Evaluation of Word Embedding via Domain Keywords

Word embeddings, unsupervisedly learned, have proven to be very effective and provide semantic and syntactic information in most NLP tasks. Most common intrinsic evaluations of word embeddings use the similarity of words as core. Notwithstanding, these frequently correspond inadequately with how well the word embeddings perform as features in actual downstream tasks. We present VECDS (Vector Domain Score) based on the corresponding domain keywords, like high frequency or extracted by human, in downstream evaluation tasks. The domain keywords is more important for downstream than other common vocabulary.