Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dual-space Hierarchical Learning for Goal-guided Conversational Recommendation

Can Chen, Hao Liu, Zeming Liu, Xue Liu, Dejing Dou

TL;DR

This work tackles proactive goal planning in conversational recommender systems by exploiting multi-level dialog goals (type, entity, attribute) in dual spaces. The authors propose Dual-space Hierarchical Learning (DHL), combining hierarchical representation learning with cross-attention to mutually reinforce goal-type and goal-entity representations, and hierarchical weight learning with bi-level optimization to adaptively weight high- and low-level goals. A soft labeling strategy gradually guides conversations toward the final recommendation, improving training signals without sacrificing stability. Empirical results on DuRecDial and TG-ReDial demonstrate DHL’s superiority over strong baselines, with ablations confirming the importance of each component and case studies illustrating interpretable attention and weighting dynamics. Overall, DHL advances proactive, natural transitions in CRS by harnessing hierarchical, dual-space relationships among dialog goals, with practical impact on more effective and engaging recommendations.

Abstract

Proactively and naturally guiding the dialog from the non-recommendation context (e.g., Chit-chat) to the recommendation scenario (e.g., Music) is crucial for the Conversational Recommender System (CRS). Prior studies mainly focus on planning the next dialog goal~(e.g., chat on a movie star) conditioned on the previous dialog. However, we find the dialog goals can be simultaneously observed at different levels, which can be utilized to improve CRS. In this paper, we propose Dual-space Hierarchical Learning (DHL) to leverage multi-level goal sequences and their hierarchical relationships for conversational recommendation. Specifically, we exploit multi-level goal sequences from both the representation space and the optimization space. In the representation space, we propose the hierarchical representation learning where a cross attention module derives mutually enhanced multi-level goal representations. In the optimization space, we devise the hierarchical weight learning to reweight lower-level goal sequences, and introduce bi-level optimization for stable update. Additionally, we propose a soft labeling strategy to guide optimization gradually. Experiments on two real-world datasets verify the effectiveness of our approach. Code and data are available here.

Dual-space Hierarchical Learning for Goal-guided Conversational Recommendation

TL;DR

This work tackles proactive goal planning in conversational recommender systems by exploiting multi-level dialog goals (type, entity, attribute) in dual spaces. The authors propose Dual-space Hierarchical Learning (DHL), combining hierarchical representation learning with cross-attention to mutually reinforce goal-type and goal-entity representations, and hierarchical weight learning with bi-level optimization to adaptively weight high- and low-level goals. A soft labeling strategy gradually guides conversations toward the final recommendation, improving training signals without sacrificing stability. Empirical results on DuRecDial and TG-ReDial demonstrate DHL’s superiority over strong baselines, with ablations confirming the importance of each component and case studies illustrating interpretable attention and weighting dynamics. Overall, DHL advances proactive, natural transitions in CRS by harnessing hierarchical, dual-space relationships among dialog goals, with practical impact on more effective and engaging recommendations.

Abstract

Proactively and naturally guiding the dialog from the non-recommendation context (e.g., Chit-chat) to the recommendation scenario (e.g., Music) is crucial for the Conversational Recommender System (CRS). Prior studies mainly focus on planning the next dialog goal~(e.g., chat on a movie star) conditioned on the previous dialog. However, we find the dialog goals can be simultaneously observed at different levels, which can be utilized to improve CRS. In this paper, we propose Dual-space Hierarchical Learning (DHL) to leverage multi-level goal sequences and their hierarchical relationships for conversational recommendation. Specifically, we exploit multi-level goal sequences from both the representation space and the optimization space. In the representation space, we propose the hierarchical representation learning where a cross attention module derives mutually enhanced multi-level goal representations. In the optimization space, we devise the hierarchical weight learning to reweight lower-level goal sequences, and introduce bi-level optimization for stable update. Additionally, we propose a soft labeling strategy to guide optimization gradually. Experiments on two real-world datasets verify the effectiveness of our approach. Code and data are available here.
Paper Structure (23 sections, 10 equations, 5 figures, 4 tables)

This paper contains 23 sections, 10 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Related work
  4. Method
  5. Framework Overview
  6. Base Model
  7. Hierarchical Representation Learning
  8. Hierarchical Weight Learning
  9. Soft Labeling
  10. Experiments
  11. Datasets
  12. Baselines
  13. Implementation Details
  14. Overall Results
  15. Ablation Study
  16. ...and 8 more sections

Figures (5)

  • Figure 1: An example of muti-level goal sequences in a human-machine conversation.
  • Figure 2: Overview of DHL. Taking the goal type sequence and the goal entity sequence as input, DHL outputs the next goal entity to lead the conversation via dual-space hierarchical learning and soft labeling guidance.
  • Figure 3: An illustrative example of the soft labeling strategy. A soft parameter $0.02$ is attached to the one-hot goal label, which guides the prediction to approach the final recommendation goal News Rec.
  • Figure 4: Parameter sensitivity of DHL.
  • Figure 5: Case study.

Theorems & Definitions (3)

  • Definition 1
  • Definition 2
  • Definition 3