Exploring Imbalanced Annotations for Effective In-Context Learning

Hongfu Gao, Feipeng Zhang, Hao Zeng, Deyu Meng, Bingyi Jing, Hongxin Wei

Keywords: Imbalanced Annotations; In-Context Learning; Demonstration Selection; Large Language Models

TL;DR: The paper understands and mitigates the effect of imbalanced annotations on the performance of in-context learning.

Abstract: Large language models (LLMs) have shown impressive performance on downstream tasks through in-context learning (ICL), which heavily relies on the quality of demonstrations selected from a large annotated dataset. However, real-world datasets often exhibit long-tailed class distributions, where a few classes occupy most of the data while most classes are under-represented. In this work, we show that imbalanced annotations hurt the ICL performance by degrading the Task Learning ability and cannot be mitigated by varying the demonstration sets, selection methods, calibration methods and rebalancing methods. To circumvent the issue, we propose a simple and effective approach termed Reweighting with Importance Factors (dubbed RIF) to enhance ICL performance under class imbalance. In particular, RIF constructs a balanced subset to estimate importance factors for each class: the ratio between the joint distribution of demonstration sets selected from balanced and imbalanced datasets. Then, we leverage the factors to re-weight the scoring function (e.g., the cosine similarity score used in TopK) during demonstration selection. In effect, RIF prevents over-selection from dominant classes while preserving the efficacy of current selection methods. Extensive experiments on common benchmarks demonstrate the effectiveness of our method, improving the average accuracy of current selection methods by up to 5.60\%.

Primary Area: generative models

Submission Number: 1961