Research

Bayesian Modeling

Faculty	Jordan Boyd-Graber, Naomi Feldman, Hal Daumé III, Philip Resnik
Postdocs	Taesun Moon
Graduate Students	Viet-An Nguyen Yuening Hu, Ke Zhai

What we do

Bayesian modeling is a rigorous mathematical formalism that allows us to build systems that reflect our uncertainty about the world. Applied to language, they allow us to build models that reflect the "latent" aspects of communication such as topic, part of speech, syntax, or sentiment. Using posterior inference, we can use the models to discover the latent features that best explain observed language.

In the CLIP lab, we are interested in

• building tools that make it easier for people to work with Bayesian models
• scaling inference for Bayesian models up to Web scale
• understanding how humans interpret and understand the latent variables in Bayesian models

Publications

• Hierarchical Bayes Compiler
• Reading Tea Leaves: How Humans Interpret Topic Models
• Gibbs Sampling for the Uninitiated

Machine Translation

Faculty	Bonnie Dorr, Mary Harper, Philip Resnik, Hal Daumé III
Postdocs	Kristy Hollingshead
Graduate Students	Vladimir Eidelman

What we do

The CLIP Laboratory's current work in machine translation continues the lab's long tradition of research in this area. Like most of the field, we work within the framework of statistical MT, but with an emphasis on taking appropriate advantage of knowledge driven or linguistically informed model structures, features, and priors. Some current areas of research include syntactically informed language models, linguistically informed translation model features, the use of unsupervised methods in translation modeling, exploitation of large scale "cloud computing" methods, and human-machine collaborative translation via crowdsourcing.

Publications

• Chris Dyer and Philip Resnik. Context-free reordering, finite-state translation. In Proceedings of NAACL-HLT 2010, Los Angeles, CA, USA, 2010.
• Hendra Setiawan, Chris Dyer, and Philip Resnik. Discriminative Word Alignment with a Function Word Reordering Model. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cambridge, MA, USA, 2010.
• Matt Snover, Nitin Madnani, Bonnie Dorr, and Richard Schwartz, TER-Plus: Paraphrases, Semantic, and Alignment Enhancements to Translation Edit Rate, Machine Translation, 23:2-3, Springer Netherlands, pp. 117-127, 2009.

Some Project Pages

• Monotrans: Translation Crowdsourcing without Bilinguals

Paraphrase

Yakov Kronrod

What we do

Paraphrase, the ability to express the same meaning in multiple ways, is an active area of research within the NLP community and here in the CLIP Laboratory. Our work in paraphrase includes the use of paraphrase in MT evaluation and parameter estimation, lattice and forest translation, and collaborative translation, as well as research on lexical and phrasal semantic similarity measures, meaning preservation in machine translation and summarization, and large-scale document similarity computation via cloud computing methods.

• Monotrans: Crowdsourcing Translation without Bilinguals

Publications

• Generating Phrasal & Sentential Paraphrases: A Survey of Data-Driven Methods. 2010. Computational Linguistics, 36(3). Nitin Madnani and Bonnie Dorr.
• Philip Resnik, Olivia Buzek, Chang Hu, Yakov Kronrod, Alex Quinn, Benjamin B. Bederson. Improving Translation via Targeted Paraphrasing, 2010 Conference on Empirical Methods in Natural Language Processing, October 2010.
• Yuval Marton, Saif Mohammad, and Philip Resnik. Estimating Semantic Distance Using Soft Semantic Constraints in Knowledge-Source / Corpus Hybrid Models'. Conference on Empirical Methods in Natural Language Processing (EMNLP). Singapore, August 6-7, 2009.
• Nitin Madnani, Necip Fazil Ayan, Philip Resnik, Bonnie Dorr. Using Paraphrases for Parameter Tuning in Statistical Machine Translation. 2007. Proceedings of the Second ACL Workshop on Statistical Machine Translation (WMT-07).

Text Summarization

Faculty

Bonnie Dorr David Zajic Hal Daumé III

What we do

Text Summarization is the creation of a short document to serve as a surrogate for a longer document. The CLIP Laboratory's approach to summarization enhances the extractive method of selecting source document sentences for inclusion in a summary by using sentence compression to enlarge the pool of available sentences, and by combining fluent text with topic terms. Our sentence compression technology has encompassed both statistical and linguistic methodologies. We have developed an extrinsic evaluation measure for summarization, Relevance Prediction, which is grounded in a real-world task using summarized documents. The CLIP Laboratory, in collaboration with BBN, has been a regular participant in NIST's summarization evaluations (Document Understanding Conferences and Text Analysis Conferences), and has contributed summarization components to DARPA Translingual Information Detection, Extraction and Summarization (TIDES), Surprise Language Exercise (SLE), and Global Autonomous Language Exploitation (GALE) programs, and to the iOpener project.

Parsing and Tagging

Graduate Students Vladimir Eidelman Publications Huang and Harper, EMNLP 2009: Self-Training PCFG Grammars with Latent Annotations Across Languages Huang, Eidelman and Harper NAACL 2009: Improving A Simple Bigram HMM Part-of-Speech Tagger by Latent Annotation and Self-Training Computational Social Science Faculty Jordan Boyd-Graber, Bonnie Dorr, Jimmy Lin, Douglas W. Oard, Louiqa Raschid, Philip Resnik, Amy Weinberg Graduate Students Viet-An Nguyen Hassan Sayyadi Shanchan Wu What we do Computational social science involves the use of computational methods and models to leverage "the capacity to collect and analyze data at a scale that may reveal patterns of individual and group behaviors". Research in the CLIP Laboratory is at the forefront of this emerging area, and includes sentiment analysis (computational modeling and prediction of opinions, perspective, and other private states), automatic analysis and visualization of the scientific literature, modeling the diffusion of technological innovations, modeling and prediction of social goals and actions such as persuasion, monitoring and prediction (tracking events, predicting new links or articles) and recommendation (personalized recommendations, learning to rank). Publications Greene and Resnik, NAACL 2009: More Than Words: Syntactic Packaging and Implicit Sentiment Boyd-Graber and Resnik, EMNLP 2010: Holistic Sentiment Analysis Across Languages: Multilingual Supervised Latent Dirichlet Allocation Joshi, Y., Rand, W. and Raschid, L. Diffusion and Ranking in Digital Social Media Information Retrieval Faculty Jimmy Lin, Douglas W. Oard Postdocs Earl Wagner, William Webber Graduate Students Mossaab Bagdouri, Sergey Golitsynskiy, Ferhan Ture, Tan Xu What we do The goal of information retrieval is to help people find what they are looking for. Information retrieval research in the CLIP lab focuses principally on retrieval based on the language contained in text, in speech, and in document images. We work across a broad range of content types, from tweets to tomes, from talking to texting, and from Cebuano to Chinese. Three perspectives inform our work: we integrate a broad range of computational linguistics techniques, we focus on scalable techniques that can accommodate very large collections we sometimes draw the boundaries of our “systems” very broadly to include both the automated tools that we create and the process by which users can best employ those tools. One example that illustrates these perspectives is our work with “cross-language information retrieval,” in which close coupling of machine translation and information retrieval techniques make it possible for people to find and use information written in languages that they can neither read nor write. Another example is our work on the design and evaluation of “question answering” systems that can automatically find and present answers to complex questions, which serves as a bridge between our work on information retrieval and summarization. Publications Douglas W. Oard, "Multilingual Information Access," in Encyclopedia of Library and Information Sciences, 3rd Ed., 2009. Project: Development and Evaluation of Search Technology for Discovery of Evidence in Civil Litigation Disambiguation Faculty Jordan Boyd-Graber, Judith Klavans, Philip Resnik Graduate Students Raul David Guerra What we do Disambiguation is the process of determining the meaning or senses of a word in its context; disambiguation remains one of the most challenging NLP problems since discovering word senses involves syntactic, semantic and pragmatic contextual inferencing, along with a rich knowledge base to base selection upon. For example, the word "wing" in the theater differs from airplanes, yet another sense for furniture ("wing chair") applies to some usages. Often disambiguation can be based on windows of two and three words, but usually involves larger computation. Techniques for disambiguation range from the use of large scale thesaural resources (such as WordNet) to purely statistical methods. Publications T3: Text, Tags, Trust A Topic Model for Word Sense Disambiguation Philip Resnik and David Yarowsky, "Distinguishing Systems and Distinguishing Senses: New Evaluation Methods for Word Sense Disambiguation", Natural Language Engineering 5(2), pp. 113-133. Annotation and Sense-making Faculty Judith Klavans Louiqa Raschid Graduate Students Hassan Sayyadi Shanchan Wu }| Annotation and tagging are ways to enhance knowledge in structured or semi-structured resources. Annotation typically references terms from a controlled vocabulary or ontology and is popular in bibliographic, scientific or museum collections. Tagging is more common in social media to tag images and documents and of course the now ubiquitous hashtags tweets. Sense-making or discovery is the process of extracting knowledge from these annotated or tagged resources and could range from simple counting to data/text mining to graph pattern recognition. In the CLIP lab, we are interested in the following tasks: Tagging and sense-making Pattern discovery in annotated graph datasets from the biomedical domain. Data mining with Linked Data. Representative Publications and Project Pages: NSF DBI 1147144 Methodology for Pattern Creation, Imprint Validation, and Discovery from the Annotated Biological Web (PattArAn) Pattern Discovery, Validation, and Hypothesis Development from the Annotated Biological Web