Answer Extraction Based On Logic Representation And Reasoning For Reading Comprehension

As an important research area in Question Answering, reading comprehension (RC) can analyze a given passage and generate/extract the corresponding answers from the passage. Knowledge representation and reasoning in natural language understanding can formalize the knowledge of context and make automatic reasoning easier. This paper applied knowledge representation and reasoning to RC system. The first step is to convert questions of reading comprehension to appropriate logical questions, and the second step is to extract the answer to the question from the proof generated by logical inference.This paper uses text predicate representation as the knowledge representation of natural language text. Text predicate representation is a first order predicate logic. In this method, different words have different definitions of logical predicates, and the position of their parameters indicates the syntax and semantic information. In order to automatically generate logic representation of text, this paper builds a logic representation transformation system based on the results of syntax parsing. What's more, the automatic reasoning system of text predicate representation is also implemented based on logic prover tools Prover9.Firstly, passage and questions of reading comprehension are sent to logic representation transformation system to generate their corresponding logic representations. Question logic representation needs further processing to maintain the semantic information of original question. Then, logic representations and their related inference knowledge are sent to automatic reasoning system to get corresponding proof of logical reasoning procedure.Knowledge base plays a vital role in the process of logical reasoning. This paper adopts XWN knowledge base and uses some inference rules as complementary rules. Experiments show that these complementary rules raise the accuracy of logic reasoning.Finally, answer sentences are extracted from logic proof by bag-of-words (BOW) method or extracted directly from BOW method if proof isn't given. This approach achieves 39.0% HumSent accuracy on the Remedia corpora. In another approach, answer sentences generated from Multi-answer BOW method are re-ranked using logical reasoning system. This approach achieves 40.0% HumSent accuracy on the Remedia corpora.