// the find
liuhuanyong/EventTriplesExtraction
An experiment and demo-level tool for text information extraction (event-triples extraction), which can be a route to the event chain and topic graph, 基于依存句法与语义角色标注的事件三元组抽取，可用于文本理解如文档主题链，事件线等应用。
A Chinese-language NLP tool that extracts subject-verb-object triples from text using three different backends: LTP (dependency parsing + semantic role labeling), Baidu's DDParser, and a POS-tag rule pattern approach. Aimed at researchers building event chains or topic graphs from Chinese text. This is explicitly an experiment/demo, not a production library.
Comparing three extraction strategies side-by-side on the same examples is genuinely useful for understanding the tradeoffs — LTP gives richer SRL-augmented output, DDParser is heavier but accurate, pattern-based is fastest but brittle. The SRL integration in method 1 is the most interesting part: using semantic roles as a fallback when dependency parsing misses agent-patient structure is a real practical technique. The project is honest about what it is — the README calls it an experiment, not a framework. Code is minimal enough that you can actually read and modify it.
Last commit was 2022; the Python 3.6 bytecode in `__pycache__` committed to the repo tells you the maintenance posture. LTP and Baidu DDParser have both released breaking API changes since this was written, so setup will require version archaeology before anything runs. There are no tests, no error handling, and no packaging — this is four standalone scripts you paste into your project. The comparison outputs in the README reveal real extraction failures (wrong subjects, truncated objects, spurious triples) with no discussion of precision/recall or any benchmark, so you have no idea how it performs on your actual data.