AGENDA:

1. 10:00-10:30 Zhao Jin, "Math Information Retrieval: User Requirements and Prototype Implementation" (JCDL)
2. 10:30-10:50 Kan Min-Yen, "Slide Image Retrieval: A Preliminary Study" (JCDL, Short paper)
3. 10:50-11:20 Michael Brown, "User-Assisted Ink-Bleed Correction for Handwritten Documents" (JCDL)
4. 11:20-11:40 Kan Min-Yen, "The ACL Anthology Reference Corpus: A Reference Dataset for Bibliographic Research in Computational Linguistics" (LREC)
5. 11:40-12:00 Kan Min-Yen, "ParsCit: An open-source CRF reference string parsing package" (LREC)

ABSTRACTS:

Talk #1: We report on the user requirements study and preliminary implementation phases in creating a digital library that indexes and retrieves educational materials on math. We first review the current approaches and resources for math retrieval, then report on the interviews of a small group of potential users to properly ascertain their needs. While preliminary, the results suggest that meta-search and resource categorization are two basic requirements for a math search engine. In addition, we implement a prototype categorization system and show that the generic features work well in identifying the math contents from the webpage but perform less well at categorizing them. We discuss our long term goals, where we plan to investigate how math expressions and text search may be best integrated.

Talk #2: We consider the task of automatic slide image retrieval, in which slide images are ranked for relevance against a textual query. Our implemented system, SLIDIR caters specifically for this task using features specifically designed for synthetic images embedded within slide presentation. We show promising results in both the ranking and binary relevance task and analyze the contribution of different features in the task performance.

Talk #3: We describe a user-assisted framework for correcting ink-bleed in old handwritten documents housed at the National Archives of Singapore (NAS). Our approach departs from traditional correction techniques that strive for full automation. Fully automated approaches make assumptions about ink-bleed characteristics that are not valid for all inputs. Furthermore, fully-automated approaches often have to set algorithmic parameters that have no meaning for the end-user. In our system, the user needs only to provide simple examples of ink-bleed, foreground ink, and background. These training examples are used to classify the remaining pixels in the document to produce a computer generated result that is equal or better than existing fully-automated approaches.

To offer a complete system, we provide additional tools to allow any remaining errors to be easily cleaned up by the user. The initial training markup, computer-generated results, and manual edits are all recorded with the final output, allowing subsequent viewers to see how a corrected document was created and to make changes or updates. While an on-going project, our feedback from the NAS staff has been overwhelmingly positive that this user-assisted approach is a practical and useful way to address the ink-bleed problem.

Talk #4: The ACL Anthology is a digital archive of conference and journal papers in natural language processing and computational linguistics. Its primary purpose is to serve as a reference repository of research results, but we believe that it can also be an object of study and a platform for research in its own right. We describe an enriched and standardized reference corpus derived from the ACL Anthology that can be used for research in scholarly document processing. This corpus, which we call the ACL Anthology Reference Corpus (ACL ARC), brings together the recent activities of a number of research groups around the world. Our goal is to make the corpus widely available, and to encourage other researchers to use it as a standard testbed for experiments in both bibliographic and bibliometric research.

Talk #5: We describe ParsCit, a freely available, open-source implementation of a reference string parsing package. At the core of ParsCit is a trained conditional random field (CRF) model used to label the token sequences in the reference string. A heuristic model wraps this core with added functionality to identify reference strings from a plain text file, and to retrieve the citation contexts. The package comes with utilities to run it as a web service or as a standalone utility. We compare ParsCit on three distinct reference string datasets and show that it compares well with other previously published work.

AGENDA:

1. 2:00-3:30 Chan Yee Seng, "MAXSIM: A Maximum Similarity Metric for Machine Translation Evaluation"
2. 2:30-3:00 Chia Tee Kiah, "Lattice-Based Approach to Query-by-Example Spoken Document Retrieval"
3. 3:00-3:30 Tan Yee Fan, "Efficient Web-Based Linkage of Short to Long Forms"

ABSTRACTS:

Talk #1: We propose an automatic machine translation (MT) evaluation metric that calculates a similarity score (based on precision and recall) of a pair of sentences. Unlike most metrics, we compute a similarity score between items across the two sentences. We then find a maximum weight matching between the items such that each item in one sentence is mapped to at most one item in the other sentence. This general framework allows us to use arbitrary similarity functions between items, and to incorporate different information in our comparison, such as n-grams, dependency relations, etc. When evaluated on data from the ACL-07 MT workshop, our proposed metric achieves higher correlation with human judgements than all 11 automatic MT evaluation metrics that were evaluated during the workshop.

Talk #2: Recent efforts on the task of spoken document retrieval (SDR) have made use of speech lattices: speech lattices contain information about alternative speech transcription hypotheses other than the 1-best transcripts, and this information can improve retrieval accuracy by overcoming recognition errors present in the 1-best transcription. In this paper, we look at using lattices for the query-by-example spoken document retrieval task -- retrieving documents from a speech corpus, where the queries are themselves in the form of complete spoken documents (query exemplars). We extend a previously proposed method for SDR with short queries to the query-by-example task. Specifically, we use a retrieval method based on statistical modeling: we compute expected word counts from document and query lattices, estimate statistical models from these counts, and compute relevance scores as divergences between these models. Experimental results on a speech corpus of conversational English show that the use of statistics from lattices for both documents and query exemplars results in better retrieval accuracy than using only 1-best transcripts for either documents, or queries, or both. In addition, we investigate the effect of stop word removal which further improves retrieval accuracy. To our knowledge, our work is the first to have used a lattice-based approach to query-by-example spoken document retrieval.

Talk #3: Abbreviations, acronyms, initialisms, and shortenings frequently occurin many texts found on the Web, such as publication metadata, stock ticker codes, and biological articles. To connect these disparate forms together for knowledge discovery, short forms must be properly linked to their canonical long forms. In this paper, we demonstratehow a search engine can be efficiently utilized in mining the requiredcontextual information, so that short forms can be effectively linked to long forms. We show that a count-based method consistently outperforms other methods, and that using the snippets is better thanusing the full web pages. We also consider adaptively combining a query probing algorithm together with our count-based method. This reduces running time and network bandwidth, while maintaining the strong linkage performance.

ABSTRACT:

Hierarchical clustering of data is one of the most widely used machine learning techniques. Traditional hierarchical clustering techniques construct a single tree in a greedy fashion, either in a top-down or a bottom-up agglomerative fashion. Sometimes we are interested in how reliable the constructed tree is, i.e. how much we believe that the structure of the tree reflects true underlying structure in the data rather than spurious effects due to noise. Such a question can be answered using a Bayesian approach where we define a prior over trees and compute a posterior distribution over trees which captures the uncertainty in the learned tree structure.

However past Bayesian models for hierarchical clustering either do not give a posterior over trees (Heller and Ghahramani 2005, Friedman 2003), not infinitely exchangeable (Williams 2000), or is simply too complex to have widespread appeal (Neal 2003). In this talk we present a model that
1) gives a posterior distribution over trees,
2) is easy to implement, and
3) has the additional nice property that it is infinitely exchangeable.

Our model is based upon a standard model in population genetics called Kingman's coalescent. We propose both greedy and sequential Monte Carlo inference algorithms for the model. We show that our model performs well compared to previous approaches on a number of small datasets, and apply it to document clustering and phylolinguistics.

BIODATA:

Dr Teh Yee Whye is a lecturer at the Gatsby Computational Neuroscience Unit, University College London in the United Kingdom. Prior to this appointment he worked with Prof Lee Wee Sun as Lee Kuan Yew Postdoctoral Fellow at the National University of Singapore, and with Prof. Michael I. Jordan as a postdoc at University of California at Berkeley. He obtained his PhD from the University of Toronto under Prof. Geoffrey E. Hinton. His research interests are in Bayesian machine learning and probabilistic graphical models.

ABSTRACT:

With the explosion of the amount of textual data in the information age, natural language processing (NLP) has become increasingly important, with direct applications in areas such as Web mining and biomedical literature mining. Currently, the most effective approach to solving most NLP problems is supervised learning coupled with linguistic knowledge. However, standard supervised learning requires the training and the test corpora to be similar, and therefore falls apart in real NLP applications because obtaining labeled data for every new domain is expensive and thus infeasible. In this talk, I will present the major line of my PhD research on domain adaptation in NLP, which aims at adapting classifiers trained on one domain to another domain. We have proposed two frameworks to achieve domain adaptation, both having been evaluated on real NLP problems and outperformed standard learning methods. I will also briefly mention the future plan to incorporate knowledge bases and expert interactions into the domain adaptation process, with applications in large-scale information extraction from biomedical literature.

BIODATA:

Ms Jing Jiang is a final year PhD student in the Text Information Management Group in the Computer Science Department at the University of Illinois at Urbana-Champaign, working with Professor ChengXiang Zhai. Her research interests include natural language processing, information retrieval, machine learning, and biomedical literature mining. She received her B.S. degree and her M.S. degree in Computer Science from Stanford University in 2002 and 2003, respectively.

ABSTRACT:

I will describe how one useful aspect of the structure of scientific articles can be discovered with reasonably shallow means, namely the prototypical argumentation for the validity of the current research. Reference to other people's work, and reasonably standardised statements about this work, are a staple part of the argumentation, and citation analysis can exploit this fact. AZ-discourse analysis is the robust machine-learning of this structure, based on the extraction of correlated, and often linguistically interesting, features. I will show results of AZ on two domains (computational linguistics and chemistry), and discuss several search and summarisation applications using AZ. I will also speculate on more sequence-based methods for recognising AZ-type structures in text.

BIODATA:

Simone Teufel is a senior lecturer in the Computer laboratory at Cambridge University, where she has worked since 2001. Her main research interests are in corpus-linguistic approaches to discourse theory, and in the application of such information to summarisation, information retrieval and citation analysis. She has a background in computer science (1994 Diploma from University Stuttgart) and in cognitive science (2000 PhD from Edinburgh University), and has also experience in medical information processing and search, from a postdoctoral stay at Columbia University, and in collocation extraction, from a research post at Xerox Europe. Her lastest research interests include lexical acquisition, and the visualisation and language generation of the analysis results of scientific articles.

ABSTRACT:

While bioinformatics has far advanced in the past years and recognisers for gene and protein names and interactions have been built, biochemistry is a new field for computational linguistics to move into. I will be talking about the recognition strategy for scientific papers in general which the NLIP group at Cambridge University is developing, while concentrating on the research done in the project SciBorg, on chemical name parsing, ontology discovery, and discourse-related search. I will also talk a bit about the role of citations in this recognition effort, and about quite unusual infrastructure that our project is built on -- robust semantic representations, encoded as XML standoff.

BIODATA:

Simone Teufel is a senior lecturer in the Computer laboratory at Cambridge University, where she has worked since 2001. Her main research interests are in corpus-linguistic approaches to discourse theory, and in the application of such information to summarisation, information retrieval and citation analysis. She has a background in computer science (1994 Diploma from University Stuttgart) and in cognitive science (2000 PhD from Edinburgh University), and has also experience in medical information processing and search, from a postdoctoral stay at Columbia University, and in collocation extraction, from a research post at Xerox Europe. Her lastest research interests include lexical acquisition, and the visualisation and language generation of the analysis results of scientific articles.

ABSTRACT:

Coreference resolution is the task of finding different mentions of the same entity in the word. In the past decade, knowledge-lean approaches are widely adopted, in which only simple morpho-syntactic cues as knowledge sources are employed in the resolution process. Although these approaches have achieved reasonable success, researchers have found that deeper syntactic or semantic knowledge is necessary in order to reach the next level performance. In this talk, we will introduce our knowledge-rich approaches to coreference resolution, including a tree-kernel-based method for the syntactic knowledge, and web-based methods for the semantic knowledge. These sources of enriched knowledge are acquired automatically without too many human efforts, and have proved effective for the coreference resolution task.

ABSTRACT:

A Scenario Template is a data structure that reflects the salient aspects shared by a set of similar events, which are considered as belonging to the same scenario. These salient aspects are typically the scenario's characteristic actions, the entities involved in these actions and the related attributes of them.

In this talk, I will first brief about our approach to scenario template creation and update the latest evaluation results. Then I will discuss one of the possible applications of scenario templates, namely, open-domain question and answering. For Q&A systems, query expansion is a common strategy while sentence selection is an important process. I will show how scenario templates might help in these two aspects.

BIODATA:

Qiu Long is a Doctoral Student at SoC, NUS, co-supervised by Professor Chua Tat-Seng and Dr. Min-Yen Kan. He got his Master of Science (SM) in CS from Singapore - MIT Alliance. He is interested in Natural Language Processing (NLP) and the related machine learning techniques.

ABSTRACT:

In recent years, speech processing products had been widely distributed all over the world, reflecting a general believe that speech technologies have a huge potential to overcome language barriers and to let everyone participate in today's information revolution. However, in spite of vast improvements in speech and language technologies, the development of speech processing systems still requires significant skills and resources to carry out. Consequently, with more than 6500 languages in the world, the current costs and effort in building speech support is prohibitive to all but the most economically viable languages.

In this talk I will discuss the challenges and limitations of rapidly developing automatic speech processing systems for a large number of languages and dialects. I will describe solutions to system development based on sharing data and system components across languages. Practical implementations and recent results are presented in the light of our SPICE project, which aims to bridge the gap between language and technology experts by providing innovative strategies and tools for non-expert users. These tools enable the user to easily collect appropriate text and speech data, to quickly develop acoustic models, pronunciation dictionaries, and language models based on very limited resources, and to monitor progress and performance allowing for iterative improvements with the user in the loop.

BIODATA:

Tanja Schultz received her Ph.D. and Masters in Computer Science from University Karlsruhe, Germany in 2000 and 1995 respectively and got a German Masters in Mathematics, Sports, and Education Science from the University of Heidelberg, Germany in 1990. She joined Carnegie Mellon University in 2000 and is a faculty member of the Language Technologies Institute as a Research Computer Scientist. Since 2007 she also holds a full professorship at Karlsruhe University, Germany.

Her research activities center around language independent and language adaptive speech recognition but also include large vocabulary continuous speech recognition systems, human-machine interfaces using speech and various biosignals, speech translation, as well as language and speaker identification approaches. With a particular area of expertise in multilingual approaches, she performs research on portability of speech processing systems to many different languages. In 2001 Tanja Schultz was awarded with the FZI price for her outstanding Ph.D. thesis on language independent and language adaptive speech recognition. In 2002 she received the Allen Newell Medal for Research Excellence from Carnegie Mellon for her contribution to Speech-to-Speech Translation and the ISCA best paper award for her publication on language independent acoustic modeling. In 2005 she was awarded the Carnegie Mellon Language Technologies Institute Junior Faculty Chair. Tanja Schultz is the author of more than 100 articles published in books, journals, and proceedings.

She is a member of the IEEE Computer Society, the European Language Resource Association, the Society of Computer Science (GI) in Germany, and currently serves on the ISCA board and several program and review panels.

Abstract:

The idea of using humans to teach computers is not a new one, but it has been largely impractical and largely ignored. Modern-day computers tend to "learn" by either sifting through large amounts of data or by being programmed/endowed with expert knowledge. Typically there is little interaction between man and machine. Our recent project, called "Wubble World", capitalizes on the availability of free hands-on human teaching as a means for machine learning of language and concepts.

The basic premise of this work begins with an online game situated in a virtual 3D environment. Language is generated as children interact with their personal creature, called a wubble, or with other children. By virtue of the virtual environment, this language is situated and forms a rich corpus of matched scenes and sentences upon which to learn language and concepts. In one part of the environment, children interact with their wubble by teaching it to accomplish certain given tasks. The wubble, like a toddler, initially knows little about the world, and must acquire concepts and labels by interacting with the child. I'll describe this environment and the basic concept learning that happens inside the wubble. In another part of the world, children play a competitive team game against other children. The game is designed to require cooperation among team members, typically using spoken language. This language, combined with a log of the game state, generates a rich sentence-scene corpus. This richness could potentially enable natural language processing to move beyond current statistical techniques by incorporating data that reveals underlying meaning. I'll demonstrate the game, describe the data we have collected so far, and discuss some of the possible approaches for learning from this data.

BIODATA:

Dr. Yu-Han Chang is a Computer Scientist at the Information Sciences Institute of the University of Southern California (USC ISI). His current research interests span topics from reinforcement learning, game theory, natural language understanding, interactive technologies, and traditional AI. Recent and ongoing projects include harnessing the power of the Internet to train intelligent agents via human teaching, transfer learning, and the development of efficient no-regret algorithms for non-cooperative learning domains. Dr. Chang holds undergraduate degrees in Mathematics and Economics, as well as a S.M. in Computer Science, from Harvard University. He received his Ph.D. in Electrical Engineering and Computer Science from MIT, focusing his efforts on developing algorithms for multi-agent learning in the context of machine learning and game theory.

Abstract:

The task of referring expression generation is concerned with determining what semantic content should be used in a reference to an intended referent so that the hearer will be able to identify that referent. The task has been a focus of interest within natural language generation at least since the early 1980s, in part because the problem appears relatively well-defined. Over the last 25 years, a range of algorithms and approaches have been proposed and explored, making this the most intensely studied problem in natural language generation; and yet, even a casual analysis of real human-authored texts suggests that we have a long way to go in terms of providing an explanation for the range of real linguistic behaviour that we find. In this talk, I'll review research in the area to date, try to characterise where we are now, and point to directions for future research in the area.

BIODATA:

Robert Dale received his PhD in Computational Linguistics from the University of Edinburgh in 1989. His research interests include low-cost approaches to intelligent text processing tasks; practical natural language generation; the engineering of habitable spoken language dialog systems; and computational, philosophical and linguistic issues in reference and anaphora. He is Director of the Centre for Language Technology at Macquarie University, Convenor of the Australian Research Council's Human Communication Science Network, and editor-in-chief of the Journal of Computational Linguistics.

Abstract:

This work deals with the problem of event annotation in social networks. The problem is made difficult due to variability of semantics and due to scarcity of labeled data. Events refer to real-world phenomena that occur at a specific time and place, and media and text tags are treated as facets of the event metadata. We are proposing a novel mechanism for event annotation by leveraging related sources (other annotators) in a social network. Our approach exploits event concept similarity, concept co-occurrence and annotator trust. We compute concept similarity measures across all facets. These measures are then used to compute event-event and user-user activity correlation. We compute inter-facet concept co-occurrence statistics from the annotations by each user. The annotator trust is determined by first requesting the trusted annotators (seeds) from each user and then propagating the trust amongst the social network using the biased PageRank algorithm. For a specific media instance to be annotated, we start the process from an initial query vector and the optimal recommendations are determined by using a coupling strategy between the global similarity matrix, and the trust weighted global co-occurrence matrix. The coupling links the common shared knowledge (similarity between concepts) that exists within the social network with trusted and personalized observations (concept co-occurrences). Our initial experiments on annotated everyday events are promising and show substantial gains against traditional SVM based techniques.

Co-authors: Amit Zunjarwad (AME), Lexing Xie (IBM)

BIODATA:

Hari Sundaram is currently an assistant professor, at Arizona State University. This is a joint appointment with the department of Computer science and the Arts Media and Engineering program. He received his Ph.D. from the Department of Electrical Engineering at Columbia University in 2002. He received his MS degree in Electrical Engineering from SUNY Stony Brook 1995 and a B.Tech in Electrical Engineering from Indian Institute of Technology, Delhi in 1993.

Jointly organized by CHIME, I2R and PREMIA.

This talk will examine some assumptions in media semantics under three broad categories - (a) aspects of meaning (b) rethinking semantic construction (c) learnability contradictions. A re-examination of the assumptions behind media semantics is useful, as the mechanisms by which people create and consume media have changed significantly in the last decade. These changes offer fresh insight into the familiar problem of the semantic gap - how to go from sensory data to meaning. There are three aspects of meaning of interest - context, approximations, and variability. We need to examine the construction of meaning in a manner very different from the familiar Marr model - specifically we shall examine embodiment and networked construction. A significant challenge to the learnability of semantics lies in re-examining within the multimedia context, of what Chomsky calls "the poverty of input" problem. How is it possible to learn a large number of concepts with very few / or even non-existent training examples? We will examine the role of context and semantic approximation with an application to media retrieval. The issues of embodiment and its relation to semantics will be discussed with respect to an educational application. We hope to provide a partial answer to the issue of semantic construction and learnability in an application related to social networks.

BIODATA:

Hari Sundaram is currently an assistant professor, at Arizona State University. This is a joint appointment with the department of Computer science and the Arts Media and Engineering program. He received his Ph.D. from the Department of Electrical Engineering at Columbia University in 2002. He received his MS degree in Electrical Engineering from SUNY Stony Brook 1995 and a B.Tech in Electrical Engineering from Indian Institute of Technology, Delhi in 1993.

His research group works on developing computational models and systems for situated communication. There are two complementary (but coupled) directions - (a) designing intelligent multimedia environments that exist as part of our physical world (e.g. an intelligent room) (b) developing new algorithms and systems to understand the media artifacts resulting from human activity (e.g. emails, photos / video). Specific projects include - context models for action, resource adaptation, interaction architectures, communication patterns in media sharing social networks, collaborative annotation, as well analysis of online communities.

Prof. Sundaram's research has won several awards - the best student paper award at JCDL 2007, the best ACM Multimedia demo award in 2006. The best student paper award at ACM conference on Multimedia 2002, the 2002 Eliahu I. Jury Award for best Ph.D. dissertation. He has also received a best paper award on video retrieval, from IEEE Trans. On Circuits and Systems for Video Technology, for the year 1998. He is an active participant in the Multimedia community - he is an associate editor for ACM Transactions on Multimedia Computing, Communications and Applications (TOMCCAP), as well as the IEEE Signal Processing magazine. He has co-organized workshops at acm multimedia on experiential telepresence (ETP 2003, ETP 2004), archival of personal experiences (CARPE 2004, CARPE 2005) and a conference of image and video retrieval (CIVR 2006).

The problem of domain adaptation for statistical classifiers arises when our labeled training examples and unlabeled test examples come from different domains. This problem is commonly encountered in natural language processing (NLP) tasks. For example, we may train a named entity recognition (NER) system on news articles but apply the system to blog or email text. It is generally observed that the performance of a classifier tends to drop significantly when it is applied to a different domain.

In this talk, I will present our recent work addressing the domain adaptation problem. We have proposed two frameworks, corresponding to two different perspectives on this problem: feature selection and instance weighting. In the feature selection framework, we seek to identify .generalizable features. that behave similarly across domains; in the instance weighting framework, our idea is to re-weight the examples in order to minimize the expected loss on the test domain. In both frameworks, we have also incorporated semi-supervised learning to make use of the unlabeled test domain examples. Experiment results on a number of NLP tasks, including NER, part-of-speech (POS) tagging, and spam filtering, show the effectiveness of both frameworks. At the end of the talk, I will briefly mention our current effort of unifying the two perspectives, as well as some future directions to pursue.

BIODATA: Jing Jiang is a Ph.D. candidate in the Department of Computer Science at the University of Illinois at Urbana-Champaign. She is a member of the Information Retrieval Group led by Professor ChengXiang Zhai. Her research interests include information extraction, information retrieval, biomedical text mining, and machine learning. She received her B.S. degree and her M.S. degree in Computer Science from Stanford University in 2002 and 2003, respectively.

1. 1:00-1:25 Chia Tee Kiah, "A Statistical Language Modeling Approach to Lattice-Based Spoken Document Retrieval"
2. 1:25-1:50 Zhao Shanheng, "Identification and Resolution of Chinese Zero Pronouns: A Machine Learning Approach"
3. 1:50-2:15 Hendra Setiawan, "Ordering Phrases with Function Words"
4. 2:15-2:30 15 minute break
5. 2:30-2:55 Dave Kor, TBA
6. 2:55-3:20 Yang Xiaofeng, "Coreference Resolution Using Semantic Relatedness Information from Automatically Discovered Patterns"
7. 3:20-3:35 Tan Yee Fan, "PSNUS: Web People Name Disambiguation by Simple Clustering with Rich Features"

ABSTRACTS:

Title: A Statistical Language Modeling Approach to Lattice-Based Spoken Document Retrieval

Abstract: Speech recognition transcripts are far from perfect; they are not of sufficient quality to be useful on their own for spoken document retrieval. This is especially the case for conversational speech. Recent efforts have tried to overcome this issue by using statistics from speech lattices instead of only the 1-best transcripts; however, these efforts have invariably used the classical vector space retrieval model. This paper presents a novel approach to lattice-based spoken document retrieval using statistical language models: a statistical model is estimated for each document, and probabilities derived from the document models are directly used to measure relevance. Experimental results show that the lattice-based language modeling method outperforms both the language modeling retrieval method using only the 1-best transcripts, as well as a recently proposed lattice-based vector space retrieval method.

Title: Identification and Resolution of Chinese Zero Pronouns: A Machine Learning Approach

Abstract: In this paper, we present a machine learning approach to the identification and resolution of Chinese anaphoric zero pronouns. We perform both identification and resolution automatically, with two sets of easily computable features. Experimental results show that our proposed learning approach achieves anaphoric zero pronoun resolution accuracy comparable to a previous state-of-the-art, heuristic rule-based approach. To our knowledge, our work is the first to perform both identification and resolution of Chinese anaphoric zero pronouns using a machine learning approach.

Title: Ordering Phrases with Function Words

Abstract: Function words are a class of words with little intrinsic meaning but is vital in expressing grammatical relationships among phrases within a sentence. Such encoded grammatical information, often implicit, makes function words pivotal in modeling structural divergences, as projecting them in different languages often result in long-range structural changes to the realized sentences. This distinctive feature has not been fully-utilized to address phrase ordering problem in the context of statistical machine translation (SMT). We observe that just like foreign language learner often makes mistakes in using function words, current SMT system often perform poorly in ordering function words' arguments; lexically correct translations often end up reordered incorrectly. In this talk, I will present a Function Words centered, Syntax-based (FWS) solution to address the phrase ordering problem, including its statistical formalism, its implementation and experimental results.

Title: Coreference Resolution Using Semantic Relatedness Information from Automatically Discovered Patterns

Abstract: Semantic relatedness is a very important factor for the coreference resolution task. To obtain this semantic information, corpus-based approaches commonly leverage patterns that can express a specific semantic relation. The patterns, however, are designed manually and thus are not necessarily the most effective ones in terms of accuracy and breadth. To deal with this problem, in this paper we propose an approach that can automatically find the effective patterns for coreference resolution. We explore how to automatically discover and evaluate patterns, and how to exploit the patterns to obtain the semantic relatedness information. The evaluation on ACE data set shows that the pattern based semantic information is helpful for coreference resolution.

Title: PSNUS: Web People Name Disambiguation by Simple Clustering with Rich Features

Abstract: We describe about the system description of the PSNUS team for the SemEval-2007 Web People Search Task. The system is based on the clustering of the web pages by using a variety of features extracted and generated from the data provided. This system achieves F_alpha=0.5 = 0.75 and F_alpha=0.2 = 0.78 for the final test data set of the task.

AGENDA:

1. 1:00-1:25 Chan Yee Seng, "Word Sense Disambiguation Improves Statistical Machine Translation"
2. 1:25-1:50 Chan Yee Seng, "Domain Adaptation with Active Learning for Word Sense Disambiguation"
3. 1:50-2:15 Li Haizhou, "Semantic Transliteration of Personal Names"
4. 2:15-2:30 15 minute break - refreshments to be served.
5. 2:30-2:55 Min Zhang, "A Grammar-driven Convolution Tree Kernel for Semantic Role Classification"
6. 2:55-3:20 Mstislav Maslennikov, "ARE&D: A Discourse-based Multi-resolution Framework for Information Extraction on Free Text"

ABSTRACTS:

Title: Word Sense Disambiguation Improves Statistical Machine Translation

Abstract: Recent research presents conflicting evidence on whether word sense disambiguation (WSD) systems can help to improve the performance of statistical machine translation (MT) systems. In this paper, we successfully integrate a state-of-the-art WSD system into a state-of-the-art hierarchical phrase-based MT system, Hiero. We show for the first time that integrating a WSD system improves the performance of a state-of-the-art statistical MT system on an actual translation task. Furthermore, the improvement is statistically significant.

Title: Domain Adaptation with Active Learning for Word Sense Disambiguation

Abstract: When a word sense disambiguation (WSD) system is trained on one domain but applied to a different domain, a drop in accuracy is frequently observed. This highlights the importance of domain adaptation for word sense disambiguation. In this paper, we first show that an active learning approach can be successfully used to perform domain adaptation of WSD systems. Then, by using the predominant sense predicted by expectation-maximization (EM) and adopting a count-merging technique, we improve the effectiveness of the original adaptation process achieved by the basic active learning approach.

Title: Semantic Transliteration of Personal Names

Abstract: Words of foreign origin are referred to as borrowed words or loanwords. A loanword is usually imported to Chinese by phonetic transliteration if a translation is not easily available. Semantic transliteration is seen as a good tradition in introducing foreign words to Chinese. Not only does it preserve how a word sounds in the source language, it also carries forward the word's original semantic attributes. This paper attempts to automate the semantic transliteration process for the first time. We conduct an inquiry into the feasibility of semantic transliteration and propose a probabilistic model for transliterating personal names in Latin script into Chinese. The results show that semantic transliteration substantially and consistently improves accuracy over phonetic transliteration in all the experiments.

Title: A Grammar-driven Convolution Tree Kernel for Semantic Role Classification

Abstract: Convolution tree kernel has shown very promising results in semantic role classification. However, this method considers less linguistic knowledge and only carries out hard matching between substructures, which may lead to over-fitting and less accurate similarity measure. To remove the constraints, this paper proposes a grammar-driven convolution tree kernel for semantic role classification by introducing more linguistic grammar information into the standard convolution tree kernel. The proposed grammar-driven convolution tree kernel displays two advantages over the previous one: 1) grammar-driven approximate substructure matching and 2) grammar-driven approximate tree node matching. The two improvements enable the proposed grammar-driven tree kernel explore more linguistically motivated substructure features than the previous one. Experiments on the CoNLL-2005 SRL shared task show that the proposed grammar-driven tree kernel significantly outperforms the previous non-grammar-driven one in semantic role classification. Moreover, we present a composite kernel to integrate feature-based and tree kernel-based methods. Experimental results show that the composite kernel outperforms the previous best-reported methods.

Title: ARE&D: A Discourse-based Multi-resolution Framework for Information Extraction on Free Text

Abstract: Extraction of relations between entities is an important part of Information Extraction on free text. Previous methods are mostly based on statistical correlation and dependency relations between entities. This paper re-examines the problem at the multi-resolution layers of phrase, clause and sentence using dependency and discourse relations. Our multi-resolution framework ARE&D (Anchor and Relation and Discourse analysis) uses clausal relations in 2 ways: 1) to filter noisy dependency paths; and 2) to increase reliability of dependency path extraction. The resulting system outperforms the previous approaches by 3%, 7%, 4% on MUC4, MUC6 and ACE RDC domains respectively.

In this talk, I will present a Function Words centered, Syntax-based (FWS) solution to address the phrase ordering problem, including its statistical formalism, its implementation and experimental results.

2nd talk: In this talk, I will present a new graph-based approach to text understanding and summarization. Current graph-based approaches to automatic text summarization, such as LexRank and TextRank, assume a static graph which does not model how the input texts emerge. A suitable evolutionary text graph model may impart a better understanding of the texts and improve the summarization process. We give simplified assumptions of human writing and reading processes, and then propose a timestamped graph (TSG) model that is motivated by these processes and show how text units in this model emerge over time. This model not only captures the evolving process of text within a document, but also the evolving process across documents. In our model, the graphs used by LexRank and TextRank are specific instances of our timestamped graph with particular parameter settings.

We have investigated several widely-used unsupervised and supervised term weighting methods on several popular data collections in combination with SVM and kNN algorithms. In consideration of the distribution of relevant documents in the collection and analysis of the term's discriminating power, we have proposed a new term weighting scheme, namely tf.rf. The controlled experimental results showed that the term weighting methods show mixed performance in terms of different category distribution data sets and different learning algorithms. Most of the supervised term weighting methods which are based on information theory have not shown satisfactory performance according to our experimental results. However, the newly proposed tf.rf method shows a consistently better performance than other term weighting methods. On the other hand, the popularly used tf.idf method has not shown a uniformly good performance with respect to different category distribution data sets.

Manually defining scenario template is expensive and we aim to automatize this template generation process. We argue that context is valuable to identify semantically similar text spans, from which template slots could be generalized. To leverage context, we convert news articles into a graphical representation and then apply a generic context-sensitive clustering (CSC) framework to get meaningful clusters of text spans by examining the intrinsic and extrinsic similarities between them. We use the Expectation-Maximization algorithm to guide the clustering process. The experiments show that: 1) our approach generates high quality clusters, and 2) information extracted from the clusters is adequate to build high coverage templates.

The goal of our research is to reduce the manual effort and automate the process of relation extraction. To realize this intention, we investigate semi-supervised learning and unsupervised learning solutions to rival supervised learning methods to resolve the problem of relation extraction with minimal human cost and still achieve comparable performance to supervised learning methods.

First, we presented a Label Propagation (LP) based semi-supervised learning algorithm for relation extraction problem to learn from both labeled and unlabeled data. It represents labeled and unlabeled examples and their distances as the nodes and the weights of edges of a graph, then propagating the label information from any vertex to nearby vertices through weighted edges iteratively, finally inferring the labels of unlabeled examples after the propagation process converges.

Secondly, we introduced an unsupervised learning algorithm based on model order identification for automatic relation extraction. The model order identification is achieved by resampling based stability analysis and used to infer the number of relation types between entity pairs automatically.

Thirdly, we further investigated unsupervised learning solution for relation disambiguation using graph based strategy. We defined the unsupervised relation disambiguation task for entity mention pairs as a partition of a graph so that entity pairs that are more similar to each other, belong to the same cluster. We apply spectral clustering to resolve the problem, which is a relaxation of such NP-hard discrete graph partitioning problem. It works by calculating eigenvectors of an adjacency graph's Laplacian to recover a submanifold of data from a high dimensionality space and then performing cluster number estimation on such spectral information.

The thesis evaluates the proposed methods for extracting relations among named entities automatically, using the ACE corpus. The experimental results indicate that our methods can overcome the problem of not having enough manually labeled relation instances for supervised relation extraction methods. The results show that when only a few labeled examples are available, our LP based relation extraction can achieve better performance than SVM and another bootstrapping method. Moreover, our unsupervised approaches can achieve order identification capabilities and outperform the previous unsupervised methods. The results also suggest that all of the four categories of lexical and syntactic features used in the study are useful for the relation extraction task.

and

2nd talk: Name origin recognition is to identify the original source of a name. It is a necessary step for name translation/transliteration because of different origins need different translation strategies. It is more important when translating across languages with different alphabets and sound inventories. Previous works used rule based methods or statistics based methods to solve this problem. In this work, we cast name origin recognition as a multi-class classification task and propose to use Maximum Entropy model to solve it. Experiments show that our approach can achieve an overall accuracy 98.35% for name written in English and 98.10% for name written in Chinese, which are much better than previous methods.

This talk borrows materials from my recent work in IQIS'05 JCDL'06, ICDM'06, and ICDE'07, that are the results of joint work with several students and collaborators:

Ergin Elmacioglu (Penn State), Min-Yen Kan (NUS), Jaewoo Kang (Korea U.), Nick Koudas (U. Toronto), Byung-Won On (Penn State), Jian Pei (Simon Fraser U.), Divesh Srivastava (AT&T Labs -- Research), Yee Fan Tan (NUS)

This talk is part of a tutorial given with Kevin Knight at ACL 2006

As far as extrinsic methods go, leading QA systems employ a variety of means to boost accuracy. Such methods include redundancy (getting the same answer from multiple documents/sources), inferencing (proving the answer from information in texts plus background knowledge) and sanity-checking (verifying that answers are consistent with known facts). To our knowledge, however, no other QA system deliberately asks additional questions in order to derive constraints on the answers to the original questions. We present two variations on this idea. The first is the method of QA-by-Dossier-with-Constraints (QDC), which is an extension of the simpler method of QA-by-Dossier, in which definitional questions ("Who/what is X?") are addressed by asking a set of questions about anticipated properties of X. In QDC, the collection of Dossier candidate answers is subjected to satisfying a set of naturally-arising constraints. For example, for a "Who is X?" question, the system will ask about birth, accomplishment and death dates, which if they exist, must occur in that order, and also obey other constraints such as lifespan. Temporal, spatial and kinship relationships seem to be parti