Research And Application Of Word Sense Disambiguation Method Based On Contextual Semantic

Human language plays a very important role in human civilization and social development,and it is an important cornerstone of human social development.However,ambiguity in human language is a common phenomenon.The same word may have multiple different meanings,and understanding the correct meaning is indispensable for effective communication and communication,so word sense disambiguation（WSD）is essential in natural language processing tasks.WSD refers to finding the definite meaning of ambiguous words in a specific context.To understand WSD tasks from linguistics,context is the main factor that determines the meaning of a word,and morphology also contributes to WSD.Current WSD methods mainly use open source data sets and additional relational knowledge to disambiguate with the help of pre-trained models.Due to the limitations of labeled data and open source knowledge base,the current English WSD method mainly disambiguates the data of Word Net annotation system.However,there are some problems in the Word Net annotation system,such as the granularity of word meanings,and the interpretation of word meanings is not a complete sentence but a phrase or word.This is not good for language learning,language communication,and language cognition.Based on the above questions,this thesis has the following research contents:（1）Considering that part-of-speech information can help the model disambiguate target words,we use part-of-speech tools for part-of-speech tagging in context,and select appropriate part-of-speech granularity based on Sem Cor corpus.At the same time,Collins Dictionary annotation system with moderate meaning granularity and complete meaning interpretation is adopted.We construct a Collins Dictionary dataset using the online Collins Dictionary and part-of-speech tools.（2）A WSD method based on BERT integrating context features and parts of speech features was proposed,including a WSD method based on BERT integrating external context features and parts of speech features and a WSD method based on BERT integrating internal context features and parts of speech features.Compared with the methods without part-of-speech information,they all show better disambiguation performance on the Collins Dictionary test data set,which verifies that part-of-speech information has a positive effect on word sense disambiguation task.（3）A contrastive learning based on WSD method was proposed.And a training template was constructed based on WSD task,which improved the learning of the similarities and differences between sentence pairs into the learning of the similarities and differences between context and candidate word meanings,so that the model could learn the correct meaning of ambiguous words in context.The₁ values of this method reached 81.3%and 76.0%in the Collins Dictionary test and validation data sets.（4）The desktop platform of the WSD dictionary of the Collins annotation system was built.The platform is based on the Flask framework,and the disambiguation results are displayed in the form of an Excel table.It can be used for word meaning tagging,as well as manual correction.