Calibrating language model ordinal outputs with few labels improves accuracy
CORDIAL: Calibrating Ordinal LLM Outputs from Few Labels
Computation and LanguageMachine Learning
Summary
Large language models can predict ratings or scores on an ordered scale from text, but their raw predictions are often biased or noisy. The authors propose CORDIAL, a way to adjust these predictions using only a few labeled examples to make them more accurate and reliable. Their method works well even with just 5 to 100 labels and outperforms many other calibration techniques on Amazon reviews and spoken transcript datasets. This calibration also allows combining outputs from different models or learning from other tasks to improve performance.
What this means in practice
- •For sentiment analysis teams: Improve the accuracy of sentiment scores from language models using only a few labeled examples for calibration on customer reviews.
- •For multimodal analytics developers: Combine outputs from multiple language models for better ordinal predictions on video and audio-transcript data using calibrated posteriors.
Authors
Xiangwei Wang, Peng Wang, Saman Halgamuge
Abstract
A large language model (LLM) can turn a text into a distribution over an ordered scale, but that distribution is a noisy measurement: saturated, compressed or exaggerated, and biased in a consistent direction. We propose CORDIAL, which treats the model's output as a noisy reading of the true label and corrects it with a channel of five interpretable parameters. The channel is small enough for its posterior to be averaged from a handful of labels, and we prove that the resulting calibration preserves first-order stochastic order. On Amazon reviews and CMU-MOSEI transcripts with four LLMs, CORDIAL has the lowest log loss among nine calibrators in 76 of 80 settings with 5 to 100 labels; with 20 labels and the main 7B reader, it matches the strongest baseline using 28-54 labels. The same posterior lets us learn priors from other tasks and fuse several LLMs. Unrestricted calibrators such as Dirichlet calibration overtake it only as the calibration set grows into the hundreds or thousands.