Victor Lavrenko

Research Interests: My primary research interest is in developing effective and efficient ways of finding relevant information in large semi-structured databases. I am particularly interested in challenging scenarios, where relevance cannot be captured by conventional methods of word (or feature) overlap. To give an example of such scenarios, consider the case of a novice user searching for highly technical information. Unaware of the domain-specific lingo, the user is unlikely to include pertinent keywords in her query, leading to unsatisfactory results from a conventional search engine. An extreme version of this scenario may involve an English-speaking user trying to locate relevant information in a database of Chinese documents -- even if the user is an experienced searcher, she will have a hard time formulating an effective query. Further aggravating the scenario, we could hypothesize that the Chinese documents are actually scans of handwritten messages, the quality of handwriting being beyond the capabilities of optical character recognition.

To address the scenarios described above, I have focused my research on creating a new framework for modeling the notion of relevance. The new framework is based on the idea that user's requests and items relevant to those requests should be modeled as independent random samples generated by the same underlying distribution. I developed a nonparametric Bayesian method for estimating this underlying relevance model and successfully applied the idea to several different retrieval settings.

SIGIR'09 Tutorial [Part I], [Part II]

Projects I've worked on:

Generative Approaches to Modeling Relevance.
° A Generative Theory of Relevance (dissertation)
° Relevance Models in Information Retrieval (book chapter)
° Optimal Mixture Models in IR (ECIR 2002) best student paper
° Relevance-Based Language Models (SIGIR 2001)
° Relevance Feedback and Personalization: A Language Modeling Perspective
° Localized Smoothing for Multinomial Language Models (tech. report)
° Formal Multiple-Bernoulli Models for Language Modeling (SIGIR 2004)

Activity Modeling. I have collaborated with Anton Leuski on applying relevance modeling to the task of modeling human actions in a social environment. Our goal is to analyze communications between the participants and pinpoint messages relevant to certain collaborative activities. One example of such activity could be a group of players in an online role-playing game organizing a raid on a hostile castle. Our approach involves constructing a joint distribution of message content and relevant actions taken by the sender and recipient after communicating.
° Proposal for a WS'06 summer workshop at Johns Hokpins University

Handwriting Recognition and Retrieval. In a joint work with Toni Rath and R. Manmatha, I adapted the relevance modeling framework to the problem of searching collections of highly-degraded handwritten documents. Our approach relied on a joint model of word shape and word meaning and was the first successful solution to this challenging problem.
° Holistic Word Recognition for Handwritten Historical Documents (DIAL 2004)
° A Search Engine for Historical Manuscript Images (SIGIR 2004)
° Retrieving Historical Manuscripts using Shape (CIIR tech. report)

Automatic Image and Video Annotation I have collaborated with R.Manmatha, Jiwoon Jeon and Shaolei Feng to extend the relevance modeling framework to include real-valued variables, such as feature functions used in computer vision. Our research resulted in a highly accurate method for automatically assigning keywords to unlabeled photographs and video segments. The algorithm currently represents the best-performing way for content-based searching of unlabeled images.
° Associating Words and Pictures Using Relevance Models (under review)
° Multiple Bernoulli Relevance Models for Image and Video Annotation (CVPR 2004)
° Statistical Models for Automatic Video Annotation and Retrieval (ICASSP 2004)
° A Model for Learning the Semantics of Pictures, (NIPS 2003)
° Image Annotation and Retrieval using Cross-Media Relevance Models (SIGIR 2003)

AEnalyst (from e-Analyst) is a market-forecasting technology that combines advances in the fields of Information Retrieval and Time Series Analysis. Ænalyst uses piecewise regression to identify trends in stock prices and employs language modeling techniques to associate trends with content of news stories.
° Project Information
° Electronic Analyst of Stock Behavior (CIIR tech. report)
° Mining of Concurrent Text and Time Series (KDD 2000)
° Mining of Concurrent Text and Time Series (full version)
° Language Models for Financial News Recommendation (CIKM 2000)

Cross-language Information Retrieval. Together with Martin Choquette and Bruce Croft, I have extended relevance modeling to cross-language retrieval, where an English query is used to find relevant documents in Chinese. The algorithm relies on a parallel corpus to estimate a joint distribution of English and Chinese word sets, which is used to model the user's information need. The algorithm is significantly more accurate than approaches based on a dictionary or on machine translation.
° Cross-Lingual Relevance Models (SIGIR 2002)

Bounds in Stochastic Processes.
° First Story Detection in TDT Is Hard (CIKM 2000)
° A Mathematical Model of Vocabulary Growth (tech. report)
° Comparing Effectiveness in TDT and IR
° Detections, Bounds, and Timelines: UMass and TDT-3

Topic Detection and Tracking.
° A Month to Topic Detection and Tracking in Hindi (TOIS 2003)
° Explorations within Topic Tracking and Detection (Kluwer 2002)
° Language-specific Models in Multilingual Topic Tracking (SIGIR 2004)
° Relevance Models for Topic Detection and Tracking (HLT 2002)
° Monitoring the News: a TDT demonstration system (HLT 2001)
° On-line New Event Detection and Tracking (SIGIR 1998)
° UMass TDT 2003 Research Summary
° UMass at TDT 2002
° UMass at TDT 2000
° UMASS Approaches to Detection and Tracking at TDT2
° Topic-Based Novelty Detection
° Event Tracking