講演情報
[P3-66]Linguistically-informed enhancements to word embedding training
*黒田 航1 (1. 杏林大学)
キーワード：
word embeddings、dynamic windowing、context lateralizaiton、frequency waves、CBoW vs Skip-gram
We introduce three linguistically-motivated enhancements to word embedding training: dynamic windowing, context lateralization, and IDF weighting. Experiments on Simple Wikipedia corpora show that these enhancements yield measurable improvements on both compositional benchmarks (SICK, STS) and lexical benchmarks (SimLex-999, WordSim- 353), with effect sizes depending on the specific combination employed. A key finding is that window size selectively affects STS while leaving SICK stable, revealing that these benchmarks measure qualitatively different semantic competences: global sentential coherence vs. local compositional structure. With appropriate configurations, CBoW becomes competitive with Skip-gram.