Word Sense Disambiguation using a Bidirectional LSTM

In this paper we present a model that leverages a bidirectional long short-term memory network to learn word sense disambiguation directly from data. The approach is end-to-end trainable and makes effective use of word order. Further, to improve the robustness of the model we introduce dropword, a regularization technique that randomly removes words from the text. The model is evaluated on two standard datasets and achieves state-of-the-art results on both datasets, using identical hyperparameter settings.