HKJudge: A Legal Discourse-Annotated Corpus for Interpreting What Courts Find, How They Reason, and What They Rule
Researchers introduce HKJudge, the first expert-annotated corpus of Hong Kong court judgments with ~290k sentences across all five court levels. The dataset enables analysis of judicial reasoning through 26 rhetorical roles and legal element extraction, establishing benchmarks for AI models in legal judgment prediction.
HKJudge addresses a significant gap in legal AI research by providing the first sentence-level discourse-annotated dataset for Hong Kong judgments. The dataset's construction reflects rigorous academic standards, with ten legal linguistics experts achieving substantial inter-annotator agreement (κ = 0.8) across ~6.5 million tokens spanning criminal cases. This foundational work enables computational analysis of how courts structure arguments, establish facts, and deliver rulings—processes previously inaccessible to systematic study.
The emergence of specialized legal corpora represents a broader trend toward domain-specific AI training datasets. While general-purpose language models dominate headlines, specialized applications in law require annotated data that captures nuanced professional discourse. HKJudge's two-tier schema—identifying rhetorical roles at sentence level and extracting sentencing elements at span level—models the actual cognitive architecture of legal judgment, not merely surface patterns.
For the AI industry, this work demonstrates the commercial viability of specialized legal datasets. The benchmark evaluation against BERT models, open-source LLMs, and commercial systems establishes performance baselines that guide future development. Legal tech companies developing judgment prediction systems now have quantifiable targets and validated methodologies.
The research's impact extends beyond Hong Kong's legal system. Similar corpus construction projects could emerge across jurisdictions, creating proprietary datasets valuable for localized legal AI applications. The publicly released dataset and code accelerate research velocity while establishing standards for legal NLP evaluation.
- →HKJudge contains ~290k annotated sentences from Hong Kong criminal judgments across all court hierarchies with 0.8 inter-annotator agreement
- →The dataset enables two core tasks: rhetorical role classification (26 roles) and legal element extraction (charges, sentences, fines)
- →Benchmark evaluation shows BERT, open-source LLMs, and commercial models achieve varying performance levels on legal discourse tasks
- →The public release of HKJudge dataset and code accelerates development of judgment prediction systems and legal discourse analysis research
- →Specialized legal corpora represent a growing trend toward domain-specific AI training data beyond general-purpose language models