Supervised Bidirectional Mstl Cell-Based Approach For Lexical Disambiguation

In Word Sense Disambiguation, there are still few usages of neural networks. This direction is very promising however, the results obtained by these first systems being systematically in the top of the evaluation campaigns, with a scope for improvement which seems still large. We present in this paper a new architecture based on neural networks for Word Sense Disambiguation. Our system is at the same time less difficult to train than existing neural networks, and it obtains state of the art results on most evaluation tasks in English. The focus is on the reproducibility of our system and our results, through the use of a word embeddings model, training corpora and evaluation corpora freely accessible.