A graph-based retrieval method improved some measures of legal question answering in a controlled Chinese-language research setting. It did not eliminate AI hallucinations.
What researchers studied
Jia Liu and Jianhao Tang developed a retrieval-augmented question-answering system based on academic legal publications from Chinese core journals.
The system organized material into a knowledge graph and compared vector retrieval with Cypher graph queries across three large language models.
Its closed research corpus focused on legal regulation of AI training data. The team evaluated answer quality and examined different types of hallucination produced by retrieval and generation working together.
What they found
In one configuration using Cypher retrieval with the Neo4j GraphRAG framework, GPT-4 Turbo’s unanswerable rate was 3.41%. The average BERTScore-F1 reached 80.58%.
These values describe that tested setup, not an overall accuracy guarantee for every legal question or model.
The researchers also identified five forms of hallucination specific to the GraphRAG workflow. Their analysis suggested that structured retrieval can reduce unsupported invention while introducing new errors through retrieval choices, graph structure and answer synthesis.
What it means
Connecting language models to structured legal scholarship may help them find relevant evidence and produce more grounded research answers.
However, retrieving material from a knowledge graph is not equivalent to verifying the legal validity of a conclusion. A response may still misinterpret a source or combine propositions incorrectly.
Legal researchers evaluating AI tools should inspect source traceability, coverage and the kinds of errors made, not simply whether the generated text sounds plausible.
Important limitation
The study evaluated a defined Chinese legal academic corpus and a narrow research topic. It did not test live courtroom advice, legal practice across jurisdictions or unrestricted public legal questions.
BERTScore is a text-similarity measure rather than an independent assessment of legal correctness. The reported unanswerable rate also depends on the evaluation setup.
Human expert review remains necessary before relying on generated legal interpretations.
Source Information
Study: Graph retrieval-augmented generation for legal academic viewpoint question answering: knowledge-graph retrieval, evaluation, and hallucination analysis
Authors: Jia Liu and Jianhao Tang
Journal / Source: Frontiers in Artificial Intelligence
Published: 12 November 2026
