Layout Attention Network For Question Answering Over Table And Text

The Internet contains a wealth of knowledge that people are interested in.Question answering systems,which use information on the Internet as the knowledge base and can automatically provide answers to users,have gained increasing attention from academia and industry.Traditional Question answering systems rely on rules designed by experts to provide answers to users through grammatical or syntactic structures.However,these approaches suffer from problems such as the inability to generalize.In recent years,deep learning methods have been used in many Question answering systems and have gained good results.Existing deep learning-based Q&A systems are often based on only a single type of information,such as tables or text.However,in practical scenarios,in order to answer people's questions,the system often needs to combine two types of information for multi-hop reasoning to find relevant knowledge to answer users.Therefore,this paper selects the Question Answering over Table and Text as the research topic.Based on the analysis of existing methods,the following problems are found in the current Question Answering Systems over Table and Text:(1)The existing models filter documents and tables related to user questions based on shallow semantic features such as TF-IDF and sub-character matching,ignoring the deep semantic similarity between user questions and tables and documents.(2)The contents of the same rows or columns in a table often imply important relationships such as temporal order and alignment.However,existing models perform reasoning based on pure textual information,ignoring the spatial information of the table.(3)User questions have various answer types,such as two-choice,three-choice,compare size,and generation.The existing Q&A models use semantic matching-based reading comprehension models.These models' performance are various according to question types.To address the above problems,we propose a Layout Attention Network for Question Answering over Table and Text.Our contribution can summerised as follows:(1)We introduce the Poly-Encoder as a base model for text similarity matching to achieve deep semantic similarity matching of related tables and texts by capturing multidimensional fine-grained features of user questions.(2)We propose a Layout Attention Network to fuse spatial information with textual information(3)We propose a cascaded reading comprehension model for different types of user questions.The accurate output of different types of questions and answers is achieved by designing corresponding prediction methods after triage of different types of questions.(4)Compared with the benchmark model,our model gets a 2% effect improvement on Hybrid QA.Adequate quantitative and qualitative experiments demonstrate the superiority of the form-text joint QA method with spatial information proposed in this paper.