融合语义信息与问题关键信息的多阶段注意力答案选取模型

自动问答系统可以帮助人们快速从海量文本中提取出有效信息,而答案选取作为其中的关键一步,在很大程度上影响着自动问答系统的性能.针对现有答案选择模型中答案关键信息捕获不准确的问题,本文提出了一种融合语义信息与问题关键信息的多阶段注意力答案选取模型.该方法首先利用双向LSTM模型分别对问题和候选答案进行语义表示;然后采用问题的关键信息,包括问题类型和问题中心词,利用注意力机制对候选答案集合进行信息增强,筛选Top K个候选答案;然后采用问题的语义信息,再次利用注意力机制对Top K个候选答案集合进行信息增强,筛选出最佳答案.通过分阶段地将问题的关键信息和语义信息与候选答案的语义表示相结合,有效提高了对候选答案关键信息的捕获能力,从而提升了答案选取系统的性能.在三个数据集上对本文所提出的模型进行验证,相较已知同类最好模型,最高性能提升达1.95%.

An Answer Selection Model Based on Multi-Stage Attention Mechanism with Combination ofSemantic Information and Key Information of the Question

With the rapid development of Internet technology,the amount of text information in the network increases exponentially,hence people usually use some search engines to retrieve the required information from mass data.A search engine can be regarded as a special question answering system.When a question is given,the general processing flow of the automatic question answering system is as follows:first,the system analyzes the question to obtain its type,semantics and other relevant information;then,select a candidate answer set from the answer database according to the analysis results;finally,the system will rearrange the candidate set with various sorting techniques and select the best answer or the text with the best answer to return to the user.The flow shows that the selection effect of the best answer will directly affect the overall performance of the automatic question answering system.Traditional answer selection models usually use lexical or syntactic analysis and artificial constructing feature to select answers,which is difficult to capture the semantic association information between questions and candidate answers.With the development of deep learning technology,researchers applied the deep learning framework into the answer selecting task,use the neural network model to obtain the semantic association information of the question and the candidate answer,and evaluate the matching association degree between them,then select the answer with the strongest matching relationship as the best answer.Because the selection of answers depends entirely on the information carried in the question,researchers often generate attention vector from the question semantic information to update the semantic representation of the candidate answers.Although this kind of attention model can strengthen the semantic relationship between the question and the candidate answer,it ignores the relationship of key information between them,therefore,the effectiveness of such models is affected.For different types of questions,the concerned content in best answers is often different.For example,when asking time-related questions,the best answer should be more focused on the key information of time or the information with strong time semantic association;when asking weather-related questions,the best answer should pay more attention to the key information related to weather.Also,the existing attention-based answer selection methods often establish the model of questions and answers at the same stage,which is not easy to capture the differences between the various candidate answers.To solve the problem that the answer key information capture is not accurate in the existing answer selection model,this paper proposes an answer selection model based on a multi-stage attention mechanism with a combination of semantic information and key information of the question.Firstly,this method uses a bidirectional LSTM model to represent questions and candidate answers semantically.Then the key information of the question,including the type of question and the headword of the question,is used to enhance the information of the candidate answer by attention mechanism,and the Top K candidate answers are selected.Finally,the attention mechanism with semantic information of the question is used again to enhance the information of the Top K candidate answer set to select the best answer.By combining the key information and semantic information of the question to enhance the semantic representation of the candidate answer in multi-stages,the ability to capture the key information of candidate answers is effectively improved,and the performance of the answer selection system is improved.The experimental results on three datasets show that the highest performance improvement is up to 1.95%compared with the other state of the art models.