Fine-grained Verbal Attack Detection via a Hierarchical Divide-and-Conquer Framework

TL;DR

This work tackles the challenge of detecting covert verbal attacks in online Chinese discourse by introducing HACD, a tree-structured, spatiotemporally annotated dataset. It then proposes a hierarchical divide-and-conquer framework comprising four lightweight modules that separately handle explicit and implicit attack detection and analysis, guided by controllable context from conversation trees. Empirical results show that smaller, modular models under this framework can match or exceed larger monolithic models while improving efficiency, with strong cross-lingual generalization to Catalan and English datasets. The approach offers a scalable, interpretable path for fine-grained attack analysis in dynamic multi-turn conversations, with potential applications in moderating online platforms.

Abstract

In the digital era, effective identification and analysis of verbal attacks are essential for maintaining online civility and ensuring social security. However, existing research is limited by insufficient modeling of conversational structure and contextual dependency, particularly in Chinese social media where implicit attacks are prevalent. Current attack detection studies often emphasize general semantic understanding while overlooking user response relationships, hindering the identification of implicit and context-dependent attacks. To address these challenges, we present the novel "Hierarchical Attack Comment Detection" dataset and propose a divide-and-conquer, fine-grained framework for verbal attack recognition based on spatiotemporal information. The proposed dataset explicitly encodes hierarchical reply structures and chronological order, capturing complex interaction patterns in multi-turn discussions. Building on this dataset, the framework decomposes attack detection into hierarchical subtasks, where specialized lightweight models handle explicit detection, implicit intent inference, and target identification under constrained context. Extensive experiments on the proposed dataset and benchmark intention detection datasets show that smaller models using our framework significantly outperform larger monolithic models relying on parameter scaling, demonstrating the effectiveness of structured task decomposition.

Figures (5)

• Figure 1: Data structure after preprocessing. Level 1 denotes a first-level comment, while Level 2 denotes a second-level comment. A level comment block represents a first-level comment block. (x,y) indicates that the comment belongs to level x and has a sequence number of y within that level. For example, Subcomment(2,1) in the figure refers to the first secondary comment under the primary comment Comment(1,1).
• Figure 2: An illustrative example of an annotated conversation thread. The left side displays the hierarchical discourse structure (from Level 1 to Level 3), while the right side presents the corresponding fine-grained labels across seven dimensions, including attack form, target, type, intent, and hazard level.
• Figure 3: Overview of the proposed attack detection and analysis framework. The processing pipeline first evaluates comment text for explicit attack. If such markers are absent, it incorporates structural context to detect implicit attack. All detected attacks are then routed to specialized analyzers to generate fine-grained labels (e.g., target, intent, harmfulness), while any identified non-attacks are automatically labeled with null values.
• Figure 4: Performance comparison of WTNF vs. WTWF strategies across model sizes. The shaded orange area highlights WTWF's accuracy gain over WTNF after overtaking at 9B parameters.
• Figure 5: A case study illustrating multi-dimensional analysis on a comment thread. The text on the right displays the model's output for each user comment, covering seven metrics from "Attack or not" to "Confidence level." Note that the color of the text corresponds to the specific label categories defined in the top legend.