Abstract: Aiming at insufficient generalization ability of current WSD (word sense disambiguation) model, Multi-Channel Residual Hybrid Dilated Convolution with Attention (MRHA) WSD model is proposed. Linguistic knowledge is used to construct disambiguation features, 3 vectorization methods are used to vectorize disambiguation features to form 3-channel word embedding matrix, and positional coding is deeply fused with 3-channel word embedding matrix. A complex convolutional encoder is designed to increase expressive ability of WSD model. Experiments are conducted on SemEval-2007: Task#5 and SemEval-2021: Task#2. Experimental results show that compared with the newest WSD model using Clustered Sense Labels (CSL) and Multi-Channel Convolutional Neural Networks with Multi-Head Attention (MCNN-MA), average bias of the proposed method is respectively reduced to 1.345% and 2.157%.

Key words: Word sense disambiguation, Linguistic knowledge, Hybrid Dilated Convolution, Convolutional encoder