Khái niệm cốt lõi
The author proposes the Clean-LaVe framework to leverage silver standard data for improved zero-shot classification performance by detecting clean data and finetuning pre-trained models. The approach involves Iteratively Weighted Negative Learning and Class-Aware Data Selector to address noisy data and broaden class selection.
Tóm tắt
The study introduces Clean-LaVe, a framework enhancing zero-shot classification tasks by utilizing silver standard data. It includes phases like obtaining silver data, identifying clean data, finetuning models, and inferring on test data. Experimental results show significant performance improvements across various datasets.
Recent works have converted zero-shot tasks to other NLP tasks using off-the-shelf models, generating large-scale silver standard data. However, these datasets are often underutilized due to noise issues. The proposed Clean-LaVe framework aims to address this challenge by detecting clean data and improving model performance through finetuning.
The study highlights the importance of leveraging pre-trained models as low-cost annotators to produce valuable silver standard data. By introducing novel algorithms like IWNL and CADS, Clean-LaVe demonstrates superior performance compared to traditional noisy label learning methods.
Overall, the research showcases the effectiveness of Clean-LaVe in enhancing zero-shot classification tasks through innovative approaches to handle noisy labels and improve model training with silver standard data.
Thống kê
Clean-LaVe outperforms baseline by 5% - 8% on various datasets.
Clean-LaVe shows improvement of 3% - 7% on cross-lingual relation classification tasks.
The framework achieves an 8% increase in event argument classification task.
Silver-LaVe outperforms LaVeEntail by 2% - 13%.
Trích dẫn
"The experimental results demonstrate that Clean-LaVe can outperform the baseline by 5% - 8% on various datasets."
"Our contributions are summarized as follows: We propose Clean-LaVe to first detect a small amount of clean data which are later used to fine-tune the pre-trained model."
"Clean-LaVe is a general framework that can be used in scenarios where a pre-trained model serves as an annotator."