Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CATCH: A Controllable Theme Detection Framework with Contextualized Clustering and Hierarchical Generation

Rui Ke, Jiahui Xu, Shenghao Yang, Kuang Wang, Feng Jiang, Haizhou Li

TL;DR

CATCH addresses the challenge of controllable theme detection in customer-service dialogues by integrating context-aware intra-dialogue representation, preference-guided cross-dialogue clustering, and hierarchical LLM-based label generation. The framework jointly leverages topic segmentation, a Preference Reward Model, and a three-stage label consolidation process to produce coherent, domain-adaptive theme labels aligned with user preferences. Across DSTC-12 benchmarks, CATCH outperforms strong baselines in both in-domain and cross-domain settings and achieves second place in official blind evaluations using a lightweight 8B model. Ablation and analysis demonstrate that each module contributes significantly and that hierarchical generation effectively denoises and stabilizes labels under clustering noise. The work offers a practical, resource-efficient approach to scalable, user-aware theme discovery in multi-domain dialogues with potential applications in proactive dialogue systems and knowledge retrieval.

Abstract

Theme detection is a fundamental task in user-centric dialogue systems, aiming to identify the latent topic of each utterance without relying on predefined schemas. Unlike intent induction, which operates within fixed label spaces, theme detection requires cross-dialogue consistency and alignment with personalized user preferences, posing significant challenges. Existing methods often struggle with sparse, short utterances for accurate topic representation and fail to capture user-level thematic preferences across dialogues. To address these challenges, we propose CATCH (Controllable Theme Detection with Contextualized Clustering and Hierarchical Generation), a unified framework that integrates three core components: (1) context-aware topic representation, which enriches utterance-level semantics using surrounding topic segments; (2) preference-guided topic clustering, which jointly models semantic proximity and personalized feedback to align themes across dialogue; and (3) a hierarchical theme generation mechanism designed to suppress noise and produce robust, coherent topic labels. Experiments on a multi-domain customer dialogue benchmark (DSTC-12) demonstrate the effectiveness of CATCH with 8B LLM in both theme clustering and topic generation quality.

CATCH: A Controllable Theme Detection Framework with Contextualized Clustering and Hierarchical Generation

TL;DR

CATCH addresses the challenge of controllable theme detection in customer-service dialogues by integrating context-aware intra-dialogue representation, preference-guided cross-dialogue clustering, and hierarchical LLM-based label generation. The framework jointly leverages topic segmentation, a Preference Reward Model, and a three-stage label consolidation process to produce coherent, domain-adaptive theme labels aligned with user preferences. Across DSTC-12 benchmarks, CATCH outperforms strong baselines in both in-domain and cross-domain settings and achieves second place in official blind evaluations using a lightweight 8B model. Ablation and analysis demonstrate that each module contributes significantly and that hierarchical generation effectively denoises and stabilizes labels under clustering noise. The work offers a practical, resource-efficient approach to scalable, user-aware theme discovery in multi-domain dialogues with potential applications in proactive dialogue systems and knowledge retrieval.

Abstract

Theme detection is a fundamental task in user-centric dialogue systems, aiming to identify the latent topic of each utterance without relying on predefined schemas. Unlike intent induction, which operates within fixed label spaces, theme detection requires cross-dialogue consistency and alignment with personalized user preferences, posing significant challenges. Existing methods often struggle with sparse, short utterances for accurate topic representation and fail to capture user-level thematic preferences across dialogues. To address these challenges, we propose CATCH (Controllable Theme Detection with Contextualized Clustering and Hierarchical Generation), a unified framework that integrates three core components: (1) context-aware topic representation, which enriches utterance-level semantics using surrounding topic segments; (2) preference-guided topic clustering, which jointly models semantic proximity and personalized feedback to align themes across dialogue; and (3) a hierarchical theme generation mechanism designed to suppress noise and produce robust, coherent topic labels. Experiments on a multi-domain customer dialogue benchmark (DSTC-12) demonstrate the effectiveness of CATCH with 8B LLM in both theme clustering and topic generation quality.
Paper Structure (27 sections, 9 equations, 4 figures, 9 tables, 1 algorithm)

This paper contains 27 sections, 9 equations, 4 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Intra-dialogue Theme Detection
  4. Inter-dialogue Theme Detection
  5. Methodology
  6. Task Definition
  7. Context-Aware Intra-Dialogue Theme Representation
  8. Dual-branch Topic Segmentation Framework.
  9. Contrastive Objectives for Encoder-Specific Learning.
  10. Two-Stage Adaptation Strategy.
  11. Preference-Enhanced Topic Clustering
  12. Hierarchical Theme Label Generation
  13. Experiments
  14. Experimental Settings
  15. Implementation Details
  16. ...and 12 more sections

Figures (4)

  • Figure 1: Illustration of the controllable theme detection task. Given a set of dialogues with unlabeled utterances, a theme is generated for each utterance. The theme granularity is influenced by auxiliary inputs such as user preferences dstc12.
  • Figure 2: The overall workflow of CATCH.
  • Figure 3: Illustration of the Preference-Enhanced Topic Clustering Algorithm. "$\color{teal}\checkmark$" denotes a strong linking preference, while "$\color{red}\times$" denotes a strong splitting preference.
  • Figure 4: Correlation analysis of PeC and HieGen. Figure \ref{['fig:analysis']}(a) illustrates the relationship between FCC and cluster-level metrics (NMI and Accuracy).