Research On Gaussian Distribution Representation And Learning Model Of Chinese Word Sense

Word representation is a basic task in natural language processing,and the simplest method is the one-hot representation based on vector space,which is simple and feasible,but has some shortcomings such as sparse data and inability to express the similarity between words.In order to solve these problems,some scholars have proposed "word embedding",which has achieved good results in many natural language processing tasks and has become the main method of word representation.However,words in natural language are uncertain or polysemy.When a word is represented by a vector,the word corresponds to a point in a multi-dimensional space,which will cause the semantic representation of the word to be too rigid and contain very limited semantic information,and fail to reflect the uncertainty of the expression and relation of the word.Therefore,we think that using distributions to represent words is a good choice.The Gaussian distribution is a kind of distribution with excellent mathematical properties,which can represent words well.The mean value of the Gaussian distribution can represent the sense of words,and the variance can represent the uncertain information of words.That is,relationships between words can be captured by relationships between distributions.At the same time,the word representation model based on Chinese is also paid attention to.According to the characteristics of Chinese characters,the semantic information contained in the model can be mined to effectively improve the quality of the word representation model.In addition,the Chinese knowledge base constructed by researchers can be fully utilized to further enhance the semantic expression of words.In this paper,Gaussian distribution is used to represent words.According to the characteristics of Gaussian distribution,the expected likelihood kernel is used as the energy function to measure the relationship between words,and the max-margin ranking objective is used as the loss function to train the model,and the skip-gram model of Word2 vec was used for training.At the same time,according to the characteristics of Chinese characters,the component information of Chinese characters is added to the word representation learning,and the PRL-CC model is proposed.Firstly,the feature files of Chinese character are obtained,and each Chinese character has its split components.Then,the component information corresponding to each Chinese character is added as the input of the model.Similarity,the semantic information of HowNet is integrated into word representation learning,and the PRL-SE model is proposed.The HowNet file is obtained first,each word should have its sense number and each sense have its sememes.We only take the two most commonly used sense of words,and then divide them into two dimensions to add different sense information,so that the different sense of words can be trained separately.For the two models,we give the corresponding model structure,and the results are compared with those of different models in terms of word similarity calculation,text classification,named entity recognition task and word qualitative analysis,it can be seen that our model has a good effect,and can better express the uncertainty and polysemy of words.