Question Intention Classification Based On Information Extraction

In recent years,with the development of AI technology,technology of question answering system has been more mature.Through information extraction,the question answering system can accurately understand and analyze natural language questions,then return more accurate answers.On the one hand,according to the source of answer,the question answering system can be divided into answer generation system and search-based system.On the other hand,according to the different ways of getting information,the search-based question answering system can be further divided into web retrieval system and knowledge base-based system.Now knowledge bases have reached an impressive coverage of general knowledge,the search query question system of the knowledge base is widely popular.The knowledge-based question answering system is mainly for factoid questions(entity,relation,entity).Information extraction can extract entities and relations from the question,and they will work well on question understanding.Entity and relation are important components of triples.Accurately extracting related entities and relations from questions not only helps to better analytical questions but also provides more accurate question intention categories.This paper wants to utilize the information of entities and relations from questions to analyze the questions' potential intention categories.The main contributions of this paper are as follows:Entity and relation extraction model.This paper proposes a new entity and relation extraction model,which contains two sub-tasks: entity and relation's keyword extraction,and relation mapping.For the entity and relation's keywords extraction task,we present a new labelling scheme,then design a new labelling annotation model(BI-LSTM-LSTM).For the relation mapping task,we extract relation features from the knowledge base,then build a feature vector for each relation.Final there is a feature mapping function which can measure the score of relation mapping through the feature vector.Intention classification of question.The questions' intention classification model is based on entity and relation information of the question.According to the result of the extraction of entity and relation in the question,this task is divided into two parts: the question that has extracted the complete triple information,and the question that does not extract the complete triple information.For the former,the questions intent categories depend on the answer entity types.For the latter,question intention classification based on KNN algorithm.The sentence-level feature vector is constructed by extracting the features of the sentence level to measure the distance between the questions.We evaluate our models on public datasets(Web Quest,Graph Quest).Experimental result show that under the same conditions,our entity and extraction model can get higher F1 values than other models,including entity extraction,relation extraction,and comprehensive extraction of entity and relation.On the question intention classification task,the model can achieve the task of English question intention classification.