Research And Implementation Of Answer Acquisition Algorithm For Commonsense Questions

The Information Age witnesses the incredible expansion speed of the Internet,which has become the most important infrastructure for users to obtain relevant information.Many researchers have attracted attention by the study of extracted required content from the massive data on the Internet.In the situation,the question answering system greatly improves the efficiency of information acquisition because of its ability to directly give answers and related information according to questions.The effectiveness of the answer acquisition algorithm directly affects the performance of this kind of system.Owing to the characteristics of commonsense questions that the answer acquisition algorithm cannot obtain the required deeplevel information through direct methods such as retrieval,these methods must have the ability to explore implicit information to solve such problems effectively.In this situation,the implicit information,such as the implicit evidence and the implicit relation,can help the acquisition do well in getting suitable answers from entitly-linked level.However,most of the current answer acquisition algorithms have the following problems: Firstly,most of the existing answer acquisition algorithms have insufficient reasoning ability,which leads to the fact that the algorithms can only use the existing knowledge for answer acquisition operations,instead of reasoning based on the existing knowledge.The information in the evidence for solving common-sense problems and the scope of some of the obtained evidence is limited by external knowledge resources.Secondly,most existing answer acquisition algorithms simply use explicit relationships retrieved from knowledge graphs to solve the problems but overlook the exploration of implicit relations with in-depth paths.To solve the challenges above,this thesis proposes Knowledge-Augmented Generative Commonsense Reasoning Algorithm(KGCRA)and Commonsense Question Answer Acquisition Algorithm based on Graph Reasoning Relations Network(GRRN),and designs and creates a prototype system for common-sense question answer acquisition based on the above algorithms.The main work of this thesis is as follows:(1)Aiming at solving the problem that the existing answer acquisition algorithms are insufficient in reasoning ability,unable to reason enough evidence for solving common-sense problems based on existing knowledge,and the scope of some obtained evidence is affected by the limitation of external knowledge resources,this thesis proposes KGCRA to take the keyword information extracted from the questions and options as input.After that,KGCRA encodes them through a text encoder,and then uses the knowledge-augmented graph encoder to obtain the knowledge information about entities in the knowledge reasoning graph through the graph neural network,and updates the nodes representation.Finally,new evidence information is generated through the text encoder and the knowledge-augmented graph decoder with multi-head attention mechanism.Experiment results show that the evidence information generated by this method has better performance on logicality and readability.(2)Focusing on the problem that the existing answer acquisition methods cannot explore the deep-level paths and implicit relationships in the knowledge graph,but only use the surface information in the knowledge graph through retrieval methods,this thesis proposes a GRRNbased common-sense question answer acquisition algorithm.The algorithm starts with the keywords in the questions and options,and finds the path between the entities represented by the keywords from the knowledge graph.Then,the evidence information of the path and the graph is obtained through the path evidence fusion and the graph evidence fusion,respectively.Then,the relations between the words in question and options is obtained through the text evidence fusuion.This thesis fuses the above three kinds of evidence information and the new generated evidence information obtained by the KGCRA for the second time,and obtains the final answer for prediction through the decision-making layer.Experiment results show that this method has better performance in obtaining the accuracy of the answer.(3)Based on the methods proposed above,this thesis designs and implements a prototype system for obtaining answers to common-sense questions.The system adopts the current mainstream micro-service architecture and sets up different service modules.The system is very easy to develop and maintain,and has strong scalability and high availability.According to the specific question input by the user,the prototype system obtains the answer information through the stored knowledge graph and the trained model,and returns it to the user.The system also uses the front-end visualization chart technology to display the inference path of the model in the process of obtaining the answer in detail,which verifies the effectiveness of the algorithm proposed in this thesis.In addition,the unit test of the prototype system is also carried out in this thesis,and the test results show that the system can cope with the normal use requirements of users.