Research On Knowledge Base Question Answering Technology Based On Deep Learning

The goal of automatic question answering is to allow the machine to understand the natural language questions raised by humans and return a direct and exact answer.The rapid development of large-scale knowledge bases provides rich and convenient resource support for the goal of implementing automatic question answering systems.Therefore,the knowledge base question answering(KBQA)has attracted more and more attention in both industry and academia.However,there are many challenges in knowledge base question answering.First of all,it is a key step to extract topic entities in questions and link them into the knowledge base,which is the first step to let machine understand natural language question.Secondly,natural language questions are much different with the structured semantic in knowledge base.How to bridge the semantic gap between questions and knowledge base is also an important challenge for KBQA.In this thesis,we use deep learning models to handle these challenges in KBQA.Specifically,we divide KBQA system into two steps:candidate retrieval and answer selection.In the candidate retrieval,we firstly extract the topic entity in the question and then use the topic entity to retrieve the relevant triples in knowledge base as candidate answers.In the answer selection,we use deep learning method to calculate the semantic similarities between the candidate answers and the question and rank the candidate answers according to the semantic similarities.We conduct experiment on the NLPCC-ICCPOL 2016 KBQA dataset and obtain a good performance.The main contributions are as follows:In this thesis,we propose a deep learning-based topic entity extraction model to extract topic entities in questions.We combine word and char embedding to learn representations for the words in a question.Then,the word sequence is fed into a bidirectional long short-term memory(BiLSTM)network to obtain the global features of each word.At last,a convolutional neural network is used to capture the local context features and predict the label for each word.We propose a self-attention based deep semantic representation model to learn semantic representations for natural language questions and predicates in knowledge base.Thus,we can map the questions and predicates into the same semantic space and use a cosine function to measure the semantic similarity between questions and predicates.We use three different deep learning structures to learn the deep level semantic representation,including convolutional neural network(CNN),BiLSTM and combination of BiLSTM and CNN.And we use a novel context based self-attention mechanism to learn better deep semantic representation.In order to further improve the performance of the KBQA system,we propose a deep fused model which considers the internal connections between the topic entity extraction and semantic representation learning.We regard them as two related tasks and use multi-task learning method to train them.In addition,to learn rich and comprehensive semantic representations for questions and predicates,we use a hierarchical semantic representation model which takes both shallow semantic and deep semantic information into account and uses a gate mechanism to integrate these two level semantic representations.