Abstract: Polysemous words have been considered as learning emphasis and major obstacles for foreign students who learn Chinese as a foreign language.Word Sense Disambiguation (WSD),which is mainly used to determine the specific meaning of polysemous words in a given context,has important application in human-computer interaction,machine translation,automatic essay scoring and other emerging applications.It is also the difficulty in teaching Chinese as a foreign language and HSK examination.The existing word sense disambiguation methods are shown to be with low accuracy,lack of corpus,simple features etc.Considering teaching Chinese as foreign language and its evaluation corpus,it is a hot research topic on building Chinese polysemous words WSD based on deep neural networks.It provides necessary technical support for achieving automatic HSK essay scoring.Existing researches assumed that semantic items are mutually independent and thus paid little attention to their evolutional relationships among items.To solve this problem,this paper firstly made research on semantics of typical Chinese polysemous words.Followed by semantic topology construction the basic semantic items and set-phrases were discriminated for supervised classification model training.Based on the semantic topology construction of polysemous words,the corpus samples were collected by web crawling.The supervised deep neural networks,including RNN,LSTM and GRU,were constructed subsequently.By analyzing the scrawled samples,both uni-directional and bi-directional neural networks were designed by choosing 30 words and 60 words length respectively.The final WSD classification models were obtained by multiple rounds of training and optimization to model parameters.The “Yisi”was chosen for example and used for WSD experiments within their contexts.The experimental results shows that RNN,LSTM and GRU all achieve average classification accuracy more than 75% while maximum accuracy is more than 94%.AUC under each model is more than 0.966,which shows its good performance on class imbalance among samples.Both uni-directional and bi-directional RNN models achieve best classification performance under different words length.

Key words: Deep neural networks, Long short-term memory network, Recurrent neural network, Semantic topology, Word sense disambiguation