SEAVER: Attention Reallocation for Mitigating Distractions in Language Models for Conditional Semantic Textual Similarity Measurement

ACL ARR 2024 April Submission314 Authors

Abstract: Conditional Semantic Textual Similarity (C-STS) introduces specific limiting conditions to the traditional Semantic Textual Similarity (STS) task, posing challenges for STS models. Language models employing cross-encoding demonstrate satisfactory performance in STS, yet their effectiveness significantly diminishes in C-STS. In this work, we argue that the failure is due to the fact that the redundant information in the text distracts language models from the required condition-relevant information. To alleviate this, we propose Self-Augmentation via Self-Reweighting (SEAVER), which, based solely on models' internal attention and without the need for external auxiliary information, adaptively reallocates the model's attention weights by emphasizing the importance of condition-relevant tokens. On the C-STS-2023 test set, SEAVER consistently improves performance of all million-scale fine-tuning baseline models (up to around 3 points), and even surpasses performance of billion-scale few-shot prompted large language models (such as GPT-4). Our code is available at https://github.com/NLP-LEE/SEAVER.

Paper Type: Long

Research Area: Semantics: Lexical and Sentence-Level

Research Area Keywords: semantic textual similarity, phrase/sentence embedding

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 314