Réseaux neuronaux profonds pour l'étiquetage de séquences (Deep Neural Networks for Sequence Labeling)

Deep Neural Networks for Sequence Labeling Since a couple of years, neural networks prove to be very effective on all NLP tasks. Recently, a variant of neural network particularly suited for sequence labeling has been proposed, which uses label embeddings. In this paper we propose a deep version of the above-mentioned variant of neural network, where several hidden layers allow to take into account separately the different types of information given as input to the model. We evaluate our variant on the same tasks as the basic version. Results show that our variant is not only more effective than the other neural models, but also that it outperforms all the other models evaluated on the same tasks, reaching state-of-the-art performances. MOTS-CLÉS : Réseaux neuronaux, apprentissage artificiel, compréhension de la parole, étiquetage de séquences.