Research On Commonsenseqa Based On Conceptnet

Nowadays,the pre-trained model often chooses words with higher word frequency by default when making choices.It mainly judges the relationship between the two words through context and word frequency,but it does not understand what kind of relationship does the two words have.In other words,the pre-trained model is not sufficient in semantic understanding.Therefore,this article attempts to integrate the common sense information into the pre-trained model and further improve the semantic understanding ability of the pre-trained model.The main work of this paper can be summarized as the following three points:(1)Proposed common sense question answering method based on conceptnet path.In order to obtain the triple information from the conceptnet that can support the connection between the question and the choices,and use the triple information to improve the ability of the semantic understanding of the pre-trained model.First,propose the common sense question answering model based on the conceptnet singlehop path,and add the choices' neighbors about the question concept to the input text;then propose the common sense question answering model based on the conceptnet multi-hop path,and obtain the question concept through lexical features.Then obtain the path within 4 hops between the concept in the question and the choices,and give a path scoring and path fusion mechanism;finally,a combined model based on the conceptnet single-hop path and multi-hop path is realized.(2)Propose a common sense question answering method based on concept net subgraph embedding.All common sense information of the conceptnet is in the form of graphs.The common sense question-and-answer method based on the conceptnet path only uses common sense text,but has not used graph information.In order to make full use of the information of the nodes and edges in the subgraph of the conceptnet,use the nltk(Natural Language Toolkit)tool to obtain the lexical features of the words in the question,find the concepts through the part-of-speech relationship,construct the subgraphs of the question concepts and choices in the conceptnet based on these entity concepts,and use the 300-dimensional embedding or graph embedding of the subgraph by graph convolutional network,and combine the final embedding with the hidden layer output of the pre-trained model.(3)Propose a common sense question answering model based on multi-stage learning.Psychology research found that the more the number of choices,the more the decision makers will pay attention to the shallow information of the choices.In order to make the common sense question answering model integrated with the conceptnet pay more attention to the deep semantic information of the text,the CSQA(commonsense QA)five-choose-one question and answer stage is divided into five-choose-two stage and two-choose-one stage.The common sense question and answer method based on the conceptnet path is used to construct the common sense question and answer model of the fusion concept network used in this stage.And analyze the experimental results in terms of the types of question word,the number of question words,and the types of conceptnet relations.