Year: 2024
Type: conference-paper
Abstract: The lack of negative samples is common in NLP tasks, such as essay scoring and model essay recommendation in English intelligent assistant teaching systems.Bootstrap Contrastive Learning Frameworks (BCLF) use only positive samples for training, which is suitable for scenarios lacking negative samples and reduces training costs.BCLF has been proven effective in computer vision (CV) related tasks.However, training BCLFs is challenging in NLP, and there is little related work.In this paper, we introduce some BCLFs from the CV domain into the NLP domain, making adjustments to enable them to handle sentence representation task.Furthermore, we evaluate, compare, and analyze these frameworks on the semantic text similarity tasks.Based on the analysis results, we propose a new framework called SimBarlow, leveraging the advantages of different kinds of BCLFs, which greatly surpasses previous uncontrastive learning-based baselines and is competitive with traditional contrastive learning frameworks.Experimental results show that our framework achieved better performance compared to the baseline bootstrapping frameworks on STS tasks.Additionally, we demonstrate how to build an essay recommendation system using BCLF, further validating the effectiveness of the proposed SimBarlow.
Author: Li Li
Institution: Shanghai University of Political Science and Law
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: bronze
OpenAlex ID: W4406255692