Research On Sememe Prediction Using Dictionary Definitions Based On Deep Learning

Sememe is the minimum semantic unit that can no longer be divided in human language.It plays an important role in various tasks in the field of natural language processing.HowNet is a typical sememe knowledge base,which is constructed by linguists for many years and has been widely used.With the development of society,the vocabulary and semantics in the language are constantly changing.The method of updating sememe knowledge base by manual annotation is time-consuming and laborious,and there is a problem of consistency in annotation.The task of automatic sememe prediction has become very important in the construction of knowledge bases,and current sememe prediction technologies still have many problems.Factors affecting the prediction effect of sememes include whether the model is appropriate,whether more information or knowledge is introduced,whether knowledge features can be fully extracted,whether the mapping process from features to sememes is reasonable,and whether the polysemy and the problem of low-frequency words can be solved.Existing research methods have not solved the above problems very well.In view of the challenges faced by automatic sememe prediction,this paper proposes two solutions based on deep learning technology to predict sememes by using dictionary definitions.One is to use an encoder based on attention mechanism to realize sememe prediction.The other is based on local semantic correspondence.Experiments show that the proposed method can achieve state-of-the-art performance in the task of sememe prediction.At the same time,this paper further verifies the practicability and effectiveness of the proposed method through Reverse Dictionary task which is a downstream task of NLP.The main work and contributions are as follows:(1)Proposing a method of sememe prediction using an encoder based on attention mechanism.Most of the existing sememe prediction methods do not combine rich knowledge information or make insufficient use of knowledge.Definitions in dictionaries are standard semantic description which can be used to predict sememes,but there are many shortcomings in existing methods.This paper improves the encoder by combining the attention mechanism,in order to solve the problem of mining definition information effectively.This paper further optimizes the sememe prediction model by fusing multiple information,which effectively improves the sememe prediction effect.(2)Researching on sememe prediction method based on local semantic correspondence.Encoder is used to encode a definition into a vector as semantic features.This is a finitedimensional vector with only a limited amount of information.By studying the relationship between the words in the definition and the target word and sememes,we find that it has the property of local semantic correspondence.Based on this property,a new method of sememe prediction is proposed.This method not only effectively improves the prediction effect of semems,but also solves the problems in predicting sememes of low-frequency words or polysemous words.(3)Implementing and improving reverse dictionary system using sememe prediction methods.The two sememe prediction methods proposed in this paper can be effectively applied to the reverse dictionary task,and the effect can be further improved by introducing sememes and other knowledge.It also has social significance as the first Chinese reverse dictionary system.Experiments show that the proposed methods have obvious improvement effect and good robustness in the task of sememe prediction,and further demonstrate the practical value of this research through the reverse dictionary task.They also prove the importance of sememe knowledge.The research on automatic sememe prediction has wide application value and farreaching practical significance in the field of artificial intelligence.