Word Sense Disambiguation Method Based On WIC Dataset

Word sense disambiguation is one of the most important tasks in natural language processing.This task almost affects the capabilities of all applications including text translation,search engines,and so on.With the rapid iteration and development of deep learning technology,it has become possible for computers to process a large amount of text data.With the birth of text pre-training models in recent years,people can optimize and improve algorithms quickly and easily.The goal of research has also risen from word representation to sense representation,and word sense disambiguation algorithms based on different theoretical foundations and external knowledge are emerging in endlessly.In order to compare the current popular word sense disambiguation methods and models,this paper uses the MCL-WiC dataset derived from WiC as a sample to evaluate the ability of the model.And we build two kinds of models,the first one is word embedding based on context information It uses deep learning models like ELMo,BERT,that has undergone unsupervised training from a large number of corpora.And the output word vector is matched with different classification methods for word meaning comparison and classification;Another one is using labeled text library or online dictionary,etc.External resources have carried out a supervised meaning representation of the target word,and the result is obtained according to the optimal word meaning comparison.For the classification problem like WiC dataset,the experimental results show that the accuracy rate of using word vector to directly classify is significantly higher than that of sense representation.After the embedded sense is fine-tuned to the BERT model,the expected result cannot be achieved.Sense representation is affected by resource quality and sense definition.In the selection of the pre-training model of the word embedding based on contest information,ELMo is not much different from BERT_BASE,and it is significantly better than the more complex model BERT_LARGE.After comparing various classification algorithms,the cosine distance measure and Neumf can achieve better results in most cases,especially the ELMo+Neumf model achieves70.5%accuracy and 73.4 F1-score on the verification dataset.