|
Title:
|
USING SEMANTIC RELATIONSHIPS TO ENHANCE
NEURAL WORD EMBEDDINGS |
|
Author(s):
|
Yanqin Yin, Xiaodong Sun, Huanhuan Lv, Pikun Wang, Hongwei Ma and Dongqiang Yang |
|
ISBN:
|
978-989-8704-23-8 |
|
Editors:
|
Pedro Isaías |
|
Year:
|
2020 |
|
Edition:
|
Single |
|
Keywords:
|
Neural Network Word Embeddings, Semantic Relationship, Semantic Similarity |
|
Type:
|
Short |
|
First Page:
|
150 |
|
Last Page:
|
154 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
click to dowload 
|
|
Paper Abstract:
|
Neural language models have significantly improved current natural language understanding tasks. However,
distributional semantics, derived from neural language models is less competitive in computing semantic relatedness or
similarity than other taxonomy-based methods. Although current researches seek to exploit the handcrafted semantic
knowledge in ontology to improve distributional semantics, they often ignore distinguishing different functions of
semantic relationships in updating or retrofitting neural word embeddings. This paper proposes retrofitting neural word
embedding through semantic relationships encoded in semantic networks such as WordNet and Roget's thesaurus. We
employ the hypernym/hyponym relationships to modify the asymmetric distance measure in retrofitting neural
embeddings, which can fully transfer the hierarchical semantic information contained in semantic networks. In the
evaluation with the gold-standard data sets, our method achieved the Spearman correlation value of 0.80, which is about
8% higher than the state-of-the-art methods in the literature. |
|
|
|
|
|
|
