Target-constrained Bidirectional Planning for Generation of Target-oriented Proactive Dialogue

Jian Wang,Dongding Lin,Wenjie Li

DOI: https://doi.org/10.1145/3652598

IF: 4.657

2024-03-13

ACM Transactions on Information Systems

Abstract:Target-oriented proactive dialogue systems aim to lead conversations from a dialogue context toward a pre-determined target, such as making recommendations on designated items or introducing new specific topics. To this end, it is critical for such dialogue systems to plan reasonable actions to drive the conversation proactively, and meanwhile, to plan appropriate topics to move the conversation forward to the target topic smoothly. In this work, we mainly focus on effective dialogue planning for target-oriented dialogue generation. Inspired by decision-making theories in cognitive science, we propose a novel target-constrained bidirectional planning (TRIP) approach, which plans an appropriate dialogue path by looking ahead and looking back. By formulating the planning as a generation task, our TRIP bidirectionally generates a dialogue path consisting of a sequence of pairs using two Transformer decoders. They are expected to supervise each other and converge on consistent actions and topics by minimizing the decision gap and contrastive generation of targets. Moreover, we propose a target-constrained decoding algorithm with a bidirectional agreement to better control the planning process. Subsequently, we adopt the planned dialogue paths to guide dialogue generation in a pipeline manner, where we explore two variants: prompt-based generation and plan-controlled generation. Extensive experiments are conducted on two challenging dialogue datasets, which are re-purposed for exploring target-oriented dialogue. Our automatic and human evaluations demonstrate that the proposed methods significantly outperform various baseline models.

computer science, information systems

What problem does this paper attempt to address?

The paper attempts to address the problem of how to effectively plan dialogue paths in goal-oriented proactive dialogue systems to guide the conversation smoothly from the current context to the preset goal, such as recommending a specific item or introducing a new topic. Specifically, the paper focuses on how the system can take reasonable actions and choose appropriate topics during the conversation to maintain the dialogue's attractiveness while smoothly transitioning to the target topic. This involves two main challenges: first, maintaining the dialogue's attractiveness and actively pushing the conversation forward; second, ensuring the dialogue coherently advances towards the target topic, stimulating the user's interest in the upcoming recommendation.

To address these challenges, the authors propose a novel Target-constrained Bidirectional Planning (TRIP) method, which plans the dialogue path through both forward-looking (from current to target) and backward-looking (from target to current) approaches. This method aims to construct a reasonable dialogue path by generating a series of action-topic pairs and improve planning consistency and accuracy by minimizing decision gaps and using a contrastive generation mechanism. Additionally, the authors propose a target-constrained decoding algorithm to better control the planning process and use the planned dialogue path to guide dialogue generation, exploring two variants: prompt-based generation and plan-controlled generation. Experimental results show that the proposed method significantly outperforms various baseline models in both automatic and human evaluations.