Length is a Curse and a Blessing for Document-level Semantics

Xiao, Chenghao; Li, Yizhi; Hudson, G Thomas; Lin, Chenghua; Al Moubayed, Noura

Abstract

In recent years, contrastive learning (CL) has been extensively utilized to recover sentence and document-level encoding capability from pre-trained language models. In this work, we question the length generalizability of CL-based models, i.e., their vulnerability towards length-induced semantic shift. We verify not only that length vulnerability is a significant yet overlooked research gap, but we can devise unsupervised CL methods solely depending on the semantic signal provided by document length. We first derive the theoretical foundations underlying length attacks, showing that elongating a document would intensify the high intra-document similarity that is already brought by CL. Moreover, we found that isotropy promised by CL is highly dependent on the length range of text exposed in training. Inspired by these findings, we introduce a simple yet universal document representation learning framework, LA(SER) 3 : length-agnostic self-reference for semantically robust sentence representation learning, achieving state-of-the-art unsupervised performance on the standard information retrieval benchmark. Our code is publicly available.

Citation

Xiao, C., Li, Y., Hudson, G. T., Lin, C., & Al Moubayed, N. (2023). Length is a Curse and a Blessing for Document-level Semantics.

• Presentation Conference Type: Conference Paper (Published)
• Conference Name: The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)
• Start Date: Dec 6, 2023
• End Date: Dec 10, 2023
• Acceptance Date: Nov 9, 2023
• Publication Date: 2023
• Deposit Date: Nov 23, 2023
• Publicly Available Date: Dec 8, 2023
• Pages: 1385-1396
• Public URL: https://durham-repository.worktribe.com/output/1948282
• Publisher URL: https://aclanthology.org/venues/emnlp/