Research on Sentence Embeddings for Text Matching through Multiview Interactive Features
Xiaohui Liu, Fumin Chen, Yong Le Hu, Xinjian Li · 2023
In this paper, we propose a sentence embedding for text matching method through multi-view interactive features: SBERT-Inter. This method is improved from the Sentence-BERT(SBERT) model based on semantic matching. Existing text matching models including representation-based methods get the vector representation of two sentences and calculate the similarity, but the interaction between sentences is lost. Or based on the interactive method to splice two sentences and directly output the similarity, but not through the sentence representation. Aiming at the problems of the two methods, a text matching method based on representation and interaction is proposed, which combines more fine-grained interaction modules after obtaining the representation features, through the multi-view and multi-perspective fusion of sentence and token features such as embedding difference, embeddings similarity, and token's self-attention, we can better learn the fine-grained interactive features between embeddings and we can get better semantic representation. It improves the matching performance between sentence pairs. On the public dataset semantic textual similarity benchmark(STS-B), the Spearman correlation of this method is enhanced by 6% compared with other sentence embedding methods.