Réseaux de neurones de quaternions pour le traitement du langage

Machine Learning algorithms reach great performances on different Natural Language Processing tasks. Among these methods, Neural Networks (NN or MLP) recently received a great interest from researchers due to their capability to represent complex internal structures. However, MLPs employ basic word level or thème-based features and, therefore, reveal little in way of document statistical structure. We propose to address this issue by extending the NN to Quaternion called QMLP to take into consideration features dependencies. A well-dedicated segmentation of document approach is also compared to the one proposed in (Morchid et al., 2013). Experiments made on a SLU task with spoken dialogues show that our QMLP associated with the proposed document segmentation outperforms other approaches, with an gain of 2% and 3% compared to MLP and (Morchid et al., 2013) respectively. We finally demonstrated that less iterations are needed by QMLPs to reach better accuracies than MLP. MOTS-CLÉS : Réseaux de neurones, Quaternions, Traitement du langage.