
}

Towards robust cross-linguistic comparisons of phonological networks. Philippa Shoemark, Sharon Goldwater, James Kirby and Rik Sarkar. In Proceedings of the 14th ACL SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pp. 110–120. 2016.
[ pdf | bib | abstract ]

Recent work has proposed using network science to analyse the structure of the mental lexicon by viewing words as nodes in a phonological network, with edges connecting words that differ by a single phoneme. Comparing the structure of phonological networks across different languages could provide insights into linguistic typology and the cognitive pressures that shape lan- guage acquisition, evolution, and processing. However, previous studies have not considered how statistics gathered from these networks are affected by factors such as lexicon size and the distribution of word lengths. We show that these factors can substantially affect the statistics of a phonological network and propose a new method for making more robust comparisons. We then analyse eight languages, finding many commonalities but also some qualitative differences in their lexicon structure.
```
@inproceedings{shoemark.goldwater.ea:towards,
  author = {Shoemark, Philippa and Goldwater, Sharon and Kirby, James and Sarkar, Rik},
  title = {Towards robust cross-linguistic comparisons of phonological networks},
  booktitle = {Proceedings of the 14th ACL SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology},
  pages = {110–120},
  year = {2016}
}
```

2015

Unsupervised neural network based feature extraction using weak top-down constraints. Herman Kamper, Micha Elsner, Aren Jansen and Sharon Goldwater. In Proceedings of the 40th IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 5818–5822. 2015. (Copyright 2015 IEEE.)
[ pdf | bib | abstract ]

Deep neural networks (DNNs) have become a standard component in supervised ASR, used in both data-driven feature extraction and acoustic modelling. Supervision is typically obtained from a forced alignment that provides phone class targets, requiring transcriptions and pronunciations. We propose a novel unsupervised DNN-based feature extractor that can be trained without these resources in zero-resource settings. Using unsupervised term discovery, we find pairs of isolated word examples of the same unknown type; these provide weak top-down supervision. For each pair, dynamic programming is used to align the feature frames of the two words. Matching frames are presented as input-output pairs to a deep autoencoder (AE) neural network. Using this AE as feature extractor in a word discrimination task, we achieve 64% relative improvement over a previous stateof- the-art system, 57% improvement relative to a bottom-up trained deep AE, and come to within 23% of a supervised system.
```
@inproceedings{kamper.elsner.ea:unsupervised,
  author = {Kamper, Herman and Elsner, Micha and Jansen, Aren and Goldwater, Sharon},
  title = {Unsupervised neural network based feature extraction using weak top-down constraints},
  booktitle = {Proceedings of the 40th IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2015},
  pages = {5818–5822},
  annote = {Copyright 2015 IEEE.}
}
```
Fully Unsupervised Small-Vocabulary Speech Recognition Using a Segmental Bayesian Model. Herman Kamper, Aren Jansen and Sharon Goldwater. In Proceedings of Interspeech. 2015.
[ pdf | bib | abstract ]

Current supervised speech technology relies heavily on transcribed speech and pronunciation dictionaries. In settings where unlabelled speech data alone is available, unsupervised methods are required to discover categorical linguistic structure directly from the audio. We present a novel Bayesian model which segments unlabelled input speech into word-like units, resulting in a complete unsupervised transcription of the speech in terms of discovered word types. In our approach, a potential word segment (of arbitrary length) is embedded in a fixed-dimensional space; the model (implemented as a Gibbs sampler) then builds a whole-word acoustic model in this space while jointly doing segmentation. We report word error rates in a connected digit recognition task by mapping the unsupervised output to ground truth transcriptions. Our model outperforms a previously developed HMM-based system, even when the model is not constrained to discover only the 11 word types present in the data.
```
@inproceedings{kamper.jansen.ea:fully,
  author = {Kamper, Herman and Jansen, Aren and Goldwater, Sharon},
  title = {Fully Unsupervised Small-Vocabulary Speech Recognition Using a Segmental Bayesian Model},
  booktitle = {Proceedings of Interspeech},
  year = {2015}
}
```
Talkers account for listener and channel characteristics to communicate efficiently. John K. Pate and Sharon Goldwater. Journal of Memory and Language 78, pp. 1–17. 2015.
[ preprint pdf | bib | abstract | online journal ]

A well-known effect in speech production is that more predictable linguistic constructions tend to be reduced. Recent work has interpreted this effect in an information-theoretic framework, proposing that such *predictability effects* reflect a tendency towards communicative effciency. However, others have argued that these effects are, in the terminology of Gould and Lewontin (1979), *spandrels*: incidental by-products of other processes (such as a talker-oriented tendency for low production effort). This article develops the information-theoretic framing more fully, showing that information-theoretic effciency involves different kinds of coding operations (predictability effects), not all of which are consistent with the spandrel account. Using mixed effects regressions, we analyze word durations in several spontaneous speech corpora, comparing predictability effects between infant-directed and adult-directed speech and between speech to visible and invisible listeners. We find that talkers adjust the extent to which production varies with predictability measures according to listener characteristics, and exploit an additional visual channel to eliminate phonetic redundancy. This pattern would demand multiple independent spandrel accounts, but is unified by an adaptive account. Our results broaden the scope of existing work on predictability effects and provide further evidence that these effects are tied to communicative efficiency.
```
@article{pate.goldwater:talkers,
  author = {Pate, John K. and Goldwater, Sharon},
  title = {Talkers account for listener and channel characteristics to communicate efficiently},
  journal = {Journal of Memory and Language},
  volume = {78},
  pages = {1–17},
  year = {2015}
}
```
A Comparison of Neural Network Methods for Unsupervised Representation Learning on the Zero Resource Speech Challenge. Daniel Renshaw, Herman Kamper, Aren Jansen and Sharon Goldwater. In Proceedings of Interspeech. 2015.
[ pdf | bib | abstract ]

The success of supervised deep neural networks (DNNs) in speech recognition cannot be transferred to zero-resource languages where the requisite transcriptions are unavailable. We investigate unsupervised neural network based methods for learning frame-level representations. Good frame representations eliminate differences in accent, gender, channel characteristics, and other factors to model subword units for within- and acrossspeaker phonetic discrimination. We enhance the correspondence autoencoder (cAE) and show that it can transform Mel Frequency Cepstral Coefficients (MFCCs) into more effective frame representations given a set of matched word pairs from an unsupervised term discovery (UTD) system. The cAE combines the feature extraction power of autoencoders with the weak supervision signal from UTD pairs to better approximate the extrinsic task’s objective during training. We use the Zero Resource Speech Challenge’s minimal triphone pair ABX discrimination task to evaluate our methods. Optimizing a cAE architecture on English and applying it to a zero-resource language, Xitsonga, we obtain a relative error rate reduction of 35% compared to the original MFCCs. We also show that Xitsonga frame representations extracted from the bottleneck layer of a supervised DNN trained on English can be further enhanced by the cAE, yielding a relative error rate reduction of 39%.
```
@inproceedings{renshaw.kamper.ea:comparison,
  author = {Renshaw, Daniel and Kamper, Herman and Jansen, Aren and Goldwater, Sharon},
  title = {A Comparison of Neural Network Methods for Unsupervised Representation Learning on the Zero Resource Speech Challenge},
  booktitle = {Proceedings of Interspeech},
  year = {2015}
}
```

2014

Weak semantic context helps phonetic learning in a model of infant language acquisition. Stella Frank, Naomi Feldman and Sharon Goldwater. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pp. 1073–1083. 2014.
[ pdf | bib | abstract ]

Learning phonetic categories is one of the first steps to learning a language, yet is hard to do using only distributional phonetic information. Semantics could potentially be useful, since words with different meanings have distinct phonetics, but it is unclear how many word meanings are known to infants learning phonetic categories. We show that attending to a weaker source of semantics, in the form of a distribution over topics in the current context, can lead to improvements in phonetic category learning. In our model, an extension of a previous model of joint word-form and phonetic category inference, the likelihood of word-forms is topic-dependent, enabling the model to find significantly better phonetic vowel categories and word-forms than a model with no semantic knowledge.
```
@inproceedings{frank.feldman.ea:weak,
  author = {Frank, Stella and Feldman, Naomi and Goldwater, Sharon},
  title = {Weak semantic context helps phonetic learning in a model of infant language acquisition},
  booktitle = {Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics},
  address = {Baltimore, Maryland},
  pages = {1073–1083},
  year = {2014}
}
```
Unsupervised lexical clustering of speech segments using fixed dimensional acoustic embeddings. Herman Kamper, Aren Jansen, Simon King and Sharon Goldwater. In Proceedings of the IEEE Spoken Language Technology Workshop. 2014. (Copyright 2014 IEEE.)
[ pdf | bib | abstract ]

Unsupervised speech processing methods are essential for applications ranging from zero-resource speech technology to modelling child language acquisition. One challenging problem is discovering the word inventory of the language: the lexicon. Lexical clustering is the task of grouping unlabelled acoustic word tokens according to type. We propose a novel lexical clustering model: variable-length word segments are embedded in a fixed-dimensional acoustic space in which clustering is then performed. We evaluate several clustering algorithms and find that the best methods produce clusters with wide variation in sizes, as observed in natural language. The best probabilistic approach is an infinite Gaussian mixture model (IGMM), which automatically chooses the number of clusters. Performance is comparable to that of nonprobabilistic Chinese Whispers and average-linkage hierarchical clustering. We conclude that IGMM clustering of fixed-dimensional embeddings holds promise as the lexical clustering component in unsupervised speech processing systems.
```
@inproceedings{kamper.jansen.ea:unsupervised,
  author = {Kamper, Herman and Jansen, Aren and King, Simon and Goldwater, Sharon},
  title = {Unsupervised lexical clustering of speech segments using fixed dimensional acoustic embeddings},
  booktitle = {Proceedings of the IEEE Spoken Language Technology Workshop},
  year = {2014},
  annote = {Copyright 2014 IEEE.}
}
```
POS induction with distributional and morphological information using a distance-dependent Chinese restaurant process. Kairit Sirts, Jacob Eisenstein, Micha Elsner and Sharon Goldwater. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, Volume 2: Short Papers, pp. 265–271. 2014.
[ pdf | bib | abstract ]

We present a new approach to inducing the syntactic categories of words, combining their distributional and morphological properties in a joint nonparametric Bayesian model based on the distance-dependent Chinese Restaurant Process. The prior distribution over word clusterings uses a log-linear model of morphological similarity; the likelihood function is the probability of generating vector word embeddings. The weights of the morphology model are learned jointly while inducing part-of-speech clusters, encouraging them to cohere with the distributional features. The resulting algorithm outperforms competitive alternatives on English POS induction.
```
@inproceedings{sirts.eisenstein.ea:pos,
  author = {Sirts, Kairit and Eisenstein, Jacob and Elsner, Micha and Goldwater, Sharon},
  title = {POS induction with distributional and morphological information using a distance-dependent Chinese restaurant process},
  booktitle = {Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, Volume 2: Short Papers},
  address = {Baltimore, Maryland},
  pages = {265–271},
  year = {2014}
}
```

2013

A Joint Learning Model of Word Segmentation, Lexical Acquisition, and Phonetic Variability. Micha Elsner, Sharon Goldwater, Naomi Feldman and Frank Wood. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 42–54. 2013.
[ pdf | bib | abstract ]

We present a cognitive model of early lexical acquisition which jointly performs word segmentation and learns an explicit model of phonetic variation. We define the model as a Bayesian noisy channel; we sample segmentations and word forms simultaneously from the posterior, using beam sampling to control the size of the search space. Compared to a pipelined approach in which segmentation is performed first, our model is qualitatively more similar to human learners. On data with variable pronunciations, the pipelined approach learns to treat syllables or morphemes as words. In contrast, our joint model, like infant learners, tends to learn multiword collocations. We also conduct analyses of the phonetic variations that the model learns to accept and its patterns of word recognition errors, and relate these to developmental evidence.
```
@inproceedings{elsner.goldwater.ea:joint,
  author = {Elsner, Micha and Goldwater, Sharon and Feldman, Naomi and Wood, Frank},
  title = {A Joint Learning Model of Word Segmentation, Lexical Acquisition, and Phonetic Variability},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Seattle, Washington, USA},
  pages = {42–54},
  year = {2013}
}
```
A role for the developing lexicon in phonetic category acquisition. Naomi H. Feldman, Thomas L. Griffiths, Sharon Goldwater and James L. Morgan. Psychological Review 120 (4), pp. 751–778. 2013.
[ preprint pdf | bib | abstract | online journal ]

Infants learn to segment words from fluent speech during the same period as they learn native language phonetic categories, yet accounts of phonetic category acquisition typically ignore information about the words in which sounds appear. We use a Bayesian model to illustrate how feedback from segmented words might constrain phonetic category learning by providing information about which sounds occur together in words. Simulations demonstrate that word-level information can successfully disambiguate overlapping English vowel categories. Learning patterns in the model are shown to parallel human behavior from artificial language learning tasks. These findings point to a central role for the developing lexicon in phonetic category acquisition and provide a framework for incorporating top-down constraints into models of category learning.
```
@article{feldman.griffiths.ea:role,
  author = {Feldman, Naomi H. and Griffiths, Thomas L. and Goldwater, Sharon and Morgan, James L.},
  title = {A role for the developing lexicon in phonetic category acquisition},
  journal = {Psychological Review},
  volume = {120},
  number = {4},
  pages = {751–778},
  year = {2013}
}
```
Adding sentence types to a model of syntactic category acquisition. Stella Frank, Sharon Goldwater and Frank Keller. TopiCS in Cognitive Science 5 (3), pp. 495–521. 2013.
[ preprint pdf | bib | abstract | online journal ]

The acquisition of syntactic categories is a crucial step in the process of acquiring syntax. At this stage, before a full grammar is available, only surface cues are available to the learner. Previous computational models have demonstrated that local contexts are informative for syntactic categorization. However, local contexts are affected by sentence-level structure. In this paper, we add sentence type as an observed feature to a model of syntactic category acquisition, based on experimental evidence showing that pre-syntactic children are able to distinguish sentence type using prosody and other cues. The model, a Bayesian Hidden Markov Model, allows for adding sentence type in a few different ways; we find that sentence type can aid syntactic category acquisition if it is used to characterize the differences in word order between sentence types. In these models, knowledge of sentence type permits similar gains to those found by extending the local context.
```
@article{frank.goldwater.ea:adding,
  author = {Frank, Stella and Goldwater, Sharon and Keller, Frank},
  title = {Adding sentence types to a model of syntactic category acquisition},
  journal = {TopiCS in Cognitive Science},
  year = {2013},
  volume = {5},
  number = {3},
  pages = {495–521}
}
```
Exploring the utility of joint morphological and syntactic learning from child-directed speech. Stella Frank, Frank Keller and Sharon Goldwater. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 30–41. 2013.
[ pdf | bib | abstract ]

Children learn various levels of linguistic structure concurrently, yet most existing models of language acquisition deal with only a single level of structure, implicitly assuming a sequential learning process. Developing models that learn multiple levels simultaneously can provide important insights into how these levels might interact synergistically during learning. Here, we present a model that jointly induces syntactic categories and morphological segmentations by combining two well-known models for the individual tasks. We test on child-directed utterances in English and Spanish and compare to single-task baselines. In the morphologically poorer language (English), the model improves morphological segmentation, while in the morphologically richer language (Spanish), it leads to better syntactic categorization. These results provide further evidence that joint learning is useful, but also suggest that the benefits may be different for typologically different languages.
```
@inproceedings{frank.keller.ea:exploring,
  author = {Frank, Stella and Keller, Frank and Goldwater, Sharon},
  title = {Exploring the utility of joint morphological and syntactic learning from child-directed speech},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Seattle, Washington, USA},
  pages = {30–41},
  year = {2013}
}
```
A summary of the 2012 JHU CLSP Workshop on Zero Resource Speech Technologies and Models of Early Language Acquisition. Aren Jansen, Emmanuel Dupoux, Sharon Goldwater, Mark Johnson, Sanjeev Khudanpur, Kenneth Church, Naomi Feldman, Hynek Hermansky, Florian Metze, Richard Rose, Mike Seltzer, Pascal Clark, Ian McGraw, Balakrishnan Varadarajan, Erin Bennett, Benjamin Borschinger, Justin Chiu, Ewan Dunbar, Abdallah Fourtassi, David Harwath, Chia-ying Lee, Keith Levin, Atta Norouzian, Vijay Peddinti, Rachel Richardson, Thomas Schatz and Samuel Thomas. In Proceedings of the 38th IEEE International Conference on Acoustics, Speech and Signal Processing. 2013. (Copyright 2013 IEEE.)
[ pdf | bib | abstract ]

We summarize the accomplishments of a multi-disciplinary workshop exploring the computational and scientific issues surrounding zero resource (unsupervised) speech technologies and related models of early language acquisition. Centered around the tasks of phonetic and lexical discovery, we consider unified evaluation metrics, present two new approaches for improving speaker independence in the absence of supervision, and evaluate the application of Bayesian word segmentation algorithms to automatic subword unit tokenizations. Finally, we present two strategies for integrating zero resource techniques into supervised settings, demonstrating the potential of unsupervised methods to improve mainstream technologies.
```
@inproceedings{jansen.dupoux.ea:summary,
  author = {Jansen, Aren and Dupoux, Emmanuel and Goldwater, Sharon and Johnson, Mark and Khudanpur, Sanjeev and Church, Kenneth and Feldman, Naomi and Hermansky, Hynek and Metze, Florian and Rose, Richard and Seltzer, Mike and Clark, Pascal and McGraw, Ian and Varadarajan, Balakrishnan and Bennett, Erin and Borschinger, Benjamin and Chiu, Justin and Dunbar, Ewan and Fourtassi, Abdallah and Harwath, David and Lee, Chia-ying and Levin, Keith and Norouzian, Atta and Peddinti, Vijay and Richardson, Rachel and Schatz, Thomas and Thomas, Samuel},
  title = {A summary of the 2012 JHU CLSP Workshop on Zero Resource Speech Technologies and Models of Early Language Acquisition},
  booktitle = {Proceedings of the 38th IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2013},
  annote = {Copyright 2013 IEEE.}
}
```
Modeling Graph Languages with Grammars Extracted via Tree Decompositions. Bevan K. Jones, Mark Johnson and Sharon Goldwater. In Proceedings of the 11th Conference on Finite-State Methods and Natural Language Processing. 2013.
[ pdf | bib | abstract ]

Work on probabilistic models of natural language tends to focus on strings and trees, but there is increasing interest in more general graph-shaped structures since they seem to be better suited for representing natural language semantics, ontologies, or other varieties of knowledge structures. However, while there are relatively simple approaches to defining generative models over strings and trees, it has proven more challenging for more general graphs. This paper describes a natural generalization of the n-gram to graphs, making use of Hyperedge Replacement Grammars to define generative models of graph languages.
```
@inproceedings{jones.johnson.ea:modeling,
  author = {Jones, Bevan K. and Johnson, Mark and Goldwater, Sharon},
  title = {Modeling Graph Languages with Grammars Extracted via Tree Decompositions},
  booktitle = {Proceedings of the 11th Conference on Finite-State Methods and Natural Language Processing},
  year = {2013}
}
```
Unsupervised dependency parsing with acoustic cues. John K. Pate and Sharon Goldwater. Transactions of the Association for Computational Linguistics 1(Mar), pp. 63–74. 2013.
[ pdf | bib | abstract ]

Unsupervised parsing is a difficult task that infants readily perform. Progress has been made on this task using text-based models, but few computational approaches have considered how infants might benefit from acoustic cues. This paper explores the hypothesis that word duration can help with learning syntax. We describe how duration information can be incorporated into an unsupervised Bayesian dependency parser whose only other source of information is the words themselves (without punctuation or parts of speech). Our results, evaluated on both adult-directed and child-directed utterances, show that using word duration can improve parse quality relative to words-only baselines. These results support the idea that acoustic cues provide useful evidence about syntactic structure for language-learning infants, and motivate the use of word duration cues in NLP tasks with speech.
```
@article{pate.goldwater:unsupervised*1,
  author = {Pate, John K. and Goldwater, Sharon},
  title = {Unsupervised dependency parsing with acoustic cues},
  journal = {Transactions of the Association for Computational Linguistics},
  volume = {1(Mar)},
  pages = {63–74},
  year = {2013}
}
```
Minimally-Supervised Morphological Segmentation using Adaptor Grammars. Kairit Sirts and Sharon Goldwater. Transactions of the Association for Computational Linguistics 1(May), pp. 231–242. 2013.
[ pdf | bib | abstract ]

This paper explores the use of Adaptor Grammars, a nonparametric Bayesian modelling framework, for minimally supervised morphological segmentation. We compare three training methods: unsupervised training, semisupervised training, and a novel model selection method. In the model selection method, we train unsupervised Adaptor Grammars using an over-articulated metagrammar, then use a small labelled data set to select which potential morph boundaries identified by the metagrammar should be returned in the final output. We evaluate on five languages and show that semi-supervised training provides a boost over unsupervised training, while the model selection method yields the best average results over all languages and is competitive with state-of-the-art semi-supervised systems. Moreover, this method provides the potential to tune performance according to different evaluation metrics or downstream tasks.
```
@article{sirts.goldwater:minimally-supervised,
  author = {Sirts, Kairit and Goldwater, Sharon},
  title = {Minimally-Supervised Morphological Segmentation using Adaptor Grammars},
  journal = {Transactions of the Association for Computational Linguistics},
  volume = {1(May)},
  pages = {231–242},
  year = {2013}
}
```

2012

Turning the pipeline into a loop: Iterated unsupervised dependency parsing and PoS induction. Christos Christodoulopoulos, Sharon Goldwater and Mark Steedman. In Proceedings of the NAACL-HLT Workshop on the Induction of Linguistic Structure. 2012.
[ pdf | bib ]

@inproceedings{christodoulopoulos.goldwater.ea:turning,
  author = {Christodoulopoulos, Christos and Goldwater, Sharon and Steedman, Mark},
  title = {Turning the pipeline into a loop: Iterated unsupervised dependency parsing and PoS induction},
  booktitle = {Proceedings of the NAACL-HLT Workshop on the Induction of Linguistic Structure},
  year = {2012}
}

Bootstrapping a Unified Model of Lexical and Phonetic Acquisition. Micha Elsner, Sharon Goldwater and Jacob Eisenstein. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, pp. 184–193. 2012.
[ pdf | bib | abstract ]

During early language acquisition, infants must learn both a lexicon and a model of phonetics that explains how lexical items can vary in pronunciation—for instance "the" might be realized as [Di] or [D@]. Previous models of acquisition have generally tackled these problems in isolation, yet behavioral evidence suggests infants acquire lexical and phonetic knowledge simultaneously. We present a Bayesian model that clusters together phonetic variants of the same lexical item while learning both a language model over lexical items and a log-linear model of pronunciation variability based on articulatory features. The model is trained on transcribed surface pronunciations, and learns by bootstrapping, without access to the true lexicon. We test the model using a corpus of child-directed speech with realistic phonetic variation and either gold standard or automatically induced word boundaries. In both cases modeling variability improves the accuracy of the learned lexicon over a system that assumes each lexical item has a unique pronunciation.
```
@inproceedings{elsner.goldwater.ea:bootstrapping,
  author = {Elsner, Micha and Goldwater, Sharon and Eisenstein, Jacob},
  title = {Bootstrapping a Unified Model of Lexical and Phonetic Acquisition},
  booktitle = {Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics},
  address = {Jeju Island, Korea},
  pages = {184–193},
  year = {2012}
}
```
Semantic Parsing with Bayesian Tree Transducers. Bevan K. Jones, Mark Johnson and Sharon Goldwater. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, pp. 488–496. 2012.
[ pdf | bib | abstract ]

Many semantic parsing models use tree transformations to map between natural language and meaning representation. However, while tree transformations are central to several state-of-the-art approaches, little use has been made of the rich literature on tree automata. This paper makes the connection concrete with a tree transducer based semantic parsing model and suggests that other models can be interpreted in a similar framework, increasing the generality of their contributions. In particular, this paper further introduces a variational Bayesian inference algorithm that is applicable to a wide class of tree transducers, producing state-of-the-art semantic parsing results while remaining applicable to any domain employing probabilistic tree transducers.
```
@inproceedings{jones.johnson.ea:semantic,
  author = {Jones, Bevan K. and Johnson, Mark and Goldwater, Sharon},
  title = {Semantic Parsing with Bayesian Tree Transducers},
  booktitle = {Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics},
  address = {Jeju Island, Korea},
  pages = {488–496},
  year = {2012}
}
```
A Probabilistic Model of Syntactic and Semantic Acquisition from Child-Directed Utterances and their Meanings. Tom Kwiatkowski, Sharon Goldwater, Luke Zettelmoyer and Mark Steedman. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pp. 234–244. 2012.
[ pdf | bib | abstract ]

This paper presents an incremental probabilistic learner that models the acquistion of syntax and semantics from a corpus of child-directed utterances paired with possible representations of their meanings. These meaning representations approximate the contextual input available to the child; they do not specify the meanings of individual words or syntactic derivations. The learner then has to infer the meanings and syntactic properties of the words in the input along with a parsing model. We use the CCG grammatical framework and train a non-parametric Bayesian model of parse structure with online variational Bayesian expectation maximization. When tested on utterances from the CHILDES corpus, our learner outperforms a state-of-the-art semantic parser. In addition, it models such aspects of child acquisition as "fast mapping," while also countering previous criticisms of statistical syntactic learners.
```
@inproceedings{kwiatkowski.goldwater.ea:probabilistic,
  author = {Kwiatkowski, Tom and Goldwater, Sharon and Zettelmoyer, Luke and Steedman, Mark},
  title = {A Probabilistic Model of Syntactic and Semantic Acquisition from Child-Directed Utterances and their Meanings},
  booktitle = {Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics},
  address = {Avignon, France},
  pages = {234–244},
  year = {2012}
}
```

2011

A Bayesian mixture model for part-of-speech induction using multiple features. Christos Christodoulopoulos, Sharon Goldwater and Mark Steedman. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 638–647. 2011.
[ pdf | bib | abstract ]

In this paper we present a fully unsupervised syntactic class induction system formulated as a Bayesian multinomial mixture model, where each word type is constrained to belong to a single class. By using a mixture model rather than a sequence model (e.g., HMM), we are able to easily add multiple kinds of features, including those at both the type level (morphology features) and token level (context and alignment features, the latter fromparallel corpora). Using only context features, our system yields results comparable to state-of-the art, far better than a similar model without the one-class-per-type constraint. Using the additional features provides added benefit, and our final system outperforms the best published results on most of the 25 corpora tested.
```
@inproceedings{christodoulopoulos.goldwater.ea:bayesian,
  author = {Christodoulopoulos, Christos and Goldwater, Sharon and Steedman, Mark},
  title = {A Bayesian mixture model for part-of-speech induction using multiple features},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Edinburgh, Scotland, UK.},
  pages = {638–647},
  year = {2011}
}
```
Producing power-law distributions and damping word frequencies with two-stage language models. Sharon Goldwater, Thomas L. Griffiths and Mark Johnson. Journal of Machine Learning Research 12(Jul), pp. 2335–2382. 2011.
[ pdf | bib | abstract ]

Standard statistical models of language fail to capture one of the most striking properties of natural languages: the power-law distribution in the frequencies of word tokens. We present a framework for developing statisticalmodels that can generically produce power laws, breaking generativemodels into two stages. The first stage, the generator, can be any standard probabilistic model, while the second stage, the adaptor, transforms the word frequencies of this model to provide a closer match to natural language. We show that two commonly used Bayesian models, the Dirichlet-multinomial model and the Dirichlet process, can be viewed as special cases of our framework. We discuss two stochastic processes—the Chinese restaurant process and its two-parameter generalization based on the Pitman-Yor process—that can be used as adaptors in our framework to produce power-law distributions over word frequencies. We show that these adaptors justify common estimation procedures based on logarithmic or inverse-power transformations of empirical frequencies. In addition, taking the Pitman-Yor Chinese restaurant process as an adaptor justifies the appearance of type frequencies in formal analyses of natural language and improves the performance of a model for unsupervised learning of morphology.
```
@article{goldwater.griffiths.ea:producing*1,
  author = {Goldwater, Sharon and Griffiths, Thomas L. and Johnson, Mark},
  title = {Producing power-law distributions and damping word frequencies with two-stage language models},
  journal = {Journal of Machine Learning Research},
  volume = {12(Jul)},
  pages = {2335–2382},
  year = {2011}
}
```
Formalizing Semantic Parsing with Tree Transducers. Bevan K. Jones, Mark Johnson and Sharon Goldwater. In Proceedings of the Australasian Language Technology Workshop. 2011.
[ pdf | bib | abstract ]

This paper introduces tree transducers as a unifying theory for semantic parsing models based on tree transformations. Many existing models use tree transformations, but implement specialized training and smoothing methods, which makes it difficult to modify or extend the models. By connecting to the rich literature on tree automata, we show how semantic parsing models can be developed using completely general estimation methods. We demonstrate the approach by reframing and extending one state-of-the-art model as a tree automaton. Using a variant of the inside-outside algorithm with variational Bayesian estimation, our generative model achieves higher raw accuracy than existing generative and discriminative approaches on a standard data set.
```
@inproceedings{jones.johnson.ea:formalizing,
  author = {Jones, Bevan K. and Johnson, Mark and Goldwater, Sharon},
  title = {Formalizing Semantic Parsing with Tree Transducers},
  booktitle = {Proceedings of the Australasian Language Technology Workshop},
  year = {2011}
}
```
Lexical generalization in CCG grammar induction for semantic parsing. Tom Kwiatkowski, Luke Zettelmoyer, Sharon Goldwater and Mark Steedman. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 1512–1523. 2011.
[ pdf | bib | abstract ]

We consider the problem of learning factored probabilistic CCG grammars for semantic parsing from data containing sentences paired with logical-form meaning representations. Traditional CCG lexicons list lexical items that pair words and phrases with syntactic and semantic content. Such lexicons can be inefficient when words appear repeatedly with closely related lexical content. In this paper, we introduce factored lexicons, which include both lexemes to model word meaning and templates to model systematic variation in word usage. We also present an algorithm for learning factored CCG lexicons, along with a probabilistic parse-selection model. Evaluations on benchmark datasets demonstrate that the approach learns highly accurate parsers, whose generalization performance benefits greatly from the lexical factoring.
```
@inproceedings{kwiatkowski.zettelmoyer.ea:lexical,
  author = {Kwiatkowski, Tom and Zettelmoyer, Luke and Goldwater, Sharon and Steedman, Mark},
  title = {Lexical generalization in CCG grammar induction for semantic parsing},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Edinburgh, Scotland, UK.},
  pages = {1512–1523},
  year = {2011}
}
```
Unsupervised extraction of recurring words from infant-directed speech. Fergus R. McInnes and Sharon Goldwater. In Proceedings of the 33rd Annual Conference of the Cognitive Science Society. 2011.
[ pdf | bib | abstract ]

To date, most computational models of infant word segmentation have worked from phonemic or phonetic input, or have used toy datasets. In this paper, we present an algorithm for word extraction that works directly from naturalistic acoustic input: infant-directed speech from the CHILDES corpus. The algorithm identifies recurring acoustic patterns that are candidates for identification as words or phrases, and then clusters together the most similar patterns. The recurring patterns are found in a single pass through the corpus using an incremental method, where only a small number of utterances are considered at once. Despite this limitation, we show that the algorithm is able to extract a number of recurring words, including some that infants learn earliest, such as Mommy and the child’s name. We also introduce a novel information-theoretic evaluation measure.
```
@inproceedings{mcinnes.goldwater:unsupervised,
  author = {McInnes, Fergus R. and Goldwater, Sharon},
  title = {Unsupervised extraction of recurring words from infant-directed speech},
  booktitle = {Proceedings of the 33rd Annual Conference of the Cognitive Science Society},
  year = {2011}
}
```
Predictability effects in adult-directed and infant-directed speech: Does the listener matter?. John K. Pate and Sharon Goldwater. In Proceedings of the 33rd Annual Conference of the Cognitive Science Society. 2011.
[ pdf | bib | abstract ]

A well-known effect in speech production is that more predictable words tend to be phonetically reduced. Recent work has suggested that predictability effects result from hardwired properties of the language production system, rather than active modulation by the talker to accommodate the listener. However, these studies investigated only minor manipulations of listener characteristics. Here, we examine predictability effects with two very different listener populations: adults and preverbal infants. Using mixed effects regressions on spontaneous speech corpora, we compare the effect of word frequency, probability in context, and previous mention on word duration in adult-directed and infant-directed speech. We find that the effects of preceding context and word frequency differ according to listener. Contrary to previous work, these results suggest that talkers do modulate the phonetic effects of predictability based on listener characteristics. To our knowledge, this study is also the first published analysis of predictability effects in infant-directed speech.
```
@inproceedings{pate.goldwater:predictability,
  author = {Pate, John K. and Goldwater, Sharon},
  title = {Predictability effects in adult-directed and infant-directed speech: Does the listener matter?},
  booktitle = {Proceedings of the 33rd Annual Conference of the Cognitive Science Society},
  year = {2011}
}
```
Unsupervised syntactic chunking with acoustic cues: Computational models for prosodic bootstrapping. John K. Pate and Sharon Goldwater. In Proceedings of the 2nd Workshop on Cognitive Modeling and Computational Linguistics. 2011. (Received Best Student Paper award.)
[ pdf | bib | abstract ]

Learning to group words into phrases without supervision is a hard task for NLP systems, but infants routinely accomplish it. We hypothesize that infants use acoustic cues to prosody, which NLP systems typically ignore. To evaluate the utility of prosodic information for phrase discovery, we present an HMM-based unsupervised chunker that learns from only transcribed words and raw acoustic correlates to prosody. Unlike previous work on unsupervised parsing and chunking, we use neither gold standard part-of-speech tags nor punctuation in the input. Evaluated on the Switchboard corpus, our model outperforms several baselines that exploit either lexical or prosodic information alone, and, despite producing a flat structure, performs competitively with a state-of-the-art unsupervised lexicalized parser, with a substantial advantage in precision. Our results support the hypothesis that acoustic-prosodic cues provide useful evidence about syntactic phrases for languagelearning infants.
```
@inproceedings{pate.goldwater:unsupervised,
  author = {Pate, John K. and Goldwater, Sharon},
  title = {Unsupervised syntactic chunking with acoustic cues: Computational models for prosodic bootstrapping},
  booktitle = {Proceedings of the 2nd Workshop on Cognitive Modeling and Computational Linguistics},
  year = {2011},
  annote = {Received Best Student Paper award.}
}
```

2010

Two decades of unsupervised POS induction: How far have we come?. Christos Christodoulopoulos, Sharon Goldwater and Mark Steedman. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 575–584. 2010. (This is an updated version that corrects a minor bug in the computation of the vmb measure in Figure 1 and Table 1. Thanks to Andreas Zollmann for pointing this out.)
[ pdf | bib | abstract ]

Part-of-speech (POS) induction is one of the most popular tasks in research on unsupervised NLP. Many different methods have been proposed, yet comparisons are difficult to make since there is little consensus on evaluation framework, and many papers evaluate against only one or two competitor systems. Here we evaluate seven different POS induction systems spanning nearly 20 years of work, using a variety of measures. We show that some of the oldest (and simplest) systems stand up surprisingly well against more recent approaches. Since most of these systems were developed and tested using data from the WSJ corpus, we compare their generalization abil- ities by testing on both WSJ and the multilingual Multext-East corpus. Finally, we introduce the idea of evaluating systems based on their ability to produce cluster prototypes that are useful as input to a prototype-driven learner. In most cases, the prototype-driven learner outperforms the unsupervised system used to initialize it, yielding state-of-the-art results on WSJ and improvements on non-English corpora.
```
@inproceedings{christodoulopoulos.goldwater.ea:two,
  author = {Christodoulopoulos, Christos and Goldwater, Sharon and Steedman, Mark},
  title = {Two decades of unsupervised POS induction: How far have we come?},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Cambridge, MA},
  pages = {575–584},
  year = {2010},
  annote = {This is an updated version that corrects a minor bug in the computation of the vmb measure in Figure 1 and Table 1. Thanks to Andreas Zollmann for pointing this out.}
}
```
Inducing tree substitution grammars. Trevor Cohn, Phil Blunsom and Sharon Goldwater. Journal of Machine Learning Research 11(Nov), pp. 3053–3096. 2010.
[ preprint pdf | bib | abstract | online journal ]

Inducing a grammar from text has proven to be a notoriously challenging learning task despite decades of research. The primary reason for its difficulty is that in order to induce plausible grammars, the underlying model must be capable of representing the intricacies of language while also ensuring that it can be readily learned from data. The majority of existing work on grammar induction has favoured model simplicity (and thus learnability) over representational capacity by using context free grammars and first order dependency grammars, which are not sufficiently expressive to model many common linguistic constructions. We propose a novel compromise by inferring a probabilistic tree substitution grammar, a formalism which allows for arbitrarily large tree fragments and thereby better represent complex linguistic structures. To limit the model’s complexity we employ a Bayesian non-parametric prior which biases the model towards a sparse grammar with shallow productions. We demonstrate the model’s efficacy on supervised phrase-structure parsing, where we induce a latent segmentation of the training treebank, and on unsupervised dependency grammar induction. In both cases the model uncovers interesting latent linguistic structures while producing competitive results.
```
@article{cohn.blunsom.ea:inducing,
  author = {Cohn, Trevor and Blunsom, Phil and Goldwater, Sharon},
  title = {Inducing tree substitution grammars},
  journal = {Journal of Machine Learning Research},
  volume = {11(Nov)},
  pages = {3053–3096},
  year = {2010}
}
```
Beyond transitional probabilities: Human learners impose a parsimony bias in statistical word segmentation. Michael C. Frank, Inbal Arnon, Harry Tily and Sharon Goldwater. In Proceedings of the 32nd Annual Conference of the Cognitive Science Society. 2010.
[ pdf | bib | abstract ]

Human infants and adults are able to segment coherent sequences from unsegmented strings of auditory stimuli after only a short exposure, an ability thought to be linked to early language acquisition. Although some research has hypothesized that learners succeed in these tasks by computing transitional probabilities between syllables, current experimental results do not differentiate between a range of models of different computations that learners could perform. We created a set of stimuli that was consistent with two different lexicons—one consisting of two-syllable words and one of three-syllable words—but where transition probabilities would not lead learners to segment sentences consistently according to either lexicon. Participants’ responses formed a distribution over possible segmentations that included consistent segmentations into both two- and three-syllable words, suggesting that learners do not use pure transitional probabilities to segment but instead impose a bias towards parsimony on the lexicons they learn.
```
@inproceedings{frank.arnon.ea:beyond,
  author = {Frank, Michael C. and Arnon, Inbal and Tily, Harry and Goldwater, Sharon},
  title = {Beyond transitional probabilities: Human learners impose a parsimony bias in statistical word segmentation},
  booktitle = {Proceedings of the 32nd Annual Conference of the Cognitive Science Society},
  year = {2010}
}
```
Modeling human performance in statistical word segmentation. Michael C. Frank, Sharon Goldwater, Thomas L. Griffiths and Joshua B. Tenenbaum. Cognition 117 (2), pp. 107–125. 2010.
[ preprint pdf | bib | abstract | online journal ]

The ability to discover groupings in continuous stimuli on the basis of distributional information is present across species and across perceptual modalities. We investigate the nature of the computations underlying this ability using statistical word segmentation experiments in which we vary the length of sentences, the amount of exposure, and the number of words in the languages being learned. Although the results are intuitive from the perspective of a language learner (longer sentences, less training, and a larger language all make learning more difficult), standard computational proposals fail to capture several of these results. We describe how probabilistic models of segmentation can be modified to take into account some notion of memory or resource limitations in order to provide a closer match to human performance.
```
@article{frank.goldwater.ea:modeling*1,
  title = {Modeling human performance in statistical word segmentation},
  author = {Frank, Michael C. and Goldwater, Sharon and Griffiths, Thomas L. and Tenenbaum, Joshua B.},
  journal = {Cognition},
  volume = {117},
  number = {2},
  pages = {107–125},
  year = {2010}
}
```
Using sentence type information for syntactic category acquisition. Stella Frank, Sharon Goldwater and Frank Keller. In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics at ACL. 2010.
[ pdf | bib | abstract ]

In this paper we investigate a new source of information for syntactic category acquisition: sentence type (question, declarative, imperative). Sentence type correlates strongly with intonation patterns in most languages; we hypothesize that these intonation patterns are a valuable signal to a language learner, indicating different syntactic patterns. To test this hypothesis, we train a Bayesian Hidden Markov Model (and variants) on child-directed speech. We first show that simply training a separate model for each sentence type decreases performance due to sparse data. As an alternative, we propose two new models based on the BHMM in which sentence type is an observed variable which influences either emission or transition probabilities. Both models outperform a standard BHMM on data from English, Cantonese, and Dutch. This suggests that sentence type information available from intonational cues may be helpful for syntactic acquisition crosslinguistically.
```
@inproceedings{frank.goldwater.ea:using,
  author = {Frank, Stella and Goldwater, Sharon and Keller, Frank},
  title = {Using sentence type information for syntactic category acquisition},
  booktitle = {Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics at ACL},
  year = {2010}
}
```
Which words are hard to recognize? Prosodic, lexical, and disfluency factors that increase speech recognition error rates. Sharon Goldwater, Daniel Jurafsky and Christopher D. Manning. Speech Communication 52 (3), pp. 181–200. 2010.
[ preprint pdf | bib | abstract | online journal ]

Despite years of speech recognition research, little is known about which words tend to be misrecognized and why. Previous work has shown that errors increase for infrequent words, short words, and very loud or fast speech, but many other presumed causes of error (e.g., nearby disfluencies, turn-initial words, phonetic neighborhood density) have never been carefully tested. The reasons for the huge differences found in error rates between speakers also remain largely mysterious.
Using a mixed-effects regression model, we investigate these and other factors by analyzing the errors of two state-of-the-art recognizers on conversational speech. Words with higher error rates include those with extreme prosodic characteristics, those occurring turn-initially or as discourse markers, and doubly confusable pairs: acoustically similar words that also have similar language model probabilities. Words preceding disfluent interruption points (first repetition tokens and words before fragments) also have higher error rates. Finally, even after accounting for other factors, speaker differences cause enormous variance in error rates, suggesting that speaker error rate variance is not fully explained by differences in word choice, fluency, or prosodic characteristics. We also propose that doubly confusable pairs, rather than high neighborhood density, may better explain phonetic neighborhood errors in human speech processing.
```
@article{goldwater.jurafsky.ea:which*1,
  author = {Goldwater, Sharon and Jurafsky, Daniel and Manning, Christopher D.},
  title = {Which words are hard to recognize? Prosodic, lexical, and disfluency factors that increase speech recognition error rates},
  journal = {Speech Communication},
  volume = {52},
  number = {3},
  pages = {181–200},
  year = {2010}
}
```
Inducing probabilistic CCG grammars from logical form with higher-order unification. Tom Kwiatkowski, Luke Zettelmoyer, Sharon Goldwater and Mark Steedman. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 1223–1233. 2010.
[ pdf | bib | abstract ]

This paper addresses the problem of learning to map sentences to logical form, given training data consisting of natural language sentences paired with logical representations of their meaning. Previous approaches have been designed for particular natural languages or specific meaning representations; here we present a more general method. The approach induces a probabilistic CCG grammar that represents the meaning of individual words and defines how these meanings can be combined to analyze complete sentences. We use higher-order unification to define a hypothesis space containing all grammars consistent with the training data, and develop an online learning algorithm that efficiently searches this space while simultaneously estimating the parameters of a log-linear parsing model. Experiments demonstrate high accuracy on benchmark data sets in four languages with two different meaning representations.
```
@inproceedings{kwiatkowski.zettelmoyer.ea:inducing,
  author = {Kwiatkowski, Tom and Zettelmoyer, Luke and Goldwater, Sharon and Steedman, Mark},
  title = {Inducing probabilistic CCG grammars from logical form with higher-order unification},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Cambridge, MA},
  pages = {1223–1233},
  year = {2010}
}
```

How ideal are we? Incorporating human limitations into Bayesian models of word segmentation. Lisa Pearl, Sharon Goldwater and Mark Steyvers. In Proceedings of the 34th annual Boston University Conference on Child Language Development, pp. 315–326. 2010.
[ pdf | bib ]

@inproceedings{pearl.goldwater.ea:how,
  title = {How ideal are we? Incorporating human limitations into Bayesian models of word segmentation},
  author = {Pearl, Lisa and Goldwater, Sharon and Steyvers, Mark},
  booktitle = {Proceedings of the 34th annual Boston University Conference on Child Language Development},
  pages = {315–326},
  address = {Somerville, MA},
  publisher = {Cascadilla Press},
  year = {2010}
}

Online Learning Mechanisms for Bayesian Models of Word Segmentation. Lisa Pearl, Sharon Goldwater and Mark Steyvers. Research on Language and Computation 8 (2), pp. 107-132. 2010.
[ preprint pdf | bib | abstract | online journal ]

In recent years, Bayesian models have become increasingly popular as a way of understanding human cognition. Ideal learner Bayesian models assume that cognition can be usefully understood as optimal behavior under uncertainty, a hypothesis that has been supported by a number of modeling studies across various domains (e.g., Griffiths & Tenenbaum, 2005; Xu & Tenenbaum, 2007). The models in these studies aim to explain why humans behave as they do given the task and data they encounter, but typically avoid some questions addressed by more traditional psychological models, such as how the observed behavior is produced given constraints on memory and processing. Here, we use the task of word segmentation as a case study for investigating these questions within a Bayesian framework. We consider some limitations of the infant learner, and develop several online learning algorithms that take these limitations into account. Each algorithm can be viewed as a different method of approximating the same ideal learner. When tested on corpora of English child-directed speech, we find that the constrained learner’s behavior depends non-trivially on how the learner’s limitations are implemented. Interestingly, sometimes biases that are helpful to an ideal learner hinder a constrained learner, and in a few cases, constrained learners perform equivalently or better than the ideal learner. This suggests that the transition from a computational-level solution for acquisition to an algorithmic-level one is not straightforward.
```
@article{pearl.goldwater.ea:online,
  author = {Pearl, Lisa and Goldwater, Sharon and Steyvers, Mark},
  title = {Online Learning Mechanisms for Bayesian Models of Word Segmentation},
  journal = {Research on Language and Computation},
  publisher = {Springer Netherlands},
  pages = {107-132},
  volume = {8},
  number = {2},
  year = {2010}
}
```

2009

A note on the implementation of Hierarchical Dirichlet Processes. Phil Blunsom, Trevor Cohn, Sharon Goldwater and Mark Johnson. In Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics, pp. 337–340. 2009. (This is an updated version with corrected pseudocode for Algorithm 1: line 16 was previously missing. Thanks to Weng Wei for pointing this out.)
[ pdf | bib | abstract ]

The implementation of collapsed Gibbs samplers for non-parametric Bayesian models is non-trivial, requiring considerable book-keeping. Goldwater et al. (2006a) presented an approximation which significantly reduces the storage and computation overhead, but we show here that their formulation was incorrect and, even after correction, is grossly inaccurate. We present an alternative formulation which is exact and can be computed easily. However this approach does not work for hierarchical models, for which case we present an efficient data structure which has a better space complexity than the naive approach.
```
@inproceedings{blunsom.cohn.ea:note,
  author = {Blunsom, Phil and Cohn, Trevor and Goldwater, Sharon and Johnson, Mark},
  title = {A note on the implementation of Hierarchical Dirichlet Processes},
  booktitle = {Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics},
  address = {Suntec, Singapore},
  pages = {337–340},
  year = {2009},
  annote = {This is an updated version with corrected pseudocode for Algorithm 1: line 16 was previously missing. Thanks to Weng Wei for pointing this out.}
}
```
Inducing compact but accurate tree-substitution grammars. Trevor Cohn, Sharon Goldwater and Phil Blunsom. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pp. 548–556. 2009.
[ pdf | bib | abstract ]

Tree substitution grammars (TSGs) are a compelling alternative to context-free grammars for modelling syntax. However, many popular techniques for estimating weighted TSGs (under the moniker of Data Oriented Parsing) suffer from the problems of inconsistency and overfitting. We present a theoretically principled model which solves these problems using a Bayesian non-parametric formulation. Our model learns compact and simple grammars, uncovering latent linguistic structures (e.g., verb subcategorisation), and in doing so far out-performs a standard PCFG.
```
@inproceedings{cohn.goldwater.ea:inducing,
  author = {Cohn, Trevor and Goldwater, Sharon and Blunsom, Phil},
  title = {Inducing compact but accurate tree-substitution grammars},
  booktitle = {Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics},
  address = {Boulder, Colorado},
  pages = {548–556},
  year = {2009}
}
```
Evaluating models of syntactic category acquisition without using a gold standard. Stella Frank, Sharon Goldwater and Frank Keller. In Proceedings of the 31st Annual Conference of the Cognitive Science Society. 2009.
[ pdf | bib | abstract ]

A number of different measures have been proposed for evaluating computational models of human syntactic category acquisition. They all rely on a gold standard set of manually determined categories. However, children’s syntactic categories change during language development, so evaluating against a fixed and final set of adult categories is not appropriate. In this paper, we propose a new measure, substitutable precision and recall, based on the idea that words which occur in similar syntactic environments share the same category. We use this measure to evaluate three standard category acquisition models (hierarchical clustering, frequent frames, Bayesian HMM) and show that the results correlate well with those obtained using two gold-standard-based measures.
```
@inproceedings{frank.goldwater.ea:evaluating,
  author = {Frank, Stella and Goldwater, Sharon and Keller, Frank},
  title = {Evaluating models of syntactic category acquisition without using a gold standard},
  booktitle = {Proceedings of the 31st Annual Conference of the Cognitive Science Society},
  year = {2009}
}
```
A Bayesian framework for word segmentation: Exploring the effects of context. Sharon Goldwater, Thomas L. Griffiths and Mark Johnson. Cognition 112 (1), pp. 21–54. 2009. (Results in this paper are based on a newer version of the code used in the ACL06 and BUCLD07 word segmentation papers and chapter 5 of my thesis. The new version corrects a small bug in the implementation of the bigram (HDP) model. Please cite results from this paper in future publications.)
[ preprint pdf | bib | abstract | online journal ]

Since the experiments of Saffran et al. [Saffran, J., Aslin, R., & Newport, E. (1996). Statistical learning in 8-month-old infants. Science, 274, 1926–1928], there has been a great deal of interest in the question of how statistical regularities in the speech stream might be used by infants to begin to identify individual words. In this work, we use computational modeling to explore the effects of different assumptions the learner might make regarding the nature of words – in particular, how these assumptions affect the kinds of words that are segmented from a corpus of transcribed child-directed speech. We develop several models within a Bayesian ideal observer framework, and use them to examine the consequences of assuming either that words are independent units, or units that help to predict other units. We show through empirical and theoretical results that the assumption of independence causes the learner to undersegment the corpus, with many two- and three-word sequences (e.g. what’s that, do you, in the house) misidentified as individual words. In contrast, when the learner assumes that words are predictive, the resulting segmentation is far more accurate. These results indicate that taking context into account is important for a statistical word segmentation strategy to be successful, and raise the possibility that even young infants may be able to exploit more subtle statistical patterns than have usually been considered.
```
@article{goldwater.griffiths.ea:bayesian,
  author = {Goldwater, Sharon and Griffiths, Thomas L. and Johnson, Mark},
  title = {A Bayesian framework for word segmentation: Exploring the effects of context},
  journal = {Cognition},
  volume = {112},
  number = {1},
  pages = {21–54},
  year = {2009},
  annote = {Results in this paper are based on a newer version of the code used in the ACL06 and BUCLD07 word segmentation papers and chapter 5 of my thesis. The new version corrects a small bug in the implementation of the bigram (HDP) model. Please cite results from this paper in future publications.}
}
```
Improving nonparametric Bayesian inference: Experiments on unsupervised word segmentation with adaptor grammars. Mark Johnson and Sharon Goldwater. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pp. 317–325. 2009.
[ pdf | bib | abstract ]

One of the reasons nonparametric Bayesian inference is attracting attention in computational linguistics is because it provides a principled way of learning the units of generalization together with their probabilities. Adaptor grammars are a framework for defining a variety of hierarchical nonparametric Bayesian models. This paper investigates some of the choices that arise in formulating adaptor grammars and associated inference procedures, and shows that they can have a dramatic impact on performance in an unsupervised word segmentation task. With appropriate adaptor grammars and inference procedures we achieve an 87% word token f-score on the standard Brent version of the Bernstein- Ratner corpus, which is an error reduction of over 35% over the best previously reported results for this corpus.
```
@inproceedings{johnson.goldwater:improving,
  author = {Johnson, Mark and Goldwater, Sharon},
  title = {Improving nonparametric Bayesian inference: Experiments on unsupervised word segmentation with adaptor grammars},
  booktitle = {Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics},
  address = {Boulder, Colorado},
  pages = {317–325},
  year = {2009}
}
```
Improving morphology induction by learning spelling rules. Jason Naradowsky and Sharon Goldwater. In Proceedings of IJCAI. 2009.
[ pdf | bib | abstract ]

Unsupervised learning of morphology is an important task for human learners and in natural language processing systems. Previous systems focus on segmenting words into substrings (taking) tak.ing), but sometimes a segmentation-only analysis is insufficient (e.g., taking may be more appropriately analyzed as take+ing, with a spelling rule accounting for the deletion of the stem-final e). In this paper, we develop a Bayesian model for simultaneously inducing both morphology and spelling rules. We show that the addition of spelling rules improves performance over the baseline morphology-only model.
```
@inproceedings{naradowsky.goldwater:improving,
  title = {Improving morphology induction by learning spelling rules},
  author = {Naradowsky, Jason and Goldwater, Sharon},
  booktitle = {Proceedings of IJCAI},
  year = {2009}
}
```

2008

Which words are hard to recognize? Prosodic, lexical, and disfluency factors that increase ASR error rates. Sharon Goldwater, Daniel Jurafsky and Christopher D. Manning. In Proceedings of ACL-08: HLT, pp. 380–388. 2008.
[ pdf | bib | abstract ]

Many factors are thought to increase the chances of misrecognizing a word in ASR, including low frequency, nearby disfluencies, short duration, and being at the start of a turn. However, few of these factors have been formally examined. This paper analyzes a variety of lexical, prosodic, and disfluency factors to determine which are likely to increase ASR error rates. Findings include the following. (1) For disfluencies, effects depend on the type of disfluency: errors increase by up to 15% (absolute) for words near fragments, but decrease by up to 7.2% (absolute) for words near repetitions. This decrease seems to be due to longer word duration. (2) For prosodic features, there are more errors for words with extreme values than words with typical values. (3) Although our results are based on output from a system with speaker adaptation, speaker differences are a major factor influencing error rates, and the effects of features such as frequency, pitch, and intensity may vary between speakers.
```
@inproceedings{goldwater.jurafsky.ea:which,
  author = {Goldwater, Sharon and Jurafsky, Daniel and Manning, Christopher D.},
  title = {Which words are hard to recognize? Prosodic, lexical, and disfluency factors that increase ASR error rates},
  booktitle = {Proceedings of ACL-08: HLT},
  address = {Columbus, Ohio},
  year = {2008},
  pages = {380–388}
}
```

2007

Modeling Human Performance in Statistical Word Segmentation. Michael C. Frank, Sharon Goldwater, Vikash Mansinghka, Thomas L. Griffiths and Joshua B. Tenenbaum. In Proceedings of the 29th Annual Conference of the Cognitive Science Society. 2007.
[ pdf | bib | abstract ]

What mechanisms support the ability of human infants, adults, and other primates to identify words from fluent speech using distributional regularities? In order to better characterize this ability, we collected data from adults in an artificial language segmentation task similar to Saffran, Newport, and Aslin (1996) in which the length of sentences was systematically varied between groups of participants. We then compared the fit of a variety of computational models— including simple statistical models of transitional probability and mutual information, a clustering model based on mutual information by Swingley (2005), PARSER (Perruchet & Vintner, 1998), and a Bayesian model. We found that while all models were able to successfully complete the task, fit to the human data varied considerably, with the Bayesian model achieving the highest correlation with our results.
```
@inproceedings{frank.goldwater.ea:modeling,
  title = {Modeling Human Performance in Statistical Word Segmentation},
  author = {Frank, Michael C. and Goldwater, Sharon and Mansinghka, Vikash and Griffiths, Thomas L. and Tenenbaum, Joshua B.},
  booktitle = {Proceedings of the 29th Annual Conference of the Cognitive Science Society},
  year = {2007}
}
```

Distributional cues to word segmentation: Context is important. Sharon Goldwater, Thomas L. Griffiths and Mark Johnson. In Proceedings of the 31st Boston University Conference on Language Development. 2007. (If you plan to cite results from this paper, see this note.)
[ pdf | bib ]

@inproceedings{goldwater.griffiths.ea:distributional,
  author = {Goldwater, Sharon and Griffiths, Thomas L. and Johnson, Mark},
  title = {Distributional cues to word segmentation: Context is important},
  booktitle = {Proceedings of the 31st Boston University Conference on Language Development},
  year = {2007},
  annote = {If you plan to cite results from this paper, see this note.}
}

A Fully Bayesian Approach to Unsupervised Part-of-Speech Tagging. Sharon Goldwater and Thomas L. Griffiths. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, pp. 744–751. 2007.
[ pdf | bib | abstract ]

Unsupervised learning of linguistic structure is a difficult problem. A common approach is to define a generative model and maximize the probability of the hidden structure given the observed data. Typically, this is done using maximum-likelihood estimation (MLE) of the model parameters. We show using part-of-speech tagging that a fully Bayesian approach can greatly improve performance. Rather than estimating a single set of parameters, the Bayesian approach integrates over all possible parameter values. This difference ensures that the learned structure will have high probability over a range of possible parameters, and permits the use of priors favoring the sparse distributions that are typical of natural language. Our model has the structure of a standard trigram HMM, yet its accuracy is closer to that of a state-of-the-art discriminative model (Smith and Eisner, 2005), up to 14 percentage points better than MLE. We find improvements both when training from data alone, and using a tagging dictionary.
```
@inproceedings{goldwater.griffiths:fully,
  author = {Goldwater, Sharon and Griffiths, Thomas L.},
  title = {A Fully Bayesian Approach to Unsupervised Part-of-Speech Tagging},
  booktitle = {Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics},
  address = {Prague, Czech Republic},
  year = {2007},
  pages = {744–751}
}
```
Adaptor Grammars: a Framework for Specifying Compositional Nonparametric Bayesian Models. Mark Johnson, Thomas L. Griffiths and Sharon Goldwater. In Advances in Neural Information Processing Systems 19, pp. 641–648. 2007. (This is an updated version that fixes a typo in equation 4. Thanks to Julia Hockenmaier for pointing this out.)
[ pdf | bib | abstract ]

This paper introduces adaptor grammars, a class of probabilistic models of language that generalize probabilistic context-free grammars (PCFGs). Adaptor grammars augment the probabilistic rules of PCFGs with "adaptors" that can induce dependencies among successive uses. With a particular choice of adaptor, based on the Pitman-Yor process, nonparametric Bayesian models of language using Dirichlet processes and hierarchical Dirichlet processes can be written as simple grammars. We present a general-purpose inference algorithm for adaptor grammars, making it easy to define and use such models, and illustrate how several existing nonparametric Bayesian models can be expressed within this framework.
```
@inproceedings{johnson.griffiths.ea:adaptor,
  author = {Johnson, Mark and Griffiths, Thomas L. and Goldwater, Sharon},
  title = {Adaptor Grammars: a Framework for Specifying Compositional Nonparametric Bayesian Models},
  booktitle = {Advances in Neural Information Processing Systems 19},
  pages = {641–648},
  publisher = {MIT Press},
  year = {2007},
  editor = {Schölkopf, B. and Platt, J. and Hoffman, T.},
  address = {Cambridge, MA},
  annote = {This is an updated version that fixes a typo in equation 4. Thanks to Julia Hockenmaier for pointing this out.}
}
```
Bayesian Inference for PCFGs via Markov chain Monte Carlo. Mark Johnson, Thomas L. Griffiths and Sharon Goldwater. In Proceedings of Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics, pp. 139–146. 2007.
[ pdf | bib | abstract ]

This paper presents two Markov chain Monte Carlo (MCMC) algorithms for Bayesian inference of probabilistic context free grammars (PCFGs) from terminal strings, providing an alternative to maximum-likelihood estimation using the Inside-Outside algorithm. We illustrate these methods by estimating a sparse grammar describing the morphology of the Bantu language Sesotho, demonstrating that with suitable priors Bayesian techniques can infer linguistic structure in situations where maximum likelihood methods such as the Inside-Outside algorithm only produce a trivial grammar.
```
@inproceedings{johnson.griffiths.ea:bayesian,
  author = {Johnson, Mark and Griffiths, Thomas L. and Goldwater, Sharon},
  title = {Bayesian Inference for PCFGs via Markov chain Monte Carlo},
  booktitle = {Proceedings of Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics},
  address = {Rochester, New York},
  pages = {139–146},
  year = {2007}
}
```

2006

Contextual Dependencies in Unsupervised Word Segmentation. Sharon Goldwater, Thomas L. Griffiths and Mark Johnson. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pp. 673–680. 2006.
[ pdf | bib | abstract ]

Developing better methods for segmenting continuous text into words is important for improving the processing of Asian languages, and may shed light on how humans learn to segment speech. We propose two new Bayesian word segmentation methods that assume unigram and bigram models of word dependencies respectively. The bigram model greatly outperforms the unigram model (and previous probabilistic models), demonstrating the importance of such dependencies for word segmentation. We also show that previous probabilistic models rely crucially on suboptimal search procedures.
```
@inproceedings{goldwater.griffiths.ea:contextual,
  author = {Goldwater, Sharon and Griffiths, Thomas L. and Johnson, Mark},
  title = {Contextual Dependencies in Unsupervised Word Segmentation},
  booktitle = {Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics},
  address = {Syndney, Australia},
  year = {2006},
  pages = {673–680}
}
```
Interpolating Between Types and Tokens by Estimating Power-Law Generators. Sharon Goldwater, Thomas L. Griffiths and Mark Johnson. In Advances in Neural Information Processing Systems 18, pp. 459–466. 2006. (This is a corrected version.)
[ pdf | bib | abstract ]

Standard statistical models of language fail to capture one of the most striking properties of natural languages: the power-law distribution in the frequencies of word tokens. We present a framework for developing statistical models that generically produce power-laws, augmenting standard generative models with an adaptor that produces the appropriate pattern of token frequencies. We show that taking a particular stochastic process—the Pitman-Yor process—as an adaptor justifies the appearance of type frequencies in formal analyses of natural language, and improves the performance of a model for unsupervised learning of morphology.
```
@inproceedings{goldwater.griffiths.ea:interpolating,
  author = {Goldwater, Sharon and Griffiths, Thomas L. and Johnson, Mark},
  title = {Interpolating Between Types and Tokens by Estimating Power-Law Generators},
  booktitle = {Advances in Neural Information Processing Systems 18},
  pages = {459–466},
  publisher = {MIT Press},
  address = {Cambridge, MA},
  year = {2006},
  annote = {This is a corrected version.}
}
```

Nonparametric Bayesian Models of Lexical Acquisition. Sharon Goldwater. Ph.D. Dissertation, Brown University, 2006. (This is the tree-saving version of my thesis, single-spaced with minimal front matter. The official version is double-spaced and contains more front matter. If you plan to cite results on word segmentation, see this note.)
[ pdf | bib ]

@phdthesis{goldwater:nonparametric,
  author = {Goldwater, Sharon},
  title = {Nonparametric Bayesian Models of Lexical Acquisition},
  school = {Brown University},
  year = {2006},
  annote = {This is the tree-saving version of my thesis, single-spaced with minimal front matter. The official version is double-spaced and contains more front matter. If you plan to cite results on word segmentation, see this note.}
}

A Non-Parametric Bayesian Approach to Spike Sorting. Frank Wood, Sharon Goldwater and Michael Black. In Proceedings of the 28th IEEE Conference on Engineering in Medicine and Biologicial Systems. 2006.
[ pdf | bib | abstract ]

In this work we present and apply infinite Gaussian mixture modeling, a non-parametric Bayesian method, to the problem of spike sorting. As this approach is Bayesian, it allows us to integrate prior knowledge about the problem in a principled way. Because it is non-parametric we are able to avoid model selection, a difficult problem that most current spike sorting methods do not address. We compare this approach to using penalized log likelihood to select the best from multiple finite mixture models trained by expectation maximization. We show favorable offline sorting results on real data and discuss ways to extend our model to online applications.
```
@inproceedings{wood.goldwater.ea:non-parametric,
  author = {Wood, Frank and Goldwater, Sharon and Black, Michael},
  title = {A Non-Parametric Bayesian Approach to Spike Sorting},
  booktitle = {Proceedings of the 28th IEEE Conference on Engineering in Medicine and Biologicial Systems},
  year = {2006}
}
```

2005

Representational Bias in Unsupervised Learning of Syllable Structure. Sharon Goldwater and Mark Johnson. In Proceedings of the Ninth Conference on Computational Natural Language Learning (CONLL ’05), pp. 112–119. 2005.
[ pdf | bib | abstract ]

Unsupervised learning algorithms based on Expectation Maximization (EM) are often straightforward to implement and provably converge on a local likelihood maximum. However, these algorithms often do not perform well in practice. Common wisdom holds that they yield poor results because they are overly sensitive to initial parameter values and easily get stuck in local (but not global) maxima. We present a series of experiments indicating that for the task of learning syllable structure, the initial parameter weights are not crucial. Rather, it is the choice of model class itself that makes the difference between successful and unsuccessful learning. We use a language-universal rule-based algorithm to find a good set of parameters, and then train the parameter weights using EM. We achieve word accuracy of 95.9% on German and 97.1% on English, as compared to 97.4% and 98.1% respectively for supervised training.
```
@inproceedings{goldwater.johnson:representational,
  author = {Goldwater, Sharon and Johnson, Mark},
  title = {Representational Bias in Unsupervised Learning of Syllable Structure},
  booktitle = {Proceedings of the Ninth Conference on Computational Natural Language Learning (CONLL ’05)},
  address = {Ann Arbor, Michigan},
  pages = {112–119},
  year = {2005}
}
```
Improving Statistical MT Through Morphological Analysis. Sharon Goldwater and David McClosky. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 676–683. 2005.
[ pdf | bib | abstract ]

In statistical machine translation, estimating word-to-word alignment probabilities for the translation model can be difficult due to the problem of sparse data: most words in a given corpus occur at most a handful of times. With a highly inflected language such as Czech, this problem can be particularly severe. In addition, much of the morphological variation seen in Czech words is not reflected in either the morphology or syntax of a language like English. In this work, we show that using morphological analysis to modify the Czech input can improve a Czech-English machine translation system. We investigate several different methods of incorporating morphological information, and show that a system that combines these methods yields the best results. Our final system achieves a BLEU score of .333, as compared to .270 for the baseline word-to-word system.
```
@inproceedings{goldwater.mcclosky:improving,
  author = {Goldwater, Sharon and McClosky, David},
  title = {Improving Statistical MT Through Morphological Analysis},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  address = {Vancouver, British Columbia, Canada},
  pages = {676–683},
  year = {2005}
}
```

2004

Priors in Bayesian Learning of Phonological Rules. Sharon Goldwater and Mark Johnson. In Proceedings of the Seventh Meeting of the ACL Special Interest Group in Computational Phonology (SIGPHON ’04). 2004.
[ pdf | bib | abstract ]

This paper describes a Bayesian procedure for unsupervised learning of phonological rules from an unlabeled corpus of training data. Like Goldsmith’s Linguistica program (Goldsmith, 2004b), whose output is taken as the starting point of this procedure, our learner returns a grammar that consists of a set of signatures, each of which consists of a set of stems and a set of suffixes. Our grammars differ from Linguistica’s in that they also contain a set of phonological rules, specifically insertion, deletion and substitution rules, which permit our grammars to collapse far more words into a signature than Linguistica can. Interestingly, the choice of Bayesian prior turns out to be crucial for obtaining a learner that makes linguistically appropriate generalizations through a range of different sized training corpora.
```
@inproceedings{goldwater.johnson:priors,
  author = {Goldwater, Sharon and Johnson, Mark},
  title = {Priors in Bayesian Learning of Phonological Rules},
  booktitle = {Proceedings of the Seventh Meeting of the ACL Special Interest Group in Computational Phonology (SIGPHON ’04)},
  address = {Barcelona, Spain},
  year = {2004}
}
```

2003

Statically finding errors in spreadsheets. Yanif Ahmad, Tudor Antoniu, Sharon Goldwater and Shriram Krishnamurthi. In Proceedings of the IEEE International Conference on Software Engineering. 2003.
[ pdf | bib | abstract ]

We describe a methodology for detecting user errors in spreadsheets, using the notion of units as our basic elements of checking. We define the concept of a header and discuss two types of relationships between headers, namely is-a and has-a relationships. With these, we develop a set of rules to assign units to cells in the spreadsheet. We check for errors by ensuring that every cell has a well-formed unit. We describe an implementation of the system that allows the user to check Microsoft Excel spreadsheets. We have run our system on practical examples, and even found errors in published spreadsheets.
```
@inproceedings{ahmad.antoniu.ea:statically,
  author = {Ahmad, Yanif and Antoniu, Tudor and Goldwater, Sharon and Krishnamurthi, Shriram},
  title = {Statically finding errors in spreadsheets},
  booktitle = {Proceedings of the IEEE International Conference on Software Engineering},
  year = {2003}
}
```
Learning OT Constraint Rankings Using a Maximum Entropy Model. Sharon Goldwater and Mark Johnson. In Proceedings of the Workshop on Variation within Optimality Theory, pp. 113–122. 2003.
[ pdf | bib | abstract ]

A weakness of standard Optimality Theory is its inability to account for grammars with free variation. We describe here the Maximum Entropy model, a general statistical model, and show how it can be applied in a constraint-based linguistic framework to model and learn grammars with free variation, as well as categorical grammars. We report the results of using the MaxEnt model for learning two different grammars: one with variation, and one without. Our results are as good as those of a previous probabilistic version of OT, the Gradual Learning Algorithm (Boersma, 1997), and we argue that our model is more general and mathematically well-motivated.
```
@inproceedings{goldwater.johnson:learning,
  author = {Goldwater, Sharon and Johnson, Mark},
  title = {Learning OT Constraint Rankings Using a Maximum Entropy Model},
  booktitle = {Proceedings of the Workshop on Variation within Optimality Theory},
  address = {Stockholm University},
  pages = {113–122},
  year = {2003}
}
```

2000

Building a Robust Dialogue System with Limited Data. Sharon Goldwater, Elizabeth Owen Bratt, Jean-Mark Gawron and John Dowding. In Proceedings of the NAACL Workshop on Conversational Systems. 2000.
[ pdf | bib ]

@inproceedings{goldwater.bratt.ea:building,
  author = {Goldwater, Sharon and Bratt, Elizabeth Owen and Gawron, Jean-Mark and Dowding, John},
  title = {Building a Robust Dialogue System with Limited Data},
  booktitle = {Proceedings of the NAACL Workshop on Conversational Systems},
  address = {Seattle, WA},
  year = {2000}
}

Compiling language models from a linguistically motivated unification grammar. Manny Rayner, Beth Ann Hockey, Frankie James, Elizabeth Owen Bratt, Sharon Goldwater and Jean Mark Gawron. In Proceedings of the 18th Conference on Computational linguistics, Volume 2, pp. 670–676. 2000.
[ pdf | bib | abstract ]

Systems now exist which are able to compile unification grammars into language models that can be included in a speech recognizer, but it is so far unclear whether non-trivial linguistically principled grammars can be used for this purpose. We describe a series of experiments which investigate the question empirically, by incrementally constructing a grammar and discovering what problems emerge when successively larger versions are compiled into nite state graph representations and used as language models for a medium-vocabulary recognition task.
```
@inproceedings{rayner.hockey.ea:compiling,
  title = {Compiling language models from a linguistically motivated unification grammar},
  author = {Rayner, Manny and Hockey, Beth Ann and James, Frankie and Bratt, Elizabeth Owen and Goldwater, Sharon and Gawron, Jean Mark},
  booktitle = {Proceedings of the 18th Conference on Computational linguistics, Volume 2},
  pages = {670–676},
  year = {2000}
}
```

1999

Interpreting language in context in CommandTalk. John Dowding, Elizabeth Owen Bratt and Sharon Goldwater. In Proceedings of Communicative Agents: The Use of Natural Language in Embodied Systems. 1999.
[ pdf | bib ]

@inproceedings{dowding.bratt.ea:interpreting,
  author = {Dowding, John and Bratt, Elizabeth Owen and Goldwater, Sharon},
  title = {Interpreting language in context in CommandTalk},
  booktitle = {Proceedings of Communicative Agents: The Use of Natural Language in Embodied Systems},
  year = {1999}
}

1998

Edge-based best-first chart parsing. Eugene Charniak, Sharon Goldwater and Mark Johnson. In Proceedings of the Sixth Workshop on Very Large Corpora at COLING-ACL. 1998.
[ pdf | bib ]

@inproceedings{charniak.goldwater.ea:edge-based,
  author = {Charniak, Eugene and Goldwater, Sharon and Johnson, Mark},
  title = {Edge-based best-first chart parsing},
  booktitle = {Proceedings of the Sixth Workshop on Very Large Corpora at COLING-ACL},
  year = {1998}
}

Thanks to Charles Sutton for the scripts used to generate this page automatically from my BibTeX file. You too can download the scripts here.