Improving Word Embeddings by Emphasizing Co-hyponyms
Lists
Only you see this
Abstract
Word embeddings are powerful for capturing semantic similarity between words in a vocabulary. They have been demonstrated beneficial to various natural language processing tasks such as language modeling, part-of-speech tagging and machine translation, etc. Existing embedding methods derive word vectors from the co-occurrence statistics of target-context word pairs. They treat the context words of a target word equally, while not all contexts are created equal for a target. Some recent work learns non-uniform weights of the contexts for predicting the target, while none of them take the semantic relation types of target-context pairs into consideration. This paper observes co-hyponyms usually have similar contexts and can be substitutes of one another. To this end, this paper proposes a simple but effective method to improve word embeddings. It automatically identifies possible co-hyponyms within the context window and optimizes the embeddings of co-hyponyms to be close directly. Compared to 3 state-of-the-art neural embedding models, the proposed model performs better on several datasets of different languages in terms of the human similarity judgement and the language modeling tasks.
BibTeX
@inproceedings{cai2018improving,
author = {Xiangrui Cai and Yonghong Luo and Ying Zhang and Xiaojie Yuan},
title = {{Improving Word Embeddings by Emphasizing Co-hyponyms}},
booktitle = {WISA},
year = {2018},
doi = {10.1007/978-3-030-02934-0\_20},
}Loading related content...