Combining existing methods for an applied textual entailment search in the christian bible


Kenston Choi

Date of Publication


Document Type

Master's Thesis

Degree Name

Master of Science in Computer Science


College of Computer Studies


Computer Science

Thesis Adviser

Nathalie Rose Lim-Cheng

Defense Panel Chair

Allan Borra

Defense Panel Member

Shirley Chu
Nathalie Rose Lim-Cheng


Since the first RTE Challenge, various approaches in recognizing textual entailment have been implemented and evaluated. However, there has been limited exploration in combining scored decisions from independently developed systems. In this work, class labels and features from three existing entailment systems sharing the same word alignment were fused and compared. Additional features improved an existing word aligner by at least 3.3% in Fmeasure. Various ensemblemethods combining lexical, syntactic and semantic features along with outputs from three entailment systems were shown to improve over the performance of the best individual system in the ensemble across the datasets from the first five RTE Challenges. The highest micro-average accuracy improvement of 2.6% was achieved using mean combiner within a cascading classifier. Applying this entailment combination method to the search task for the Biblical domain involving 200 test queries, an improvement of 6% in F-measure was obtained when combining entailment-decidedmonolingual parallel texts versus pairing only with a single source text. Deciding parallel texts using a non-ensemble method decreased F-measure by less than 1% but with a different precision-recall tradeoff. Using entailment as criterion for search relevance improved themean reciprocal rank by 18% however, filtering search results by entailment did not result to an improvement due to lack of recall.

Abstract Format






Accession Number


Shelf Location

Archives, The Learning Commons, 12F Henry Sy Sr. Hall

Physical Description

1 computer optical disc ; 4 3/4 in.

This document is currently not available here.