Transferable Multi-Domain State Generator for Task-Oriented Dialogue Systems (1905.08743v2)

Abstract: Over-dependence on domain ontology and lack of knowledge sharing across domains are two practical and yet less studied problems of dialogue state tracking. Existing approaches generally fall short in tracking unknown slot values during inference and often have difficulties in adapting to new domains. In this paper, we propose a Transferable Dialogue State Generator (TRADE) that generates dialogue states from utterances using a copy mechanism, facilitating knowledge transfer when predicting (domain, slot, value) triplets not encountered during training. Our model is composed of an utterance encoder, a slot gate, and a state generator, which are shared across domains. Empirical results demonstrate that TRADE achieves state-of-the-art joint goal accuracy of 48.62% for the five domains of MultiWOZ, a human-human dialogue dataset. In addition, we show its transferring ability by simulating zero-shot and few-shot dialogue state tracking for unseen domains. TRADE achieves 60.58% joint goal accuracy in one of the zero-shot domains, and is able to adapt to few-shot cases without forgetting already trained domains.

Transferable Multi-Domain State Generator for Task-Oriented Dialogue Systems

The paper presents an empirical investigation into dialogue state tracking (DST) within the domain of task-oriented dialogue systems, specifically focusing on the limitations posed by traditional reliance on predefined ontologies and lack of cross-domain knowledge sharing. The authors introduce TRADE, a novel framework designed to enhance DST by predicting dialogue states using a generative approach that leverages a copy mechanism. This methodology enables the transfer of knowledge across domains, facilitating accurate predictions of dialogue (domain, slot, value) triplets, even when encountering previously unseen data during inference.

Model Architecture

The TRADE model features three core components: an utterance encoder, a slot gate, and a state generator. The model does not rely on a predefined ontology, addressing a significant challenge in traditional DST models which struggle to enumerate all possible values and adapt to the latent dynamism of dialogue interactions. The slot gate, configured as a three-way classifier, predicates whether a slot is mentioned and distinguishes between ‘none’, ‘dontcare’, or utilizing the generated value. The state generator employs a soft-gated copy mechanism allowing for the generation of slot values directly from utterances, a strategy that enables handling unknown or variable slot values over manually enumerated lists.

Empirical Results

TRADE outperforms established DST frameworks such as MDBT, GLAD, GCE, and SpanPtr across multiple metrics on the MultiWOZ dataset, demonstrating a state-of-the-art joint goal accuracy of 48.62% and slot accuracy of 96.92%. The nuanced mechanism of TRADE allows for substantial zero-shot performance, notably achieving a joint goal accuracy of 60.58% in the taxi domain, indicating significant cross-domain knowledge transfer. Furthermore, TRADE can adapt efficiently to few-shot learning scenarios, retaining its proficiency in previously trained domains while accommodating new data—an essential feature for dynamic real-world applications where comprehensive datasets may be sparse or expensive to obtain.

Implications and Future Directions

The research progresses the field of task-oriented dialogue systems by solving pressing constraints of ontology-dependence and portraying a scalable method for cross-domain generalization. The adaptable architecture of TRADE facilitating domain transfer is an asset for artificial intelligence applications consistently interacting with dynamic and multifaceted datasets. Moving forward, employing methods from meta-learning and exploring larger-scale datasets encompassing diverse domains could enhance TRADE's adaptability and generalizability further.

The success of this approach suggests broader implications for the architectural decentralization of pretrained models, potentially eliminating the ubiquitous need for predefined schemas in many NLP applications. The continued exploration into combining TRADE with external resources could further attenuate the zero-shot learning gap and probe deeper into unsupervised dialogue system optimization.

In conclusion, TRADE presents an innovative and effective approach to DST that stands to significantly influence ongoing developments in building more robust, flexible, and capable dialogue systems.