XQA-DST: Multi-Domain and Multi-Lingual Dialogue State Tracking

Findings of the Association for Computational Linguistics: EACL 2023, pages 999–1009 May 2-6, 2023 ©2023 Association for Computational Linguistics XQA-DST: Multi-Domain and Multi-Lingual Dialogue State Tracking Han Zhou1,∗ Ignacio Iacobacci2 Pasquale Minervini3,4 1University of Cambridge 3University of Edinburgh2Huawei Noah’s Ark Lab, London 4University College London hz416@cam.ac.uk ignacio.iacobacci@huawei.com p.minervini@ed.ac.uk Abstract Dialogue State Tracking (DST), a crucial com- ponent of task-oriented dialogue (ToD) sys- tems, keeps track of all important information pertaining to dialogue history: filling slots with the most probable values throughout the con- versation. Existing methods generally rely on a predefined set of values and struggle to gen- eralise to previously unseen slots in new do- mains. To overcome these challenges, we pro- pose a domain-agnostic extractive question an- swering (QA) approach with shared weights across domains. To disentangle the complex domain information in ToDs, we train our DST with a novel domain filtering strategy by ex- cluding out-of-domain question samples. With an independent classifier that predicts the pres- ence of multiple domains given the context, our model tackles DST by extracting spans in ac- tive domains. Empirical results demonstrate that our model can efficiently leverage domain- agnostic QA datasets by two-stage fine-tuning while being both domain-scalable and open- vocabulary in DST. It shows strong transferabil- ity by achieving zero-shot domain-adaptation results on MultiWOZ 2.1 with an average JGA of 36.7%. It further achieves cross-lingual transfer with state-of-the-art zero-shot results, 66.2% JGA from English to German and 75.7% JGA from English to Italian on WOZ 2.0. 1 Introduction Task-oriented dialogue systems are designed to provide natural conversation with users and assist them in achieving daily goals. With the growth of task-oriented dialogue systems, there is an increas- ing interest in supporting dialogues among many domains and languages to fit the users’ demands. However, either modelling a multi-domain or multi- lingual dialogue system requires substantial data collected in real scenarios. This data acquisition ∗ Work done while at UCL. Code is available at https://github.com/ hanzhou032/xqa-dst procedure is extremely expensive, and it motivates us to resolve this challenge by leveraging dialogue data in rich-resource domains and languages via zero-shot transfer learning. Dialogue State Tracking (DST) is crucial for accurately extracting user intents and goals over multiple turns within the dialogue. Based on the tracked dialogue states, the dialogue manager makes corresponding next actions with back-end results, where the accuracy of the DST becomes absolutely vital. With a fully predefined ontology, traditional approaches tackle the DST as a classifi- cation problem by enumerating every combination of slot-value pairs (Mrkši´c et al., 2017; Zhong et al., 2018). Those approaches are strongly limited by their scalability, as some slots (e.g. name) have an unbounded set of slot values. Secondly, they are generally not flexible to unseen slot-value pairs, making them more difficult to adapt to zero-shot transfer learning. Moreover, a completely prede- fined ontology is hard to acquire and not scalable for ToD systems in real applications. To overcome those challenges, we take inspira- tion from Gao et al. (2020) and investigate how DST can …

Content truncated to 500 words