Reaserch On Text Reasoning Based On Integrating External Knowledge Base Infomation

Text reasoning is a basic and very important research in the field of natural language processing.Its research object is two-stage text description,usually called premise and hypothesis.The purpose of reasoning is to judge whether there a logical relationship between these two texts based on common sense information.Since the text description is composed of one or more sentences,text reasoning can also be expressed as the identification of logical relationships between sentences.The particularity of the task of text reasoning is that it is difficult to train the model to a very satisfactory level by simply using the data in the training set.Because when human beings face the same problem,they will make comprehensive judgments based on common sense information beyond a large amount of corpus.And depending on the sentence representation and the logical relationships that may exist between sentences,the types of cormmon-sense information required are also different.Therefore,it is very important to let the machine take into account the rich external knowledge information when identifying the logical relationship between sentences.In recent years,with the rapid development of deep neural networks,it also provides a lot of ways to integrate information representation,which greatly improves the accuracy of the corresponding tasks.However,the traditional text reasoning method generally focuses on learning a training set to find a general method to complete the reasoning,but there are many key information in the test set that does not appear in the training set.In order to make up for this shortcoming,the paper introduces two kinds of external knowledge base information which are crucial in the reasoning process,such as knowledge map and affiliation map,based on the perfect model:1.Using knowledge graph information,this thesis studies the text reasoning based on local information collection,and proposes the Entailment Relations Recogniton of Sentences with Knowledge Graph(KG-ER).The main focus of the KG-ER model is the acquisition of local inference information.When the Encoding representation of the premise content is introduced,the knowledge map information is introduced,including the relationship between entities and entities.Based on these knowledge,for different assumptions,Give the weight information of the premise content sentence granularity differently,and then fuse the premise knowledge to reason.2.Using event graph information,this thesis studies the text reasoning based on event evolution,and proposes a Causality Recogniton of Sentences with Event Graph(EG-CR).The traditional knowledge map stores the entities of the noun type.In the process of reasoning,it is more necessary to introduce the context information of event development.Therefore,this paper constructs an affair map.The nodes of this map are composed of the main components describing the event.The relationship between the node and the node represents the relationship before and after the event occurs.The general reasoning ability of the model is improved by the general law information of the event development contained in this graph structure.Based on the above analysis,this paper tests the two types of text reasoning tasks,including textual relationship recognition and textual causality recognition,using a combination of the above two types of external knowledge.Experiments show that the reasonable use of these two types of external knowledge can effectively improve the reasoning ability of the model.