Research On Key Technologies Of Question Answering Over Knowledge Base

Automatic question-answering system over knowledge base(KBQA)provides a direct,efficient and reliable way to get needed information.Recently,with the rapid development of Information Extraction and Data Mining technologies,some large-scale knowledge bases,such as Freebase and DBpedia,which covering multiple areas have been built.It not only provides good data resources for question answering,but also brings some new challenges.So far there are two main kind of methods for this task:Semantic Parsing based(SP-based)methods and Information Retrieval based(IR-based)methods.SP-based methods focus on converting natural language formed question to some kind of logic expression,for example lambda expression,which is exploited to query knowledge base and get answers.IR-based methods first fetch the set of possible candidate answers from knowledge base by relatively coarse information retrieval method,and then ranking technologies are often adopted to select correct answers.With the development of artificial neural network and deep learning technologies,more and more researchers attempt to exploit end-to-end neural networks to learn representations of knowledge base information,question and candidate answer,and then use these representations to get correct answers.As for the Chinese field,the NLPCC-ICCPOL 2016 KBQA task provides a large-scale opendomain knowledge base with more than 23,000 question-answer pairs.Based on this dataset,we research the two main technologies in KBQA system in this paper: recognition of named entity in question,and mapping from question to property in KB.The main content consists of following three aspects:1.LSTM language model based named entity recognition in question.To make full use of information in KB,we adapt ranking method to recognize named entity appear in question.First,we fetch the set of all possible candidate named entities by alias dictionary,then combine the corresponding output of LSTM language with two simple word surface features to calculate the scores of candidate named entities,finally choose the candidate with highest score as correct named entity.2.Convolutional neural network based question-property mapping.We exploit multi-layer CNN which based on the architecture of Siamese network to encode t he question and candidate property to corresponding semantic vectors independently,choose the candidate property whose semantic vector has highest similarity with question's as the correct relevant property.For the word alignment behavior occurred between question and correct relevant property,attention mechanism is employed to previous CNN model,combine its output with two word-surface features,the accuracy of question-property mapping is improved by large margin.3.LSTM based question-property mapping.We use LSTM to encode the relationship of semantic between question and candidate property,and combine two different kind of attention mechanism with LSTM model: static attention mechanism and input-before attention mechanism.Combine with the same word-surface features as previous,the accuracy is improved again.Then we employ the results of question-property mapping to re-rank the previous candidate named entities.Finally,we select the correct answers from KB by integrate the results of named entity recognition and question-property mapping.