The Research Of Knowledge Base Question Answering System Based On Deep Learning

KBQA has always been a popular research direction in the field of natural language processing.Due to the development of natural language processing and the emerging high-quality knowledge base,the traditional text-based KBQA system has been able to achieve a more satisfactory effect,and relevant achievements have also been widely used in various functional services in real life.In recent years,with the expansion of service scope and the upgrading of business requirements,the traditional KBQA has become increasingly difficult to meet the functional requirements of related tasks in real life.As for modality of knowledge,the traditional knowledge bases are mainly text-based,but now more and more QA tasks need to combine the multimodal information;from the perspective of knowledge content,the traditional knowledge base mainly contains the knowledge information based on the static nominal concept,lacking the dynamic and sequence knowledge of actions,events.The construction of the new knowledge base,the representation of the new knowledge and how to use it to solve the new QA tasks need to be explored.The motivation of this paper is to make a attempt to the above research problems.The main innovations and contributions of this paper can be summarized into the following two parts:In QA tasks for daily events and actions,it is often necessary to capture and reason the sequence of actions and events to get answers.Utilizing some deep learning skill for text sequence processing,this paper introduces a universal learning model for sequential knowledge representation,which can capture context semantic of sequence.To solve the problem of multimodal sequence,a sequential representation model of modal fusion is proposed.Three multimodal QA tasks in the RecipeQA dataset are selected as verification tasks.The model proposed in this paper surpasses state-of-the-art models in all three tasks.Finally,experiments show that the model can learn effective sequence semantic representation to support subsequent reasoning and judgment,and can effectively integrate sequence information of different modes.Considering the practical application,some daily service systems or robots often need to answer the questions about the sequential concepts of processes,actions and steps.In view of the lack of knowledge base for human daily activities and the lack of research on using knowledge base to solve non-text questions and answers,this paper creatively constructs a multimodal knowledge base for daily activities based on Charades dataset.The knowledge base contains multimodal knowledge about atomic actions and atomic action sequences in people's daily activities.This paper also constructs the corresponding benchmark for the new knowledge base,and the task form is to solve the question of video sequence QA based on the knowledge base.In our baseline model,a knowledge base searching method based on vector matching is proposed,which solves the problem of knowledge base matching with non-text information as query item.Finally,the validity of the method based on the knowledge base to answer the video formal questions is proved by the contrast experiment.