Machine Reading With Semantic Inference And Representation

Question answering(Q&A)is one of the research focuses in NLP.The research results of this task can reflect the level of machine understanding of nature language texts to some extent.Especially,open-domain question answering system is more challenge because it requires the machine has the ability to understand and answer questions from various domains.This paper is the research of open-domain Q&A based on inference and representation of semantic,aiming to improve the performance of open-domain Q&A system further.Firstly,the traditional lexical matching method of Q&A chooses the answer by computing the similarity between question-answer pair and the whole document.It lacks the fine-grained inference process.For example,because only a few sentences in the document are related to the question,using the above matching method will introduce many unrelated sentences to cause interference in reasoning.Compared with the traditional lexical matching method,the hidden variable model can capture the potential information of the text better.It finds the sentences related to question in the document first,and then conducts the answer reasoning on them,and a good hidden variable structure can discover the potential relationship between the question and the document better.This paper proposes two novel hidden variable models based on parse trees and semantic frames to improve the deep semantic reasoning ability of machine for the text,and experiments on MCTest dataset show that the proposed models are highly competitive with state-of-the-art machine comprehension systems.Secondly,distributed word representation is an important representation of word semantics.Compared with one-hot word representation,it can measure the similarity between words by similarity between vectors,such as cosine distance,dot product etc.Distributed word representation is also the important input feature of the neural network model of Q&A task,and the general approach is to input the distributed wordrepresentation features of document and question into the neural network for encoding,then conduct the answer reasoning by computing the similarity of the question and document's encoding representations.However,the above method ignores the interactive information between document and question.This paper implements a neural network based on multi-level interaction semantic representations to enhance the semantic representation capability of model and experiments on SQuAD demonstrate the effectiveness of multi-level interactive semantic representation for improving machine understanding.