Étude des modèles neuronaux profonds pour la compréhension automatique du langage naturel dans les habitats intelligents

Abstract : The goal of the VocADom project is to create a natural language understanding system for processing voice orders in a smart home for the elderly. Machine learning approach seems more flexible than the rule-based approach, which does not take into account orders that do not conform to grammar ; and studies have shown that older people are inclined to deviate from imposed grammar.
However there is a lack of learning data in this domain. To remedy the problem, a grammar generating the artificial corpus for the smart home domain was created. One of the goals of this work is to improve this grammar. After the improvement, the resulting artificial corpus has 42195 annotated sentences. It was then evaluated by training three state-of-the-art models on it - Tri-CRF, att-RNN, RASA, as well as Tf-seq2seq, a sequence-to-sequence model. The models were then tested on a small natural existing corpus recorded as part of the VocADom project. This allowed us to see the limitations of the artificial corpus. These models were then compared to the models that were trained and tested on the PORTMEDIA corpus of touristic domain. This comparison allowed us to know that the performance of the models trained on the artificial corpus is due to the limitations of the artificial corpus rather than the limitations of the models themselves. All the models trained and tested on PORTMEDIA gave good results but the models trained on the artificial corpus and tested on the VocADom corpus were much less efficient. This shows the difficulty of accounting for the lexical and syntactic diversity of the real data.
All models show comparable performances but Tf-seq2seq, unlike Tri-CRF, att-RNN et RASA, doesn’t require aligned data and is easier to train.
Document type :
Master Thesis
Complete list of metadatas

Cited literature [75 references]  Display  Hide  Download

https://dumas.ccsd.cnrs.fr/dumas-01846913
Contributor : Uga - Bulles <>
Submitted on : Monday, July 23, 2018 - 9:28:07 AM
Last modification on : Tuesday, April 2, 2019 - 1:47:55 AM

Identifiers

  • HAL Id : dumas-01846913, version 1

Citation

Anastasiia Mishakova. Étude des modèles neuronaux profonds pour la compréhension automatique du langage naturel dans les habitats intelligents. Sciences de l'Homme et Société. 2018. ⟨dumas-01846913⟩

Share

Metrics

Record views

49

Files downloads

60