// the find
liuhuanyong/PersonRelationKnowledgeGraph
ChinesePersonRelationGraph, person relationship extraction based on nlp methods.中文人物关系知识图谱项目,内容包括中文人物关系图谱构建,基于知识库的数据回标,基于远程监督与bootstrapping方法的人物关系抽取,基于知识图谱的知识问答等应用。
A Chinese-language person relationship extraction system that builds a knowledge graph from scraped web data, covering ~11k people and ~36k relationship pairs. It demonstrates the full pipeline from distant supervision to bootstrapping-based relation extraction, aimed at NLP researchers or students working with Chinese text. The domain is Chinese celebrities and public figures, not general-purpose.
The distant supervision approach is well-explained with worked examples — you can see exactly what the labeled training data looks like. The knowledge base (11k people, 1.1k relation types) is a real artifact, not a toy demo. Using Scrapy for corpus construction is a sensible choice that separates data collection from modeling. The README honestly benchmarks the method's limitations rather than overselling accuracy.
Last commit was December 2018 — Python 3.5/3.6 bytecode is checked in, Scrapy targets are almost certainly dead, and the Sogou relationship graph API it scraped from has changed. There's no model checkpoint or training script included, just the data pipeline and raw corpus; you cannot reproduce the relation classifier without significant reconstruction. The 1,144 relation types sounds impressive until you realize most have single-digit frequency — the long tail is noise. Documentation is entirely in Chinese with no English equivalent beyond the title, which limits its reach considerably.