A Negative Sample Enhancement Strategy to Improve Contrastive Learning for Unsupervised Sentence Representation

Contrastive learning has achieved remarkable success in sentence representation research within the field of natural language processing. Nevertheless, most existing studies focus primarily on the construction of negative samples while paying insufficient attention to the mechanisms for handling these samples. Such methods tend to treat all negative samples within a batch as equally important, neglecting the crucial role that negative samples play in semantic learning. This oversight can result in suboptimal model performance in semantic understanding. To address these issues, this study proposes a negative sample enhancement strategy that applies fine-grained processing to different types of constructed negative samples based on their importance. In the high-dimensional semantic space, hard negative samples and false negative samples are treated respectively—by increasing the distance between hard negative samples and positive samples, while treating false negative samples as pseudo-positive samples to enhance the attraction between them and the anchor sample. This strategy enables the model to perform more effective semantic differentiation and representation. Experimental results on the Semantic Textual Similarity (STS) task demonstrate that the proposed method outperforms existing baseline methods in unsupervised sentence representation learning.