Auto-encodeurs pour la compréhension de documents parlés (Auto-encoders for Spoken Document Understanding)

Auto-encoders for Spoken Document Understanding Document representations based on neural embedding frameworks have recently shown significant improvements in different natural Language processing tasks. In the context of real application framework, the automatic transcription of spoken documents may result in several word errors, especially when very noisy conditions are encountered. This paper proposes an original representation of highly imperfect spoken documents based on the bottleneck features from a Supervised Deep autoencodeur that takes advantage of both noisy automatic and clean manual transcriptions to improve the robustness of the document representation in a noisy environment. Results obtained on the DECODA theme classification task of dialogues reach an accuracy of more than 83% with a significant gain of about 6%. MOTS-CLÉS : auto-encodeur, débruitage, reconnaissance de la parole, réseaux de neurones.