Application of Recent Advances in Language Modeling to Improve Dialog Systems
Recently, Facebook AI Research introduced Fairseq, a sequence modeling toolkit based on convolutional sequence to sequence learning. This Gated Convolutional Network allows researchers to train custom models for translation, summarization, language modeling and other text generation tasks. When presented, Fairseq outperformed numerous major benchmarks in the domain of automatic translation, and large-scale language tasks. In the application context of the National Bank, we decided to investigate how this new way to exploit non-recurrent approach for language tasks could be used to improve dialog systems robustness. In this presentation, we will show how, by using FairSeq for paraphrasing, it is possible to significantly improve the precision of intent classification in a chatbot.
Eric Charton hold a Master in machine learning applied to voice recognition, and a Ph.D. in machine learning applied to Information extraction and natural language generation. He worked as scientist and research project coordinator in academic context in Europe (University of Avignon) and North America (CRIM, École Polytechnique de Montréal) before becoming head of search engine research and development at Yellow Pages Canada. Since March 2018, he is Senior AI Director at National Bank of Canada.