Research On Word Sense Disambiguation Based On Deep Learning

Word Sense Disambiguation(WSD)is a basic task in the field of natural language processing.It judges the correct meaning of an ambiguous word according to its context.The effect of word sense disambiguation has an important impact on many downstream tasks,such as machine translation,text classification,speech recognition and information retrieval.Word sense disambiguation methods are mainly divided into knowledge-based word sense disambiguation methods and supervised word sense disambiguation methods.Knowledge base based word sense disambiguation methods usually infer the meaning of words based on contextual vocabulary information,such as part of speech and word meaning.Although knowledge based word sense disambiguation methods have high disambiguation coverage,their effectiveness is often inferior to supervised word sense disambiguation methods.The supervised word sense disambiguation method converts word sense disambiguation tasks into classification tasks,uses annotated corpora as training data,and uses supervised algorithms to learn the relationship between context and word sense annotations for word sense disambiguation.The biggest difference between supervised word sense disambiguation methods is not their model architecture,but the way they use external knowledge.However,existing work often involves simple and crude use of word meaning annotations and other external knowledge.In view of this,this study aims to improve the performance of word sense disambiguation methods by using deep learning methods to mine and utilize effective information from word sense annotations and other external knowledge.The contributions of this article are mainly reflected in the following three aspects:(1)To address the issue of existing methods not fully utilizing the interrelationships between context and word meaning annotations,this thesis proposes a word meaning disambiguation method based on text matching technology.This method converts the word sense disambiguation task into a text matching task by constructing a context word meaning annotation text pair,matching the context with all candidate word meaning annotations of the target ambiguous word.Finally,the correct word meaning of the target word is determined by determining the degree of matching between the context and word meaning annotations within the same text alignment.The experimental results show that this method can effectively mine the interactive features between context and word meaning annotations,thereby improving the performance of the word meaning disambiguation model.(2)In response to the problem of existing methods only using simple and crude concatenation to utilize multiple external knowledge,this thesis proposes a word sense disambiguation method based on multi knowledge interaction.This method designs a multi knowledge interaction module that captures the relationship between example sentence knowledge and word meaning annotations to enhance the representation of word meaning annotations.Finally,the correct word meaning of the target ambiguous word is determined by calculating the similarity score between the target ambiguous word and the word meaning annotation and the target ambiguous word in the example sentence.The experimental results show that this method can capture the interactive features between various external knowledge and improve the performance of the word sense disambiguation model.(3)This thesis proposes a word sense disambiguation method based on memory enhancement mechanism to address the issue of existing methods ignoring the interrelationships between different target ambiguous words in the same context.This method designs an effective memory enhancement mechanism that enhances the representation of the target ambiguous word by storing the disambiguated word meaning annotations of other ambiguous words in the same context and interacting with the representation of the target ambiguous word.Finally,the goal of disambiguating ambiguous words is achieved by calculating the similarity score between the target ambiguous word and its candidate word annotation.The experimental results show that this method can effectively utilize the representation of identified ambiguous words in the same context to enhance the current target ambiguous word representation,thereby improving the performance of the word se nse disambiguation model.