// the find
liuhuanyong/CrimeKgAssitant
Crime assistant including crime type prediction and crime consult service based on nlp methods and crime kg,罪名法务智能项目,内容包括856项罪名知识图谱, 基于280万罪名训练库的罪名预测,基于20W法务问答对的13类问题分类与法律资讯问答功能.
A Chinese-language legal NLP demo from a researcher at the Chinese Academy of Sciences, covering crime type classification (202 categories), legal Q&A question routing (13 categories), and a retrieval-based legal consultation system. Built on 2.88M training samples for crime classification and 200K Q&A pairs. Squarely aimed at Chinese law, so its value outside that context is zero.
The crime classification model hitting 92% accuracy on 2.88M samples with doc2vec + SVM is a legitimately solid result for the era. The ES + semantic similarity hybrid scoring for Q&A retrieval is a sensible, practical architecture that avoids the brittleness of pure keyword matching. The pre-trained models ship with the repo, so you can actually run inference without sourcing your own training data. The 856-node crime knowledge graph in kg_crime.json is a real artifact — structured data of that specificity doesn't exist many places.
Last meaningful commit was late 2023 and the approach is pre-LLM-era throughout — the SVM classifier and CNN/LSTM question router have been obsoleted by fine-tuned BERT-family models, and the README doesn't acknowledge this. The QA system is purely retrieval-based (ES + cosine sim over a static corpus), meaning it can only return existing answers verbatim; it will confidently surface irrelevant responses when the query has no close match, as the demo already shows ('售卖危违禁物品' returning '没什么'). No inference API, no Docker setup, no requirements.txt — getting the environment right requires archaeology. The knowledge graph is essentially a flat JSON file with no tooling around it, so 'knowledge graph' is doing a lot of marketing work for what amounts to a lookup table.