TL;DR
This paper tackles the challenge of accurately detecting offensive language in contemporary Korean political discourse, a domain where existing datasets are outdated and LLMs struggle with nuanced interpretations. The authors introduce a novel, large-scale dataset (PoliticalK.O) and three distinct judgment methods (SEJ, PEJ, MRJ) to evaluate LLM performance, finding that prompt-based approaches (PEJ, MRJ) offer more robust and contextually aware judgments than traditional fine-tuning methods (SEJ) relying on older data.
What Problem Does This Paper Address?
The detection of offensive language, particularly in political discourse, is a persistent challenge. Existing research often relies on outdated datasets, limiting their ability to capture the evolving nature of online language and political dynamics. This is especially true for Korean, where recent studies are scarce. Furthermore, LLMs, while powerful, can struggle with the subtle nuances, sarcasm, and context-dependent nature of political commentary, leading to inconsistent or inaccurate judgments. The authors aim to bridge this gap by creating a contemporary dataset and developing robust evaluation frameworks that can assess LLM performance in this complex domain, even in the absence of traditional ground truth labels.
Methods
The paper introduces three novel judgment methods designed to assess offensive language detection by LLMs:
- Supervised Ensemble Judgment (SEJ): This method fine-tunes LLMs on five existing Korean offensive language datasets. Predictions from these fine-tuned models are then aggregated using majority voting. The goal is to leverage the human-curated annotations of these older datasets.
- Prompt-variants Ensemble Judgment (PEJ): This approach utilizes three state-of-the-art LLMs and five different prompt variants (Vanilla, Definition, Summary, Few-Shot, and a combination of the latter three). These variants provide varying levels of context, from simple instructions to article summaries and few-shot examples. The outputs from different prompts and LLMs are aggregated via majority voting.
- Multi-Debate Reasoning Judgment (MRJ): This method extends prior multi-agent frameworks. It assigns distinct personas to LLMs (offensive and non-offensive) to generate rationales for their judgments. These rationales are then debated by the agents, and a final "judge" agent makes the decision based on the arguments and stances presented. This aims to capture a broader spectrum of contextual dimensions.
Datasets
- PoliticalK.O: The authors constructed a new, large-scale dataset from Naver, South Korea's largest news platform. It comprises 114,000 articles and 9.28 million user comments from 2024, covering six political topics: Presidential Office, National Assembly/Political Parties, North Korea, Administration, National Defense/Foreign Affairs, and General Politics. This dataset is unique in its recency and focus on contemporary Korean political discourse.
- Prior Korean Offensive Language Datasets: SEJ utilized five existing datasets: K-Haters, KODOLI, KoLD, K-MHaS, and UnSmile. These datasets, however, were noted to be outdated, with the most recent collected in early 2022.
Key Insights and Ideas
- The Challenge of Outdated Data: The study strongly suggests that relying on older datasets for training LLMs leads to significantly poorer performance when evaluating against contemporary discourse. SEJ, which heavily relies on these older datasets, consistently underperformed PEJ and MRJ.
- Prompt Engineering Matters: PEJ demonstrated that carefully crafted prompts, especially those providing contextual information (like article summaries or explicit definitions), can significantly improve LLM judgment accuracy and consistency.
- Multi-Agent Debate for Nuance: MRJ's multi-debate approach proved effective in capturing nuanced interpretations, showing intermediate agreement differences and suggesting its stability in handling subjective content.
- Pseudo-Labels as a Practical Proxy: In the absence of true ground truth, the authors successfully used a "ground trust" construct, derived from majority voting across their three judgment methods, as a reliable proxy for evaluation. This highlights a practical strategy for evaluating models on new, unannotated data.
- Contextual Cues Trump Raw Offensiveness: The study found that PEJ and MRJ, which leverage contextual information, were more effective than SEJ. This implies that understanding the context and intent behind language is crucial, rather than just identifying explicit offensive terms.
Key Results
- PEJ and MRJ Outperform SEJ: When evaluated against the "ground trust," PEJ and MRJ consistently achieved higher accuracy, precision, and F1 scores compared to SEJ across all topics. PEJ, in particular, showed accuracy and F1 scores exceeding 90%.
- High Offensiveness Rates: Across all topics and judgments, a consistently high proportion of offensive comments were identified, aligning with historical patterns of verbal aggression in political discourse.
- Topic-Specific Patterns: While offensiveness was generally high, the National Defense/Foreign Affairs topic showed a lower proportion of offensive comments compared to General Politics.
- Leave-One-Out Analysis: The study revealed that MRJ showed intermediate agreement differences, suggesting greater stability. In contrast, removing components from SEJ and PEJ led to more substantial shifts, indicating the sensitivity of these methods to their constituent parts.
- Prompt Variant Effectiveness: Within PEJ, the "Definition" prompt and "Few-Shot" examples showed promise, while "Summary" prompts were less effective, potentially due to bias from article content.
Limitations and Future Work
- Generalizability: The dataset and findings are specific to Korean political discourse in 2024. Further validation is needed to assess generalizability to other languages and time periods.
- LLM Limitations: Even advanced LLMs, particularly in zero-shot settings, have limitations. The study notes that PEJ's performance was below that of fine-tuned PLMs under optimal conditions, highlighting the need for tailored detection methods.
- Dataset Bias: While efforts were made to mitigate bias, the inherent nature of online comments and political discourse means potential biases may still exist.
- Future Work: The authors suggest developing tailored methods for complex political discourse and exploring the flexible adaptability of their introduced judgments across languages. They also acknowledge the need for further discussion on how references to political figures affect perceptions of offensiveness.
Summary updated