Text Entailment Recognition Based On Integration Of Language Knowledge And Deep Learning And Its Application

With the continuous development of artificial intelligence,natural interaction continues to penetrate into people's lives,intelligent speakers,intelligent tutors,intelligent search,intelligent customer service and a series of products continue to emerge.Text entailment recognition is a basic and core task in natural language understanding task,which can enhance the ability of question answering system to understand natural language.The direct application is to verify knowledge.Compared with intention understanding in natural understanding,common sense and logical reasoning of text entailment are more challenging.In machine reading,self-learning is a challenge.Dynamic answers and automatic scoring are directly applied.With the development of in-depth learning and neural networks,and thanks to the release of large-scale data sets SNLI,the models of text entailment recognition based on neural networks emerge in endlessly.They are mainly divided into two categories: one is text entailment model based on sentence coding,the other is text entailment model based on interactive attention mechanism.Text mainly focuses on text entailment model based on interactive attention mechanism.In addition,the latest text entailment model also has the problem of worse reasoning ability,such as antonyms and hyponyms.At present,the system can not recognize the relationship between words in common sense.Therefore,the main research point of text is to introduce artificial knowledge into the text entailment recognition model to improve this problem.Firstly,we get word-to-knowledge vectors from three perspectives.At present,the main research point is how to express a word.Few people pay attention to how to express a word pair.The expression of word pairs plays an important role in text entailment recognition.First,we try to classify word-to-word relations based on text features.We want to use word pairs in the upper and lower levels.Words,synonyms and antonyms are classified to represent word-to-word relationships.Then we try TransR,a knowledge map representation tool.We hope that the relationship between entity vectors and relational vectors can help us learn more information.Finally,we model antonyms and synonyms in text entailment reasoning.In this way,we get the word pair vectors with lexical relation knowledge.Then,the features of the three knowledge vectors we acquired introduce knowledge vectors into the part of word alignment and attention mechanism.Compared with the classical model,the introduction of antonym vectors can improve greatly in specific data sets.Secondly,aiming at the existing Chinese text entailment data set,aiming at the problem of large amount of information and possible word segmentation errors in Chinese,we improve the Chinese entailment recognition data set by introducing character features and dependency analysis features,and combining the current popular context-related vectors.Finally,we try to apply the technology of text entailment to the short text scoring task in the context of Mu lesson.Specifically,we combine the standard answers to questions and questions to judge the importance of students' answers.Combining the text matching data set and the text entailment data set,we construct a text matcher to judge questions and students' answers.The matching relationship between the students' answers and the standard answers is used to judge the entailment relationship.The two parts of the joint training are used to construct the scoring model,which has greatly improved the evaluation data set.