Contrastive Learning to rank with Weak Supervision
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: gpt
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Fine-tuned pre-trained contextual word embedding models to supervised downstream tasks have become commonplace in natural language processing. However, modern deep learning natural language processing (NLP) tasks require a large amount of labeled data. We study the problem of fine-tuned pre-trained language models without using any labeled data, but only weak supervision sources. This problem is technically challenging due to the labels generated by weak supervision being noisy, it may cause over-fitting, and is inapplicable for construction labels in a continuous sample space. To address these two challenges, we design a contrastive learning framework with weak supervision, CLWS, to enable modeling language models. The unlabeled data is modeled according to the partial order relationship of weak supervision signal, which estimates the source reliability by self-attention mechanism and then reduces label noise by fusing weak labels. The discriminant model aligned with the input pair in continuous space, using the generated labels with contrastive learning to address the above issue. Experiments on multi-turn dialogue, search ranking tasks show that our model outperforms the strongest baseline by large margins and achieves competitive performance with fully-supervised methods.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1039
Loading