// the find
liuhuanyong/ChineseEmbedding
Chinese Embedding collection incling token ,postag ,pinyin,dependency,word embedding.中文自然语言处理向量合集,包括字向量,拼音向量,词向量,词性向量,依存关系向量.共5种类型的向量
A 2018 research dump of five pre-trained Chinese NLP embedding sets — character, word, pinyin, POS tag, and dependency-relation vectors — trained on Chinese Wikipedia with word2vec (skip-gram) and PCA. Useful mainly as a quick source of off-the-shelf Chinese embeddings for someone prototyping older-style NLP pipelines, not as a library or tool.
Covers five distinct linguistic units in one place (char, word, pinyin, POS, dependency), which is an unusual combination you won't find bundled elsewhere. The README documents training method, corpus, vector dimensions, and vocabulary sizes for each embedding set. Character and word vector examples show genuinely sensible nearest neighbors (surname clusters, synonym clusters), suggesting the core token/word embeddings are reasonably trained.
Dead since December 2018 — no commits, no issues addressed, nothing resembling maintenance. It's two scripts (train_vector.py, test_vector.py) with no requirements.txt, no setup.py, no API, and no license file, so you're on your own for dependency versions and reuse. The model directory only ships 3 of the 5 advertised .bin files (dep_vec_10, postag_vec_30, token_vec_300) — pinyin_vec_300 and word_vec_300 aren't in the repo despite being the two largest vocabularies (146k and 673k entries), so the headline word vectors aren't actually included. The dependency and POS-tag embeddings are weak by the author's own admission in the README ('或许有更好的解释') — the similarity scores for things like ATT/COO/SBV read more like noise than usable structure.