Guided Generation and Fine-Tuning LLMs to Mitigate Label Biases

In this article, the authors explore the challenges of fine-tuning large language models (LLMs) for text classification tasks, particularly when dealing with a large number of labels. The authors discuss the issue of majority label bias, where the model tends to produce more instances of the most common label, and propose a solution called Guided Generation, which uses finite-state machine-based vocabulary masking to allow controlled generation from LLMs. They also examine the robustness of fine-tuning LLMs using Parameter-Efficient Finetuning Techniques (PEFT) and find that PEFT can improve the model’s performance on text classification tasks.
The authors then turn their attention to addressing the issue of data bias in text classification, particularly in the context of COVID-19 misinformation. They use a dataset of 2400 tweets classified into three categories – Misinformation, Counter-misinformation, and Irrelevant – to evaluate the performance of LLMs on different settings, such as prepending task-specific instructions in the prompt or omitting explicit instructions. The results show that incorporating Guided Generation can improve the model’s performance on the test set, particularly when dealing with a large number of labels.
The article concludes by highlighting the importance of addressing data bias in text classification tasks and providing a comprehensive understanding of the challenges involved. The authors emphasize the need for further research to overcome these challenges and improve the accuracy of LLMs on text classification tasks.