Instruction-following language models often show undesirable biases. These undesirable biases may be accelerated in the real-world usage of language models, where a wide range of instructions is used through zero-shot example prompting. To solve this problem, we first define the bias neuron, which significantly affects biased outputs, and prove its existence empirically. Furthermore, we propose a novel and practical bias mitigation method, CRISPR, to eliminate bias neurons of language models in instruction-following settings. CRISPR automatically determines biased outputs and categorizes neurons that affect the biased outputs as bias neurons using an explainability method. Experimental results demonstrate the effectiveness of our method in mitigating biases under zero-shot instruction-following settings without losing the model’s task performance and existing knowledge. The experimental results reveal the generalizability of our method as it shows robustness under various instructions and datasets. Surprisingly, our method can mitigate the bias in language models by eliminating only a few neurons (at least three).

In this paper, we mainly discuss about our submission to MultiDoc2Dial task, which aims to model the goal-oriented dialogues grounded in multiple documents. The proposed task is split into grounding span prediction and agent response generation. The baseline for the task is the retrieval augmented generation model, which consists of a dense passage retrieval model for the retrieval part and the BART model for the generation part. The main challenge of this task is that the system requires a great amount of pre-trained knowledge to generate answers grounded in multiple documents. To overcome this challenge, we adopt model pretraining, fine-tuning, and multi-task learning to enhance our model’s coverage of pretrained knowledge. We experimented with various settings of our method to show the effectiveness of our approaches.