Multi-Level Positive Sample Constrained Unsupervised Sentence Representation

Unsupervised sentence representation aims to learn the spatial structure of sentence embeddings from large-scale text data. The current work uses contrastive learning to constrain the cosine distance between different sentence vectors via the loss function built upon positive and negative examples. To better capture the continuous real number of semantic similarity, this paper proposes the maximum entropy list sorting loss to learn the distance between different positive examples by utilizing the dependency syntax tree and the key phrase extraction algorithm to construct positive examples with different degrees of similarity for each sentence. Experiments on the Semantic Text Similarity datasets show that the proposed method achieves an average Spearman coefficient of 79.79%, which is 1.03% higher than the current state-of-the-art baselines.