HKJudge dataset unlocks AI analysis of Hong Kong court rulings with 290K annotated sentences

Researchers led by Xi Xuan have released HKJudge, the first expert-annotated legal discourse dataset for Hong Kong court judgments. The corpus spans all five levels of Hong Kong's court hierarchy and contains approximately 290,000 sentences and 6.5 million tokens, each labeled by ten legal linguistics experts. The annotation schema operates on two tiers: at the sentence level, every sentence is assigned one of 26 rhetorical roles (e.g., finding of fact, legal reasoning, ruling); at the span level, sentences are further tagged for three key sentencing elements—charge, imprisonment term, and fine. The annotators achieved an inter-annotator agreement of κ=0.8, indicating strong reliability.

Using HKJudge, the team formulated two benchmark tasks: rhetorical role classification and legal element extraction. They evaluated four BERT-based models, two open-source LLMs (under both zero-shot and fine-tuning settings), and four commercial LLMs. Results highlight the value of sentence-level discourse annotation for modeling the structure of Hong Kong judgments and show that fine-tuned models significantly outperform zero-shot approaches. The dataset and code are publicly available, offering a rich foundation for future work in legal judgment prediction, AI-assisted legal research, and automated case analysis.