Dialogue Systems Group
School of Informatics
Edinburgh University

conference deadlines (by Joel Tetreault)

We normally meet every second thursday in the Informatics Forum room 4.31, at 2pm. If there is no meeting scheduled, feel free to use the room for your own discussion!

Please email Oliver Lemon : olemon@inf.ed.ac.uk if you would like to propose a talk or meeting theme.

2008 meetings

Thursday 27th November, IF 4.31/4.33, 2pm Markus Guhe:
"Adapting referring expressions to properties of the task environment"
Thursday 16th october IF4.31 / 4.33 at 2pm Manuel Giuliani
"MultiML -- A General Purpose Representation Language for Multimodal Human Utterances"
Thursday 8 may, 2pm Myrosia Dzikovska:
"Diagnosing natural language answers to support adaptive tutoring" (FLAIRS dry run, chair J. Moore, abstract below)
Tuesday 13 may, 2pm, Jason Williams
(visiting from AT&T) *note unusual day*
Thursday 22 may, 2pm: 2 talks:
- Srini Janarthanam and O. Lemon: "User simulations for online adaptation and knowledge-alignment in Troubleshooting dialogue systems" (SEMdial dry run)
- Verena Rieser and O. Lemon: "Learning Effective Multimodal Dialogue Strategies from Wizard-of-Oz data: Bootstrapping and Evaluation" (ACL dry run)

Recent talk abstracts:


Markus Guhe:
"Adapting referring expressions to properties of the task environment" Thursday 27 November, 2pm
Abstract: Speakers use referring expressions to identify an object in the environment. To generate a referring expression, features of the intended referent have to be selected that distinguish the object from the other potential referents. Current accounts of referring expressions consider a number of factors that influence the choice of features but ignore the influences of the task environment. In particular, they do not address how these influences change the generation of referring expressions over an extended period of time. I will present results of how colour terms are used to describe landmarks in a task-oriented dialogue (a route communication task) and describe a computational cognitive model of the observed adaptations over time.
Thursday 16th october IF4.31 / 4.33 at 2pm
Manuel Giuliani will present a dry run of his ICMI 2008 paper: "MultiML -- A General Purpose Representation Language for Multimodal Human Utterances"
Abstract: Humans communicate with each other in many ways. They use their whole body to give information about their current status, their emotions, and their intentions. But how can the information from different modalities - speech, gestures, gazes, etc. - be described in a way so that autonomous agents, including computers and robots, are able to understand these human utterances?
In this talk we present our approach to represent speech and gestures. With this approach, we are able to represent input from several modalities, as well as the relationships between these modalities. Since we separate general parts of representation from more context-specific aspects, it can easily be adapted for use in a wide range of contexts.
"Diagnosing natural language answers to support adaptive tutoring"
Myrosia Dzikovska, Thursday 8 may, 2pm

Understanding answers to open-ended explanation questions is important in intelligent tutoring systems. Existing systems use natural language techniques in essay analysis, but revert to scripted interaction with short-answer questions during remediation, making adapting dialogue to individual students difficult. We describe a corpus study that shows that there is a relationship between the types of faulty answers and the remediation strategies that tutors use; that human tutors respond differently to different kinds of correct answers; and that re-stating correct answers is associated with improved learning. We describe a design for a diagnoser based on this study that supports remediation in open-ended questions and provides an analysis of natural language answers that enables adaptive generation of tutorial feedback for both correct and faulty answers.
  • 13th May, Jason Williams
    Title: Recent work on POMDP-based dialog systems at AT&T
    AT&T Labs - Research
    www.research.att.com/info/jdw
    Abstract:

    Building spoken dialog systems is difficult because speech recognition errors are common and user's behavior is unpredictable, which introduces uncertainty in the current state of the conversation. At AT&T, we have been applying partially observable Markov decision processes (POMDPs) to building these systems. We model the uncertainty in the dialog state explicitly as a Bayesian network and apply machine learning techniques to determine what the system should say or do.

    In this talk, I'll review the overall approach of applying statistical techniques and then describe two recent advances: first, because the system must operate in real-time, efficient Bayesian inference is crucial, yet the set of possible dialog states is enormous. To solve this, I'll present a technique which uses a particle filter to perform approximate inference in real-time. Second, to choose actions, ideally we would like to combine the robustness of machine optimization with the expertise of human designers. To tackle this, I'll present a method which unifies human expertise with automatic optimization.

    To illustrate these techniques, I'll provide examples of two dialog systems: a voice dialer, and a troubleshooting system that helps users restore connectivity on a failed DSL connection. Graphical displays illustrate the operation of the techniques, and quantitative results show that applying statistical techniques outperforms the traditional method of building systems by hand.
    March 13th 2008:

    Ivan Meza Ruiz: "Accurate Statistical Spoken Language Understanding from limited development resources"

  • January 17th 2008 (thursday 2pm): Michael Kaisser : Working with PowerSet Labs

    I this talk I will give an overview of what I was up to during my internship at Powerset in San Francisco. I will mostly concentrate on how we employed Amazon's Mechanical Turk (MTurk - http://www.mturk.com) for several data collection and evaluation purposes.

    Amazon promotes this web-service as "Artificial Artificial Intelligence" and it is used in a wide variety of ways, including market research, psychological studies and also a fair number of language-related experiments. Subjects are invited for a small reward to work on "Human Intelligence tasks" (HITs), which can display a wide range of contents and provide many different input options. MTurk provides the benefit of a readily available, large pool of subjects, which perform tasks surprisingly fast and cheap. (Potential drawbacks exist and will be addressed in the talk.)

    At Powerset we used MTurk for several purposes:

    * To carry out relevance evaluation of (actual and manually-created) search engine results--in order to find out how users like responses from state-of-the-art search engines and also to test how they react to alternative forms of content and/or presentation.

    * To build up data collections--for example a large set of question/answer sentence pairs, which is used to study the linguistic relations involved and can also serve as a data set for machine learning purposes.

    If time permits and the audience is interested I will also show some of Powerset's first demos.

    2007 talk abstracts:

  • December 13th 2007 (thursday 2pm): Lucas Dixon

    "Plans, Actions and Dialogue using Linear Logic"

    We propose a framework, based on Linear Logic, for finding and executing plans that including dialogue with the aim of simplifying agent design. In particular, we provide a model that allows significant reuse of agent specifications and makes agents robust to unexpected events and failures. Using Linear Logic as the foundational machinery improves upon previous dialogue systems by providing a clear underlying logical model for both planning and execution. The resulting framework has been implemented and several case studies have been considered. Further applications include human-computer interfaces as well as agent interaction in the semantic web.

  • December 6th 2007 (thursday 2pm): Diane Litman

    University of Pittsburgh (Currently Leverhulme Visiting Professor, University of Edinburgh)

    "User Simulation for Spoken Dialogue Systems"

    User simulation is increasingly being used in the development of spoken dialogue systems. In contrast to experiments with human subjects, user simulation can generate large corpora of user behaviors in a low cost and time efficient manner. While many studies have verified that simulation models can be trained from small real corpora, how well these models can simulate realistic human behaviors, and how realistic the models need to be for various dialogue system tasks, are still open questions. In this talk I will present an examination of these issues in the context of ITSPOKE, a spoken dialogue tutorial system.

    First, I will first present a study examining the differentiating power of prior evaluation measures. Our results show that while prior measures can distinguish real corpora from simulated ones, the measures cannot help us to draw conclusions on the reality of the simulated corpora since even two real corpora can be very different when evaluated on the same measures.

    Next, I will explore the utility of more realistic versus more exploratory types of user simulation models, for the task of using Markov Decision Processes and Reinforcement Learning to automatically learn optimal dialogue strategies. Our results suggest that with sparse training data, a model that aims to randomly explore more dialogue state spaces with certain constraints performs better than a more complex model that simulates realistic user behaviors in a statistical way.

    Finally, for use in other types of applications (e.g. system evaluation) where a more realistic model is needed, I will present a new simulation model and an associated evaluation metric, based on user knowledge consistency and learning curves. Our experiments show that our new model performs similarly to real students and outperforms our previous simulation models.

    This is joint research with Hua Ai, University of Pittsburgh.

  • November 20th 2007 (Tuesday, 3pm: joint event with Planning and Language Group)

    "Planning and Learning in Dialogue Systems", Oliver Lemon:

    We'll first briefly survey the main uses of planning in dialogue systems:

    1) domain planning, recipes, and plan recognition for input interpretation ("what in the 'world' is the user trying to do?": e.g. COLLAGEN, WITAS, DUDE, BEETLE etc);
    2) dialogue planning ("what should I say next?": all dialogue systems);
    3) planning for NLG ("how should I say it?": e.g. FLIGHTS).

    Different planning techniques are appropriate for each of these tasks, but they must communicate via shared context representations. Focusing on dialogue planning, I'll describe recent advances in decision-theoretic/statistical dialogue planning (e.g. TALK project results) which deal with the issues of noise, uncertainty, and optimization. To illustrate, I'll demonstrate a system (REALL) learning to optimize its dialogue plans under different noise conditions and time constraints, for different types of user. I'll finally attempt to describe the main research questions and directions arising from these techniques.

  • November 8th 2007 (thursday 2pm) : Maria Wolters, Florian Goedde, Sebastian Moeller

    "Adapting an Environmental Control System to Older Users"

    MeMo project (Deutsche Telekom) / MATCH project (Edinburgh)

  • Thursday 30th August, 1pm, HCRC Seminar room Ron Petrick (Edinburgh)

    "Planning Dialogue Actions"

    The problem of planning dialog moves can be viewed as an instance of the more general AI problem of planning with incomplete information and sensing. Sensing actions complicate the planning process since such actions engender potentially infinite state spaces. We adapt the Linear Dynamic Event Calculus (LDEC) to the representation of dialog acts using insights from the PKS planner, and show how this formalism can be applied to the problem of planning mixed-initiative collaborative discourse. SIGdial 2007 paper

  • Thursday 19th July, 2pm, HCRC Seminar room Silvia Quarteroni (University of York)

    TITLE: Towards personalized, interactive Question Answering: challenges and possible solutions.

    ABSTRACT: As a growing field of interest for both the academia and the industry, Question Answering (QA) must face some of its traditional limitations. Among these,I will discuss the following three: the difficulty of responding to complex questions (such as definitions), the lack of personalization and the absence of interactivity.

    In this talk, I will briefly present the architecture of YourQA, a web QA system with a user modelling component and an interactive interface, which attempts to address the challenges above. I will first describe the user modelling component, which is able to filter and re-rank results based on the user's reading abilities and interests. I will then discuss the dialogue component, which is based on a chatbot and is able to converse with the user and perform basic anaphora and ellipsis resolution. Finally, I will illustrate the results of a study on the use of Support Vector Machines and Tree Kernel functions for the complex task of classifying and re-ranking definition answers.

  • Thursday 14 June, 2pm, HCRC Seminar room

    Dialogues with robots: overview of JAST and INDIGO projects: Amy Isard, Jon Oberlander, Mary-Ellen Foster

  • Thursday 12 April, 2pm, HCRC Seminar room Tutorial dialogue systems and adaptive response generation in BEE and LeActiveMath: Colin Matheson

  • Thursday 29 March, 2pm, HCRC Seminar room The DUDE development environment for spoken dialogue systems: Oliver Lemon, Xingkun Liu

  • Thursday 22nd March, 2pm, HCRC Seminar room

    Speaker: Bonnie Webber

    Interactive Question Answering

  • Thursday 18th January, 2pm, HCRC Seminar room

    Speaker: Mary-Ellen Foster

    Communication Frameworks for Multimodal Dialogue Systems

    I will describe and compare three different communication frameworks that can be used for implementing multimodal dialogue systems: MULTIPLATFORM (http://multiplatform.sourceforge.net/), Open Agent Architecture (http://www.openagent.com/), and the Internet Communications Engine (http://www.zeroc.com/). I will summarise the advantages and disadvantages of each framework for developing and running dialogue systems, using examples from my own experience in working on systems using all three frameworks.

  • Thursday 30th november, 2pm, HCRC Seminar room

    Speaker: Heriberto Cuayahuitl

    TITLE: Reinforcement Learning of Dialogue Strategies Using Hierarchical Abstract Machines

    ABSTRACT: In this paper we propose partially specified dialogue strategies for dialogue strategy optimization, where part of the strategy is specified deterministically and the rest optimized with Reinforcement Learning (RL). To do this we apply RL with Hierarchical Abstract Machines (HAMs). We also propose to build simulated users using HAMs, incorporating a combination of hierarchical deterministic and probabilistic behaviour. We performed experiments using a single-goal flight booking dialogue system, and compare two dialogue strategies (deterministic and optimized) using three types of simulated user (novice, experienced and expert). Our results show that HAMs are promising for both dialogue optimization and simulation, and provide evidence that indeed partially specified dialogue strategies can outperform deterministic ones (on average 4.7 fewer system turns) with faster learning than the traditional RL framework.

    READING MATERIAL: http://homepages.inf.ed.ac.uk/s0456904/publications/ham-slt2006.pdf

  • 2pm, Thursday 19th october, HCRC seminar room, 2 Buccleuch Place:

    2.00 -- 2.45 Simon Keizer (University of Tilburg)

    2.45 - 3.30 Joel Tetreault (University of Pittsburgh)

    Simon Keizer:

    "Multidimensional Dialogue Management in Interactive Question Answering"

    In the IMIX research program (Interactive Multimodal Information eXtraction), several Dutch universities participate in different projects on the topics of speech recognition, question answering, information presentation and dialogue management. The projects collaborate in developing a demonstrator system for interactive question answering on the domain of medical encyclopedic information.

    The PARADIME project (PARallel Agent-based DIalogue Management Engine) is one of two dialogue management projects in IMIX. In the project, we develop a dialogue manager that takes the multidimensional nature of communication into account, by means of the representation of user and system utterances in terms of dialogue acts from a multidimensional taxonomy. The taxonomy distinguishes between dimensions of task/domain acts, auto- and allo-feedback, several dimensions of interaction-management (like turn- and time-management), and social obligations management acts.

    In the talk, I will discuss our dialogue management approach in the context of interactive question answering, the problems we run into and ideas on how to tackle these problems.

    2.45 - 3.30

    Joel Tetreault (University of Pittsburgh)

    "How much data is enough? (Experiments with Confidence Bounds and MDPs)"

    Data sparsity is one of the major issues that NLP researchers always wrestle with. That is, does one have enough data to make reliable conclusions in an experiment? Using Reinforcement Learning to improve a spoken dialogue system is no exception. Past approaches in this area have simply assumed that there was enough collected data to support a certain state and action space, or used thousands of user simulations to overcome the sparsity issue. In this talk, we present a methodology of confidence bounds on the expected reward to address the problem of data sparsity in MDP's. We show how this methodology works by apply it to a prior experiment of using MDP's to predict the best features to include in a model of the dialogue state. We also show how this approach has applications in model switching and user simulations.

  • HCRC Conference Suite, 2pm, thursday 5th October 2006
    Mary McGee-Wood (Manchester University), a corpus of multi-party tutorial dialogues

    Tutorial Dialogues: the pleasures and perils of real data

    Real-world human-human, keyboard-to-keyboard tutorial dialogues, conducted in a chat-room environment, support distance learning courses in the Department of Computer Science at the University of Manchester. I will present representative samples from some 300 hours of tutorials, and highlight ways in which their use of language differs from data sets elicited by experiments. Our material suggests a range of research questions, on corpus management, dialogue strategies, tutorial strategies, and on-line communities: I will look in particular at the definition and manifestations of "initiative", and at information level.

  • Possible other talks in 2006/7
    Masja Kempen

    Personalization and Personality in NLG: M-PIRO, JAST, CRAG, and COMIC: Mary-Ellen Foster and Amy Isard

    Bayesian user simulations: Roi Georgila

    Programming by Voice

    Demos from EACL (DUDE and TALK)

    Hierarchical Reinforcement Learning of dialogue strategies: Reall-Dude (Lemon and Liu)

    Statistical dialogue management using POMDPs: Oliver Lemon, Jamie Henderson

    Previous meetings (2006)

  • HCRC seminar room, 2pm , thursday 14th September 2006
    "User Simulation for Spoken Dialogue Systems: Learning and Evaluation"

    Kallirroi Georgila, James Henderson, Oliver Lemon

    We propose ``advanced'' n-grams as a new technique for simulating user behaviour in spoken dialogue systems, and we compare it with two methods used in our prior work, i.e. linear feature combination and ``normal'' n-grams. All methods operate on the intention level and can incorporate speech recognition and understanding errors. In the linear feature combination model user actions (lists of "speech act, task" pairs) are selected, based on features of the current dialogue state which encodes the whole history of the dialogue. The user simulation based on ``normal'' n-grams treats a dialogue as a sequence of lists of "speech act, task" pairs. Here the length of the history considered is restricted by the order of the n-gram. The ``advanced'' n-grams are a variation of the normal n-grams, where user actions are conditioned not only on speech acts and tasks but also on the current status of the tasks, i.e. whether the information needed by the application (in our case flight booking) has been provided and confirmed by the user. This captures elements of goal-directed user behaviour. All models were trained and evaluated on the COMMUNICATOR corpus, to which we added annotations for user actions and dialogue context. We then evaluate how closely the synthetic responses resemble the real user responses by comparing the user response generated by each user simulation model in a given dialogue context (taken from the annotated corpus) with the actual user response. We propose the expected accuracy, expected precision, and expected recall evaluation metrics as opposed to standard precision and recall used in prior work. We also discuss why they are more appropriate metrics for evaluating user simulation models compared to their standard counterparts. The advanced n-grams produce higher scores than the normal n-grams for small values of n, which proves their strength when little amount of data is available to train larger n-grams. The linear model produces the best expected accuracy but with respect to expected precision and expected recall it is outperformed by the large n-grams even though it is trained using more information. As a task-based evaluation, we also run each of the user simulation models against a system policy trained on the same corpus. Here the linear feature combination model outperforms the other methods and the advanced n-grams outperform the normal n-grams for all values of n, which again shows their potential. We also calculate the perplexity of the different user models.

    "Cluster-based User Simulations for Learning Dialogue Strategies and the SUPER evaluation metric"

    Verena Rieser and Oliver Lemon

    Good dialogue strategies in spoken dialogue systems help to ensure and maintain mutual understanding and thus play a crucial role in robust conversational interaction. We focus on clarification strategies and build user simulations which are critical for reinforcement learning, which is a cheap and principled way to automatically optimise dialogue management. In this paper we present a novel cluster-based technique for building user simulations which show varying, but complete and consistent behaviour with respect to real users. We use this technique to build user simulations and we also introduce the SUPER evaluation metric which allows us to evaluate user simulations with respect to these desiderata. We show that the cluster-based user simulation technique performs significantly better (at p<0.01) than decisions made using either the one most likely action or a random baseline. The cluster-based user simulations reduce the average error of these other models by 53% and 34% respectively.

  • HCRC seminar room, 11am, thursday 7th September 2006
    Myroslava O. Dzikovska, Charles B. Callaway, Matthew Stone, Johanna D. Moore

    "Understanding student input for tutorial dialogue in procedural domains"

    We present an analysis of student language input in a corpus of tutoring dialogue in the domain of symbolic differentiation. Our focus on procedural tutoring makes the dialogue comparable to collaborative problem-solving (CPS). Existing CPS models describe the process of negotiating plans and goals, which also fits procedural tutoring. However, we provide a classification of student utterances and corpus annotation which shows that approximately 28% of non-trivial student language in this corpus is not accounted for by existing models, and addresses other functions, such as evaluating past actions or correcting mistakes. Our analysis can be used as a foundation for improving models of tutoring dialogue.

  • HCRC seminar room, 2pm , thursday 7th September 2006
    Pei-Yun (Sabrina) Hsueh

    "Automatic decision detection in multiparty dialogue"

    This study addresses the problem of detecting and labeling decision-making dialogue (DM detection and labeling) from a lengthy archive of meeting recordings. The problem is central to the automatic extraction and summarization of critical information from the recorded conversation speech. Yet it has posed a challenge to the mainstream spoken language understanding and summarization techniques, which have made assumptions violated by the spontaneous, face-to-face dialogues in meetings. Also, these mainstream techniques aim to understand the overall meaning of the information conveyed throughout a meeting, whereas this study aims to provide a different type of speicialized summaries, focused on only the opinionated speech related to what have been decided. The ambiguous borderline between the decision-making dialogue and general discussion has introduced another dimension of difficulty in applying the previously developed techniques on the decision detection task.

    However, there are characteristic features unique to the DM dialogue that can be exploited for its detection. In this study, I have first conducted an empirical analysis on the potentially characteristic features (e.g., word use, prosody, discourse, meeting context) of DM and constructed models that combine these features for DM classification. Results have suggested that DM do exhibit demonstrable differences in a wide range of features. Therefore, the first of this study aims to integrate these characteristic features to detect DM at both the subdialogue level and the topic segment level. To further examine the correspondence between the decision characteristic features and the level of DM tendency in conversation units, this study also proposes to apply a regression- based analysis framework for modeling such correspondence. To evolve a better DM labeling component, the second part of this study is in search of a scheme to better represent what a decision is about in the dialogue. The search has led us to the schema theory, which supports abstracting the contexts into a collection of knowledge arranged in diagrammatic form. A manually derived schema-based model is thus proposed for extracting DM-related subdialogues.

    If time allows, I will give a brief introduction to our task-based evaluation plan, which aims to examine the effectiveness of embedding an automatic decision detection and labeling component for supporting computer-supported collaborative work in a group context. As a more amibitous goal is to develop a DM decision that can be easily generalizable to new contexts for online processing, I will also talk about the problem of developing computational models with minimum labeled data in meeting contexts, with a focus on exploring unsupervised lexical approaches motivated by research in sentiment analysis for detecting opinioated speech, and machine learning strategies that exploit existing labeled data.

  • HCRC seminar room, 2pm, thursday 20th July 2006
    Theresa Wilson

    Talk: Subjectivity Analysis and Recognizing Contextual Polarity

    The goals of subjectivity analysis are to extract opinions, sentiments, emotions, and other private states expressed in natural language discourse and to recognize their components and properties. This is currently a very active area of research in natural language processing, with the potential to develop tools supporting information analysts in governmental, commercial, and political domains who want to automatically track attitudes and feelings in the news and other forums. In this talk, I will briefly describe a corpus annotated with rich information about opinions and sentiments. I will then present experiments using that data to develop and evaluate an automatic system for recognizing the "contextual polarity" of expressions, i.e., whether a phrase is being used to express a positive or negative sentiment, considering the context in which it appears. Finally, I will present highlights from some recent experiments exploiting opinion-type analysis in question answering.

  • HCRC seminar room, 2pm, thursday 22nd June 2006
    "Cooperative, symmetrical human-robot dialogue in JAST"

    Mary Ellen Foster

    The overall goal of the JAST project ("Joint Action Science and Technology"; http://www.euprojects-jast.net/) is to investigate the cognitive and communicative aspects of jointly-acting agents, both human and artifcial. The human-robot dialogue system being built as part of the project is designed to be a platform to integrate the project's empirical fndings on cognition and dialogue with its work on autonomous robots, by supporting multimodal human-robot collaboration on a joint construction task. A distinctive feature of the JAST system is that the robot and the user are true peers in the interaction: either can decide how to proceed, and -- in principle -- either can also perform any of the required actions.

    In this talk, I will describe the construction task that the JAST system addresses, show a video of the current system capabilities, and describe the sorts of interaction that will be supported as the system develops. I will then describe the system architecture, concentrating on the dialogue manager, which is based on Blaylock and Allen's (2005) collaborative problem-solving model of dialogue.

  • HCRC seminar room, 2pm, thursday 1st June 2006
    Myroslava O. Dzikovska, 2 talks

    "Increasing the coverage of a domain independent dialogue lexicon with VerbNet"

    This paper investigates how to extend coverage of a domain independent lexicon tailored for natural language understanding. We introduce two algorithms for adding lexical entries from VerbNet to the lexicon of the TRIPS spoken dialogue system. We report results on the efficiency of the method, discussing in particular precision versus coverage issues and implications for mapping to other lexical databases.

    "Backbone Extraction and Pruning for Speeding Up a Deep Parser for Dialogue Systems"

    In this paper we discuss issues related to speeding up parsing with wide-coverage unification grammars. We demonstrate that state-of-the-art optimisation techniques based on backbone parsing before unification do not provide a general solution, because they depend on specific properties of the grammar formalism that do not hold for all unification based grammars. As an alternative, we describe an optimisation technique that combines ambiguity packing at the constituent structure level with pruning based on local features.

  • HCRC seminar room, 2pm, thursday 18th May 2006
    Joe Polifroni

    "Learning Database Content for Spoken Dialogue System Design"

    One of the most common applications for spoken dialogue systems is as an interface to structured databases, where database entities are represented in terms of a set of attributes and their values. To provide access to such databases, many spoken dialogue systems are configured to have an initial information-gathering phase, where the values for certain attributes are elicited from the user. The attributes (also referred to as "constraints"), as well as a default order of elicitation, are specified by a system developer familiar with the domain. The goal of the information gathering phase is to reduce the number of database tuples to a subset of the data that can be easily described. Any information-providing utterances by the system are thus delayed until the system has elicited enough constraints to query the database.

    Dialogue systems designed in this way have several limitations. First, users may wish to browse the data, either because they are unfamiliar with the domain or because they do not have strong preferences. Second, this type of dialogue interaction often leads to the generation of over-specific constraints with unsatisfiable queries. Third, a default order of constraint elicitation may encourage the user to provide information that will not narrow down their choices in the most efficient way.

    Our hypothesis is that these limitations can be addressed in a domain independent way by applying machine learning techniques to the automation of constraint order elicitation and content selection. This is an extension of earlier work on data-driven content selection for generation of summaries in spoken dialogue systems. We have extended that work to include a machine learning component and applied it to dialogue management in a London restaurant domain. In this talk, I will describe the techniques used and provide some preliminary results of our experiments in using these techniques for generation of spoken dialogue system responses.

  • 2pm, thursday 27th april, seminar room Buccleuch place
    Speaker: Heriberto Cuayahuitl

    Title: "Learning Multi-Goal Dialogue Strategies Using Reinforcement Learning With Reduced State-Action Spaces"

    Abstract: Learning dialogue strategies using the reinforcement learning framework is problematic due to its expensive computational cost. In this paper we propose an algorithm that reduces a state-action space to one which includes only valid state-actions. We performed experiments on full and reduced spaces using three systems (with 5, 9 and 20 slots) in the travel domain using a simulated environment. The task was to learn multi-goal dialogue strategies optimizing single and multiple confirmations. Average results using strategies learnt on reduced spaces reveal the following benefits against full spaces: 1) less computer memory (94% reduction), 2) faster learning (93% faster convergence) and better performance (8.4% less time steps and 7.7% higher reward).

  • HCRC seminar room, 2pm, thursday 23rd March 2006
    Interpretation and Generation in a Knowledge-Based Tutorial System

    Myroslava O. Dzikovska, Charles B. Callaway, Elaine Farrow

    We discuss how deep interpretation and generation can be integrated with a knowledge representation designed for question answering to build a tutorial dialogue system. We use a knowledge representation known to perform well in answering exam-type questions and show that to support tutorial dialogue it needs additional features, in particular, compositional representations for interpretation and structured explanation representations.

  • HCRC seminar room, 2pm, thursday 9th March 2006
    Nuria Bertomeu (joint talk with QA group)

    "Contextual phenomena and thematic relations in database QA dialogues: data from a Wizard-of-Oz experiment"

    In this talk I will present a corpus of interactive database Question Answering obtained from a Wizard-of-Oz experiment and the annotation scheme designed for it. Then I will show data regarding the thematic connectedness of questions with the preceding discourse, on the one hand, and data regarding the occurrence of contextual phenomena, concretely ellipsis, anaphora, and definite descriptions, on the other hand. I will look at the coocurrence of certain types of thematic relations and contextual phenomena and will address the question of whether the resolution of contextual phenomena can benefit from the tracking of thematic relations.

  • HCRC seminar room, 2pm, thursday 9th February 2006
    Report on FP7 meeting, "implicit" and "perceptual" interfaces, e.g. Gwindows: Oliver Lemon
  • HCRC seminar room, 2pm, thursday 23rd February 2006
    POMDPs and reinforcement learning of dialogue management strategies: Jamie Henderson

    The Young Researchers' Roundtable (YRR 06): Verena Rieser

  • HCRC seminar room, 2pm, thursday 26th january 2006
    Organizational meeting

    Previous meetings (2005)

  • 2pm thursday 8th december 2005, HCRC seminar room
    Jamie Henderson will speak on: "Reinforcement Learning in Information State Update dialogue systems: new results on learning from data and simulations"
  • HCRC seminar room, 2pm, thursday 24th november
    Dave Toney will discuss Michael English and Peter Heeman: "Learning Mixed Initiative Dialog Strategies By Using Reinforcement Learning On Both Conversants" in Proc HLT/NAACL 2005 paper available at: "Learning Mixed Initiative Dialog Strategies By Using Reinforcement Learning On Both Conversants"
  • 3rd November 2005, HCRC seminar room, 2pm
    Meeting theme: Simulating users of dialogue systems: techniques and evaluation methodologies
    Speakers: Roi Georgila will discuss work following on from "Learning User Simulations for Information State Update Dialogue Systems", Eurospeech 2005
  • 13th October 2005, HCRC seminar room, 2pm
    Meeting theme: Dialogue interfaces to Question-answering Systems
    Discussion paper: A. Popescu and O. Etzioni and H. Kautz, "Towards a theory of natural language interfaces to databases" In Proceedings of the conference on Intelligent User Interfaces, 2003
    See also: article on the practical difficulties of natural language interfaces to databases, written by Ami Kronfeld: "Why you still can't talk to your computer"
  • 11th October 2005, 2pm, Conference Suite, Buccleuch Place
    "Language as action in a Social Setting", Peter Wallis
    Abstract:
    In 2001 we set out to create an ECA - embodied conversational agent - that would act as a virtual assistant for command and control. The intention was not to produce a useful product, but to have a user scenario that would allow us to develop and test new tools, and to evaluate new commercial products in context. In preliminary experiments we set out to determine the amount and type of real world knowledge an ECA would need in our particular context. The conclusion was that information about the world was relatively easy; what mattered were social skills. Whereas most software is written as a tool to be used by a person, the agent model of software focuses on systems that perform autonomous action, situated in an environment. In the case of ECA, that environment is full of people who live and work with other people. When Searle was writing about speech acts, his interests were, in his own words, somewhere between a theory of language and a theory of action. What does it mean to "reason for action rather than knowledge" in a social setting? My interests today are primarily in giving dialog systems some social intelligence. Politeness, gossip, the use of humour, and playing roles are all important given we treat computers as social actors. In this talk I start with some examples from my initial work and show how Brown and Levinson's concept of "face" applies to ECA. Finding examples of social skills that ECA lack is interesting (and often amusing) but what is needed is a systematic framework. The talk finishes with a proposal for such a framework based on a layered model of social norms.
  • "Designing Information-Seeking Conversational Interfaces using Reinforcement Learning", Dave Toney
    Wednesday 15th June, 10am, HCRC Conference Suite.
    Abstract:
    Designing an information-seeking conversational interface is a time-consuming and difficult process. The principal objective of my PhD is to develop a notation that will reduce development time and require less expertise on the part of a system designer. I will make use of reinforcement learning techniques to help automate the design of spoken dialogue applications. I will investigate a variety of state representations and reward functions. Finally, I will evaluate the authoring notation using established performance criteria.
  • "Probabilistic Head-Driven Parsing for Discourse Structure", Alex Lascarides
    Thursday 28th april, 2pm , HCRC Conference Suite.
    joint work with Jason Baldridge
    Abstract:
    We propose a data intensive approach to building interpretable discourse structures for appointment scheduling dialogues. Our strategy involves a tree-based representation of discourse over which we define several probabilistic head-driven parsing models. We evaluate performance on recovering labelled and unlabelled discourse relations. Our results show that discourse-based information, such as turn-taking and domain specific goals, has a large positive impact on model performance. The best model performs significantly better on labelled relations than a baseline that connects an utterance to its context in the most frequent way.
  • "The TALK project", Oliver Lemon, Roi Georgila, James Henderson
    Thursday 14th april, 2pm , HCRC Seminar room
    i) overview of TALK's research themes and current progress:
    - unifying multimodality and multilinguality (GF grammars)
    - reconfigurable systems and ontology-based dialogue
    - multimodal generation strategies
    - in-home and in-car dialogue systems
    ii) learning dialogue policies from COMMUNICATOR data - initial results.
    see http://www.talk-project.org for more details, project publications, etc.
  • "Alignment between humans and computers during dialogue", Jamie Pearson
    Thursday 17th February, 2pm, HCRC seminar room
    People: Martin Pickering, Holly Branigan, Jamie Pearson, Janet McLean (Edinburgh), Cliff Nass, John Hu (Stanford).
    Abstract:
    In conversation, people tend to align linguistically with their interaction partner, using them same words and phrases. Additionally, people treat computers as social actors, so that they interact with them and evaluate them just like other people.
    We are applying psychological models of human-human dialogue to human-computer dialogue to determine the extent to which people align linguistically with computers in the way that they do with other people.
    In a number of experiments, we have shown that people align with computers' language. This alignment occurs with spoken input/output or text input/output. It is sensitive to the user's belief about what they are interacting with (computer vs. human via CMC link) and characteristics of the interface (e.g., presence of a human face).
  • "Anytime Example-Driven Realization with CCG", Michael White, HCRC, Edinburgh University
    HCRC seminar room, thursday 27/1/05 at 2pm
    Abstract:
    In this talk I'll describe an anytime surface realization algorithm for Combinatory Categorial Grammar (CCG) and its implementation in the OpenCCG open source realizer. For motivation, I'll begin by showing how a CCG realizer can generate contextually appropriate specifications of prosody --- based on the information structure of its input messages --- as part of a spoken language dialogue system that tailors the information it presents to user preferences. Next, I'll give an overview of the novel hybrid symbolic-statistical algorithm, which brings together previous work on both (i) chart realization with unification grammars and (ii) overgeneration and ranking with n-gram models. I'll then describe an example-driven approach to developing n-gram models for use in a dialogue system, inspired by analogy to unit selection in speech synthesis and by memory-based approaches to NLP tasks. In a case study, I'll show how such n-gram models guide the anytime search towards preferred realizations, yielding substantial reductions in realization times compared to previous two-stage, packing/unpacking approaches. I'll also show that since the grammar overgenerates only mildly, the n-gram ranking delivers very high quality outputs, as measured against the target realizations by the number of exact matches and by BLEU scores. To conclude, I'll discuss how the example-driven approach provides a simple way to loosely couple language generation with speech synthesis, and speculate on future directions for exploiting this connection.
  • "The FLIGHTS project", Oliver Lemon, Johanna Moore, Kallirroi Georgila, HCRC, Edinburgh University
    HCRC seminar room, thursday 20/1/05 at 2pm
    FLIGHTS overview
    DEMO and walkthrough
    - Oliver , Roi OAA monitor overview and log
    DIPPER .is and .urules
    O-Plan content and information structure planning
    ATK recognizer
    Open CCG
    Next steps (for TALK project):
    porting/extending to "in-car" domain
    Multimodality
    Integrating learned dialogue policies
    Last update: OL, September 2005