dalinvip/cw2vec

Implements stroke n-gram composition alongside classical word2vec variants (skipgram, CBOW) and fastText subword approaches, with the substoke model leveraging Chinese character stroke sequences as compositional features. Built in C++ with CMake, it outputs dual embeddings—context vectors and stroke n-gram averages—and includes pre-extracted stroke features for 20,901 simplified Chinese characters. Evaluated on Chinese word similarity benchmarks, demonstrating improved performance over character-agnostic baselines.

274 stars. No commits in the last 6 months.