Research And Implementation Of Commonsense Reasoning Technology Based On Path Mining

The intelligent model based on machine learning has been widely used in the fields of travel,consumption,medical treatment and so on,and has a far-reaching impact on people's life,study and work,which is due to the rapid development of the three pillars of artificial intelligence,algorithm,computing power and data.When human beings think about problems,they will analyze problems in combination with background knowledge such as spatial relationship,causality,scientific facts and common sense of social customs.This kind of knowledge is insignificant to human beings,but the current artificial intelligence model is still unavailable.In recent years,a large number of studies have begun to try to integrate common sense knowledge into intelligent models,Common sense reasoning Q A has become a new research direction in the field of artificial intelligence.Compared with traditional Q A methods,common sense reasoning Q A does not provide background knowledge,and faces problems such as insufficient evidence,complex semantics,knowledge reasoning and so on.In order to solve the challenges and problems in the question and answer task of common sense reasoning,the main research work of this thesis as follows:(1)A common sense reasoning question answering method based on expanding background knowledge evidence is proposed.Through the common sense correlation analysis of the common sense question and answer data set Commonsense QA and the existing knowledge base,a relationship based knowledge triplet selection algorithm in the knowledge graph Concept Net is proposed,combined with the Wikdictionary interpretation information crawled by the web crawler to jointly construct high-quality background knowledge,and the pre training model ALBERT is used to extract complex text semantic features,Finally,a vector reasoning method based on attention weight sum is proposed to model the complete common sense reasoning question answering system,and the common sense reasoning ability of the model is verified by comparative experiments.(2)A relationship extraction method of two-stage fine-tuning pre trained model is proposed.In order to further improve dictionary knowledge,the de-noising and purification of text knowledge is reduced to a relationship extraction task.Firstly,a relationship similarity experiment is designed to fine tune ALBERT in the first stage to learn the hidden relationship between relationship type and text structure,and then use the data set with relationship type label to fine tune ALBERT in the second stage to complete the relationship extraction task,Finally,the ablation experiment verifies that the relationship similarity experiment and the pre trained model can effectively complete the relationship extraction task.(3)A knowledge path mining method based on relationship extraction model is proposed.In order to fully mine the rich common sense knowledge in the dictionary to establish the relationship between problems and options,this thesis explores the application method of migrating the relationship extraction model to the Commonsense QA dataset and Wikdictionary,so as to provide more accurate knowledge for the common sense reasoning model and enrich the background knowledge at the same time,Finally,a comparative experiment is designed to verify that path mining can effectively improve the common sense reasoning ability of question and answer model.