Enhancing the Performance of WSD Task Using Regularized GNNs With Semantic Diffusion

Abstract

The ambiguity of a polysemous word, regardless of its language, is referred to as the ability to have more than one meaning. Many techniques from Word Sense Disambiguation (WSD) have been used to clarify the ambiguity. With these techniques, the correct sense of a word according to its particular context can be determined. In English, there exist many ambiguous words. For example, the word cell may have different meanings based on its context. For instance, it may mean the basic structural unit of all organisms in biology domains, or it may mean a device that delivers an electric current as the result of a chemical reaction in textual materials about technology or electricity, or it completely means another different thing: a room where a prisoner is kept in textual materials about prison. In this study, to enhance the performance of the WSD task, we develop a methodology with Graph Convolutional Neural Networks (GCN), including normalization, thresholding, and regularization modules. We attempt to improve the traditional Text GCN algorithm with a semantic diffusion process, which increases the classification performance of the WSD task. As far as we know, there is no work for such a comprehensive classifier for the WSD task in English. To show the effect of the suggested model, we performed experiments on the SensEval dataset, which is a very popular dataset and benchmark in the WSD domain. The experiment results show that the regularization effect and diffusion process in GCN and Text GCN architectures are powerful strategies for the WSD task.

Keywords

• Diffusion algorithm
• graph convolutional neural networks
• graph neural networks
• graph regularization
• word sense disambiguation