Classifying Questions Into Fine-grained Categories Using Semantic Expansion And Multi-layer Attention Network

With the rapid development of the Internet,a large number of Community Question Answering(CQA)websites have emerged in recent years,such as Yahoo!Answers,Stack Overflow and Quora,etc.Users can ask questions freely on these websites as well as answer other users' questions.Every user wants his needs to be satisfied as fast and accurately as possible when using Q & A sites.Therefore,how to classify the existing questions accurately and finely has become an important research topic.At present,question fine-grained classification mainly faces the following difficulties:(1)The questions are relatively short,it is hard to understand the precise subject contained in a question by extracting adequate information from limited text.(2)The label contains few information,so detecting the subject and the scope of a category using the label of the category is very hard.(3)In the fine-grained classification task,the differences among the features of each category are small.Making the way to bring in more information in order to match the questions' subject and specific fine-grained categories to be the key point of fine-grained question classifying research.To solve these problems,this paper proposes an approach of Classifying Questions into Fine-Grained Categories using Semantic Expansion and Multi-layer Attention Network(SEMAN).The semantic expansion solves the problem of less information in the original question text,and the multi-layer attention model can get the slight feature difference between the categories,thus improving the classification effect.The main contributions of this paper are as follows:(1)We propose a semantic unit selection method based on dependency parsing tree.The method can find the central component of the question accurately,and around which can also find the key phrases or core words including the semantic information.(2)We propose a semantic unit expansion method based on Word2 Vec,which can adequately extend the semantic information of the original question because of closer distance between the more similar words in Word2 Vec model.(3)Based on the semantic expansion,we propose a multi-layer attention model,which can avoid the problem that the traditional attention model can only extract a single aspect of the semantic information.The multi-layer attention model can fully extract different aspects of the semantic information.The experiment in the paper is based on Yahoo!Answers dataset in the health field and Yahoo!Answers dataset in the education field.To verify the validity of our method,we design a series of comparative experiments.The final experimental results show that SEMAN can effectively solve the problem of sparse question features and small differences among the features of each category.SEMAN achieves the best results in the fine-grained classification.