Computation And Evaluation Of Semantic Relations In Retrieval Based Question Answering System

Question answering system plays an important role in the current research of natural language processing.End-to-end question answering system based on deep-learning receives natural language sentences without lexical and syntac-tic processing,automatically extracts semantic features,and selects the correct answers that meet the requirements of the question from the candidates.Un-derstanding the semantic relationship between natural language sentences is the most difficult link in question answering system,which is also the core prob-lem faced by natural language processing tasks at present.This dissertation studies how to calculate and evaluate the semantic relationship between natural language sentences in the end-to-end retrieval question answering system,and how to combine the calculation process with the evaluation process to improve the retrieval efficiency.The innovations of this thesis are as follows:firstly,a two-way weight allocation mechanism is designed to overcome the deficiencies of one-way weight allocation in the previous question answering model,which dynamically adds the influence of the answer when calculating the representation of the question;secondly,a feature enhancement mechanism at the sentence level is designed to extend the shortcomings of the previous question answering system which only strengthens the feature in the word dimension.Thirdly,we design an algorithm and model to enhance the features of sentences and words from multiple perspectives at the same time.Fourthly,we design a new feature enhancement algorithm for words only and add features hierarchically to the network.The experimental results on several public datasets show that the four innovations of this topic have outperformed the benchmark model and reached the state-of-the-art level at present.Four papers for the innovations have been published or accepted by international conferences.