Reducing Negative Effects of the Biases of Language Models in Zero-Shot Setting.

Pre-trained language models (PLMs) such as GPTs have been revealed to be biased towards certain target classes because of the prompt and the model's intrinsic biases. In contrast to the fully supervised scenario where there are a large number of costly labeled samples that can be used to fine-tune model parameters to correct for biases, there are no labeled samples available for the zero-shot setting. We argue that a key to calibrating the biases of a PLM on a target task in zero-shot setting lies in detecting and estimating the biases, which remains a challenge. In this paper, we first construct probing samples with the randomly generated token sequences, which are simple but effective in detecting inputs for stimulating GPTs to show the biases; and we pursue an in-depth research on the plausibility of utilizing class scores for the probing samples to reflect and estimate the biases of GPTs on a downstream target task. Furtherly, in order to effectively utilize the probing samples and thus reduce negative effects of the biases of GPTs, we propose a lightweight model Calibration Adapter (CA) along with a self-guided training strategy that carries out distribution-level optimization, which enables us to take advantage of the probing samples to fine-tune and select only the proposed CA, respectively, while keeping the PLM encoder frozen. To demonstrate the effectiveness of our study, we have conducted extensive experiments, where the results indicate that the calibration ability acquired by CA on the probing samples can be successfully transferred to reduce negative effects of the biases of GPTs on a downstream target task, and our approach can yield better performance than state-of-the-art (SOTA) models in zero-shot settings.