Jongdeog Lee, Sangsu Kim, Jisoo Lee, Journal of Internet Computing and Services, Vol. 26, No. 5, pp. 39-48, Oct. 2025
10.7472/jksii.2025.26.5.39, Full Text:  HTML
Keywords: Natural Language Process, Word Embedding, Skip-gram, Sentence Similarity, dependency parsing

Abstract

Sentence similarity analysis is a fundamental task in natural language processing (NLP) with wide applications in information retrieval, text summarization, machine translation, and sentiment analysis. Word embedding techniques, such as Skip-gram in Word2Vec, have been widely used to numerically capture semantic similarity between sentences. In this study, we propose a novel word embedding approach that improves the efficiency and effectiveness of the traditional Skip-gram algorithm by incorporating head-dependent linguistic relationships. The proposed model replaces the conventional multi-context word training with a single head word per target word, significantly enhancing training efficiency without expanding the context window. We evaluated our model on Korean sentence similarity datasets, including KorSTS and KLUE-STS, and found that it achieves comparable or superior performance to conventional CBOW and Skip-gram models while requiring fewer computational resources. We also highlight the importance of improving the accuracy and speed of head word prediction as a future direction for optimizing the proposed algorithm.

Statistics
Cite this article