Less Likely Brainstorming: Using Language Models to Generate Alternative Hypotheses
This addresses bias in human decision-making for domains like radiology and commonsense reasoning, though it is incremental as it builds on existing controlled text generation methods.
The paper tackles the problem of AI assistants reinforcing human biases by only predicting likely outcomes, introducing a 'less likely brainstorming' task to generate relevant but less likely hypotheses. They propose a contrastive learning method for controlled text generation, showing improved capability in generating less likely outputs compared to baselines.
A human decision-maker benefits the most from an AI assistant that corrects for their biases. For problems such as generating interpretation of a radiology report given findings, a system predicting only highly likely outcomes may be less useful, where such outcomes are already obvious to the user. To alleviate biases in human decision-making, it is worth considering a broad differential diagnosis, going beyond the most likely options. We introduce a new task, "less likely brainstorming," that asks a model to generate outputs that humans think are relevant but less likely to happen. We explore the task in two settings: a brain MRI interpretation generation setting and an everyday commonsense reasoning setting. We found that a baseline approach of training with less likely hypotheses as targets generates outputs that humans evaluate as either likely or irrelevant nearly half of the time; standard MLE training is not effective. To tackle this problem, we propose a controlled text generation method that uses a novel contrastive learning strategy to encourage models to differentiate between generating likely and less likely outputs according to humans. We compare our method with several state-of-the-art controlled text generation models via automatic and human evaluations and show that our models' capability of generating less likely outputs is improved.