XL-NBT: A Cross-lingual Neural Belief Tracking Framework
share
Task-oriented dialog systems are becoming pervasive, and many companies heavily rely on them to complement human agents for customer service in call centers. With globalization, the need for providing cross-lingual customer support becomes more urgent than ever. However, cross-lingual support poses great challenges---it requires a large amount of additional annotated data from native speakers. In order to bypass the expensive human annotation and achieve the first step towards the ultimate goal of building a universal dialog management system, we set out to build a cross-lingual state tracking framework without requiring any human labor. Specifically, we assume that there exists a source language with dialog belief tracking annotations while having no access to any form of dialogue data for the other target languages. Then, we pre-train a state tracker for the source language as a teacher, which is able to exploit easy-to-access parallel data and distill its own knowledge to the student state tracker in target languages. In this paper, we specifically discuss two different types of common parallel resources (bilingual corpus and bilingual dictionary) and design different strategies to realize our transfer learning framework. Experimentally, we successfully use English state tracker as the teacher to transfer its knowledge to both Italian and German trackers and achieve promising results.
READ FULL TEXT
