Automatic Question Answering Over Tables

Question answering(QA)is an important computer science discipline within the field of natural language processing,which is concerned with automatically answering the natural language query questions posed by humans.According to different data sources,QA can be divided into knowledge graph based QA,document based QA,and table based QA.Compared with complex knowledge graph data,tabular data is highly time sensitive,easy to maintain,and has huge volume of data on Web.Compared with document data,tabular data contains less ambiguous information due to its structured characteristics.The above characteristics of tabular data make it an important data source of answers for search engines and QA systems.In this thesis,we focus on automatic question answering for tabular data.We summarize our research works in this thesis into the following four aspects:Content-based table search.In some cases,the table itself can be regarded as a satisfactory answer for natural language queries posed by users.In this work,we focus on content-based table retrieval.Given a natural language query,this task is to find the most relevant table in a given table set.To solve this problem,we propose a rank-based approach.In our approach,we implement a feature-based model and a neural networkbased model to measure the correlation between natural language queries and the contents of tables.Single turn neural semantic parsing over tables.In most cases,returning the entire table as an answer can not directly answer the user's question.The users need to further analyze the table to find a specific value in the table or reason based on the tabular data to obtain a satisfactory answer.Due to the semi-structured and low ambiguity of tabular data,in this work we use a method based on semantic parsing to solve this problem.We convert the natural language query proposed by the user into a logical expression which can be directly executed on the corresponding table by an existing engine to obtain a fine-grained answer.We propose a generative model to map natural language queries to SQL queries.The proposed model generates high-quality SQL queries by comprehensively considering the structure of the table and the syntax of SQL language.Multi-turn neural semantic parsing over tables.A user-friendly system should allow users to ask continuous questions over tables in the form of dialogs.In this scenario,for a single natural language query,we need to comprehensively consider its context to understand its semantics accurately.In this work,we used a semantic parsing-based approach to solve the problem of multi-turn QA over tables.We propose a knowledgeaware semantic parser to improve semantic parsing performance by integrating various types of knowledge,including grammar knowledge,expert knowledge,and external resource knowledge.Our proposed model effectively handles the co-reference and ellipsis phenomena in a sequence of natural language queries posed by usersLow-resource neural semantic parsing over tables.Semantic parsing based methods for QA over tables rely heavily on the availability of large amounts of supervised data.Due to the complexity of the target logical form for tables,it is costly to obtain such data with rich supervision.In order to alleviate this problem,this work further studied how to perform semantic parsing over tables under the low resource setting.Specifically,we propose an algorithm that can build a neural semantic parser when only prior knowledge about a limited number of simple rules is available,without access to either annotated programs or execution results.