Recognizing Textual Entailment with synthetic analysis based on SVM and feature value control

Abstract

Recognizing Textual Entailment, as one of the branches of Nature Language Processing, has been widely used in Human Computer Interaction, Question Answering System, etc. RTE is trying to build an intelligent system which can analyze the content of an input text (T), and then raises a hypothesis (H) based on that. My self-design RTE system which is called SNRTE combines lexical, syntax, and semantic 3 levels of analysis, under the support of Stemmer, Tokenizer, Parser, POS Tag, Name Finder, WordNet2.1, and Support Vector Machine, etc. All these modules fetch useful information elements in the target text to define 49 feature values, which finally adopted into SVM to make judgments. The training data is taken from RTE official contest including 1600 pairs of test and hypothesis P(T, H). The average correct judgment rates are 67.5%, far above the average system correctness in RTE1 contest (55.12%) and better than the 2nd system (60.6%).