Word Sense Disambiguation Based on RegNet With Efficient Channel Attention and Dilated Convolution

Abstract

Word sense disambiguation (WSD) is one of key problems in field of natural language processing. Ambiguous word often has different meanings in different contexts. WSD is the process of determining semantic category of ambiguous word according to its context. It has a great impact on machine translation, speech recognition, topic detection, search engine, text classification, emotion analysis and so on. It is important for WSD task to extract effective discriminative features and design high-quality disambiguation models. This paper proposes a WSD method based on RegNet with Efficient Channel Attention (ECA) and dilated convolution (ECADCRegNet). Ambiguous word is viewed as the center, and two units on the left and two units on the right of ambiguous word are selected. Then, word, part of speech and semantic categories of these four units are used as disambiguation features. Firstly, disambiguation features are vectorized and the result is input into RegNet space. Secondly, dilated convolution is used to extract discriminative features from words, parts of speech and semantic categories for constructing feature matrix. And ECA is introduced into the feature extraction layer. Thirdly, softmax function is used to calculate the probability of ambiguous word under each semantic category and the category with the largest one is selected. Training corpus of SemEval-2007: Task#5 is adopted to optimize the proposed network. Test corpus of Semeval-2007: Task#5 is used to test the performance of the proposed network. Experimental results show that average accuracy of the proposed method achieves 81.83%, which is higher than those of other methods.

Keywords

• Word sense disambiguation
• RegNet
• neural network
• efficient channel attention
• disambiguation features