Believe It or Not: Solving the TAC 2009 Textual Entailment Tasks through an Artificial Believer System


The Text Analysis Conference (TAC) 2009 competition featured a new textual entailment search task, which extends the 2008 textual entailment task. The goal is to find information in a set of documents that are entailed from a given statement. Rather than designing a system specifically for this task, we investigated the adaptation of an existing artificial believer system to solve this task. The results show that this is indeed possible, and furthermore allows to recast the existing, divergent tasks of textual entailment and automatic summarization under a common umbrella.

A Belief Revision Approach to Textual Entailment Recognition


An artificial believer has to recognize textual entailment to categorize beliefs. We describe our system – the Fuzzy Believer system – and its application to the TAC/RTE three-way task.

ERSS at TAC 2008


An Automatically Generated SummaryAn Automatically Generated Summary
ERSS 2008 attempted to rectify certain issues of ERSS 2007. The improvements to readability, however, do not re?ect in signi?cant score increases, and in fact the system fell in overall ranking. While we have not concluded our analysis, we present some preliminary observations here.

Generating Update Summaries for DUC 2007


Update summaries as defined for the new DUC 2007 task deliver focused information to a user who has already read a set of older documents covering the same topic. In this paper, we show how to generate this kind of summary from the same data structure—fuzzy coreference cluster graphs—as all other generic and focused multi-document summaries. Our system ERSS 2007 implementing this algorithm also participated in the DUC 2007 main task, without any changes from the 2006 version.

An Initial Fuzzy Coreference Cluster Graph

Fuzzy Clustering for Topic Analysis and Summarization of Document Collections

Montreal 2007


Large document collections, such as those delivered by Internet search engines, are difficult and time-consuming for users to read and analyse. The detection of common and distinctive topics within a document set, together with the generation of multi-document summaries, can greatly ease the burden of information management. We show how this can be achieved with a clustering algorithm based on fuzzy set theory, which (i) is easy to implement and integrate into a personal information system, (ii) generates a highly flexible data structure for topic analysis and summarization, and (iii) also delivers excellent performance.

Next-Generation Summarization: Contrastive, Focused, and Update Summaries

Conference Hotel, Borovets, Bulgaria


Classical multi-document summaries focus on the common topics of a document set and omit distinctive themes particular to a single document—thereby often suppressing precisely that kind of information a user might need for a specific task. This can be avoided through advanced multi-document summaries that take a user's context and history into account, by delivering focused, contrastive, or update summaries. To facilitate the generation of these different summaries, we propose to generate all types from a single data structure, topic clusters, which provide for an abstract representation of a set of documents. Evaluations carried out on five years' worth of data from the DUC summarization competition prove the feasibility of this approach.

Fuzzy Coreference Resolution for Summarization



We present a fuzzy-theory based approach to coreference resolution and its application to text summarization.

Automatic determination of coreference between noun phrases is fraught with uncertainty. We show how fuzzy sets can be used to design a new coreference algorithm which captures this uncertainty in an explicit way and allows us to define varying degrees of coreference.

The algorithm is evaluated within a system that participated in the 10-word summary task of the DUC 2003 competition.

Using Knowledge-poor Coreference Resolution for Text Summarization


We present a system that produces 10-word summaries based on the single summarization strategy of outputting noun phrases representing the most important text entities (as represented by noun phrase coreference chains). The coreference chains were computed using fuzzy set theory combined with knowledge-poor corefernce heuristics.

Multi-ERSS and ERSS 2004


Last year, we presented a system, ERSS, which constructed 10 word summaries in form of a list of noun phrases. It was based on a knowledge-poor extraction of noun phrase coreference chains implemented on a fuzzy set theoretic base. This year we present the performance of an improved version, ERSS 2004 and an extension of the same basic system: Multi-ERSS constructs 100-word extract summaries for clusters of texts. With very few modifications we ran ERSS 2004 on Tasks 1 and 3 and Multi-ERSS on Tasks 2, 4, and 5, scoring generally above average in all but the linguistic quality aspects.

ERSS 2005: Coreference-Based Summarization Reloaded


Friendly Meetings in Vancouver
We present ERSS 2005, our entry to this year's DUC competition. With only slight modifications from last year's version to accommodate the more complex context information present in DUC 2005, we achieved a similar performance to last year's entry, ranking roughly in the upper third when examining the ROUGE-1 and Basic Element score.

We also participated in the additional manual evaluation based on the new Pyramid method and performed further evaluations based on the Basic Elements method and the automatic generation of Pyramids. Interestingly, the ranking of our system differs greatly between the different measures; we attempt to analyse this effect based on correlations between the different results using the Spearman coefficient.

Syndicate content