Year: 2025
Type: conference-paper
Abstract: Unsupervised contrastive learning models typically treat samples from the same language as positive pairs, which can lead to language bias—causing the model to assign higher semantic similarity scores to sentences in the same language. To address this issue, this paper proposes a multilingual sentence similarity computation method based on IS-CSE. First, independent language tags are explicitly added to sentences, enabling the model to perceive linguistic characteristics and enhance the consistency of multilingual sentence representations. Then, a dynamic memory buffer mechanism is introduced to reduce computational overhead while maintaining a balanced distribution of languages within the buffer, preventing any single language from dominating. Additionally, a language enhancement coefficient is incorporated to adjust the weighting of different languages in generating smooth sentence embeddings. Finally, experiments on the Tatoeba and BUCC datasets demonstrate the effectiveness of the proposed method in cross-lingual sentence retrieval, particularly in scenarios without parallel corpora, where it significantly improves sentence semantic similarity computation. The results show that the proposed model exhibits superior adaptability and robustness in multilingual tasks, providing an effective solution for semantic computation in low-resource languages.
Authors: Degen Huang, Yiquan Wang
Institution: Dalian University of Technology
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413812568