Research And Implementation Of Entailment-Based Text Exploration

Finding information in a large body of text is becoming increasingly more difficult. Standard search engines output a set of documents for a given query, but do not allow any exploration of the thematic structure in the retrieved information. Thus the need to effectively sift through a target set of documents is becoming ever more important.Textual entailment is a semantic relation between text T and hypothesis H:if the meaning of H, as interpreted in the context of T, can be inferred from the meaning of T, we say T entails H. Textual entailment recognition is a generic task that captures major semantic inference needs across many natural language processing applications, such as Question Answering, Information Retrieval, Information Extraction, and multi-document Summarization. Therefore, text entailment learning is important.This paper conducts in-depth research on recognizing and acquiring the text entailment relations, and completes entailment rules learning. Text search system combined with these rules can explore the result space, by drilling down/up from one proposition to another, according to a set of entailment relations.The research and implementation of entailment-based text exploration combine the search engines with the entailment rules, allowing users to select the appropriate search results by entailment relations.The contributions of this paper are as follows:1. This paper proposes a method of learning entailment relations based on word vectors. This method employs the word vectors of predicates and learns entailment relations by calculating the semantic similarity between the vectors. The experiment results suggest that this method could learn entailment relations effectively. The mean average precision (MAP) of this method reaches60.74%, nearly5percentage points higher than the BInc one.2. This paper also combines this word vector based method with the previous word based methods and proposes a united method to learn entailment relations based on both word and word vector. The experiment results indicate that the F value of this method was increased from30.67%to34.49%, comparing with that of the BInc method.3. This paper implements an entailment-based text retrieval system working on news corpus by combining textual entailment with text retrieval. This system is an expansion of the traditional faceted search. It could reflect the practicability of the textual entailment and improve the usability of the system, because one could search the result space further according to entailment relations.