Cross-lingual transfer learning for spoken language understanding

  • 2019-04-03 08:11:57
  • Quynh Ngoc Thi Do, Judith Gaspers
  • 1

Abstract

Typically, spoken language understanding (SLU) models are trained onannotated data which are costly to gather. Aiming to reduce data needs forbootstrapping a SLU system for a new language, we present a simple buteffective weight transfer approach using data from another language. Theapproach is evaluated with our promising multi-task SLU framework developedtowards different languages. We evaluate our approach on the ATIS and areal-world SLU dataset, showing that i) our monolingual models outperform thestate-of-the-art, ii) we can reduce data amounts needed for bootstrapping a SLUsystem for a new language greatly, and iii) while multitask training improvesover separate training, different weight transfer settings may work best fordifferent SLU modules.

 

Introduction (beta)

None

 

Conclusion (beta)

None