Multi-Task Pre-Training for Plug-and-Play Task-Oriented Dialogue System (2109.14739v2)

Abstract: Pre-trained LLMs have been recently shown to benefit task-oriented dialogue (TOD) systems. Despite their success, existing methods often formulate this task as a cascaded generation problem which can lead to error accumulation across different sub-tasks and greater data annotation overhead. In this study, we present PPTOD, a unified plug-and-play model for task-oriented dialogue. In addition, we introduce a new dialogue multi-task pre-training strategy that allows the model to learn the primary TOD task completion skills from heterogeneous dialog corpora. We extensively test our model on three benchmark TOD tasks, including end-to-end dialogue modelling, dialogue state tracking, and intent classification. Experimental results show that PPTOD achieves new state of the art on all evaluated tasks in both high-resource and low-resource scenarios. Furthermore, comparisons against previous SOTA methods show that the responses generated by PPTOD are more factually correct and semantically coherent as judged by human annotators.

Overview of Multi-Task Pre-Training for Task-Oriented Dialogue Systems

The paper presents PPTOD, a unified plug-and-play model designed for task-oriented dialogue (TOD) systems. The primary motivation behind the paper is to address the limitations in existing TOD approaches, which largely depend on a cascaded generation framework. This traditional framework can result in error propagation across various sub-tasks, such as dialogue state tracking (DST), policy learning (POL), and natural language generation (NLG), and also involve extensive data annotation workloads. PPTOD seeks to innovate by offering a unified architecture facilitated by a dialogue multi-task pre-training strategy.

Methodology and Key Innovations

PPTOD targets the integration of dialogue modules into a single neural architecture, leveraging pre-trained LLMs (PLMs) to eliminate the need for manual annotation across all sub-tasks. This is achieved through a multi-task pre-training approach, wherein the model is trained across a diverse set of TOD-related tasks, enabling it to derive skills from partially annotated datasets. The essence of PPTOD lies in a plug-and-play framework that decouples sub-domains via task-specific prompts, enabling parallel generation processes.

The paper adopts T5 variants (small, base, large) to initialize PPTOD and involves pre-training using a heterogeneous dialog corpora comprising over 2.3 million utterances from 80 domains. Eleven curated datasets with varied annotations are utilized to simulate different TOD sub-tasks including NLU, DST, POL, and NLG.

Experimental Evaluation

PPTOD is evaluated against several benchmark datasets, primarily focusing on MultiWOZ 2.0 and 2.1 for end-to-end dialogue modeling, DST, and user intent classification tasks.

Numerical Results

1. End-to-End Dialogue Modeling: PPTOD demonstrates superior performance in full-data conditions for MultiWOZ by achieving high Inform, Success, BLEU, and Combined scores. Particularly, the model yields notable improvements in low-resource setups (as minimal as 1% of training data), outperforming baselines by substantial margins.
2. Dialogue State Tracking: Although classification-based models slightly outperform PPTOD in joint goal accuracy, PPTOD’s generation-based approach offers more scalability, adapting effortlessly to new ontology labels.
3. Intent Classification: The model exhibits robust accuracy both in limited and full training scenarios, underscoring its efficiency in task-oriented dialogues without necessitating extra parameters for new tasks.

Implications and Future Directions

The paper underscores the transformative potential of employing a unified model like PPTOD for TOD tasks. By effectively reducing inference latency and minimizing error accumulation typically observed in cascaded methods, PPTOD sets a precedence for future research to explore unsupervised and few-shot learning paradigms within TOD systems. The implications are particularly significant in real-world applications where frequent ontology updates necessitate adaptive dialogue models. Additionally, the methodological insights into task-specific prompt utilization could inspire innovations in multilingual and cross-domain dialogue systems.

Future research may delve into enhancing the model’s understanding capabilities by integrating more refined NLU modules or exploring semi-supervised learning pathways to optimize the performance under scarce data conditions. The theoretical foundation established by PPTOD promises scalable dialogue systems capable of sustaining complex, multi-domain conversations, thus paving avenues for robust conversational agents.