Research On Meta-Knowledge Motif Representation And Memory Network Algorithm For Complex-Factoid Question Answering

With the advent of the era of intelligence,intelligent question answer ing has gradually become an important supplement to traditional information access methods.In recent years,knowledge base question answering(KBQA)has attracted the interest of a large number of researchers in the field of natural language processing and has received widespread attention.With the development and innovation of question answer ing systems,it is often desired to automatically answer more complex questions associated with multiple facts,referred to as complex-factoid questions,based on a large-scale knowledge base.Such questions are more difficult to handle because they involve multiple topic entities and are associated with multiple facts,and require multiple jumps of reasoning.Even though significant progress has been made in the field o f knowledge base question answering systems,existing question answering systems have difficulty in capturing the complex relationships and features implied by complex-factoid questions and answers due to the complexity and diversity of natural language an d knowledge graphs.Therefore,how to design a question answering system that can answer rich complex-factoid questions becomes an important and realistic problem nowadays.The thesis mainly focuses on complex-factoid questions,which can be mainly classified into path and conjunctive questions.Answering complex-factoid questions mainly faces two difficulties: first,these questions usually involve multiple entities,which brings uncertainty in the initial state and makes it difficult to locate the initial state of the question;second,some complex-factoid questions require the intersection of two(or more)related sub-paths in the knowledge base,often requiring iterative exploration of path inference,matching and combination.Too many combination s of candidate subpaths hinder the effective reasoning about the correct answer.In order to solve the above mentioned problems,this thesis introduces knowledge meta-knowledge motif and a series of methods oriented to complex-factoid questions.The main work and contributions of the thesis are:(1)Unlike the traditional path-based matching and inference strategies,this thesis introduces knowledge metamodules as the basic components of semantic representation,inference and matching.Meta-knowledge motif have the following two advantages:(i)multiple entities can be contained by meta-knowledge motif,and(ii)the intersection of subpaths can be naturally encoded by the inference of meta-knowledge motifs one by one.In other words,meta-knowledge motifs involve specific substructures that can encode higherlevel semantics.Inspired by the Trans E model of knowledge representation learning,this thesis proposes a flexible trans E-based embedding module to combine entity-relational content and structure to construct meta-knowledge motifs.(2)In this thesis,we propose a supervised question-and-answer method based on meta-knowledge motif representation for answering complex-factoid questions.The main work is to compose and represent the meta-knowledge motif using the trans E-based embedding module for the given natural language questions and based on the relational content and structure of the entities in their questions.On the bas is of this,the metaknowledge motifs that match the question are filtered by calculating the similarity to the question,and the model also contains a learning framework to update its parameters by learning from positive and negative samples,so that matching answers can be effectively performed.However,the supervised model does not take full advantage of the inference ability of meta-knowledge motif,and this thesis makes an optimization based on this model and proposes a Motif-based Memory Networks for meta-knowledge motif representation,which extends a memory network for knowledge inference and matching to facilitate the joint learning of knowledge metamodule representation and natural language questions.matching followed by joint learning,and final ly its matched answer will be used as the answer to the question.(3)The results of experiments with three datasets,Path Question,World Cup2014,and SPACES-C,show that the algorithm proposed in this thesis can answer complexfactoid questions effectively and efficiently compared with existing related knowledge graph question and answer methods.