Transformer Language Models Handle Word Frequency in Prediction Head
This work addresses the overlooked prediction head component in Transformer models, offering insights for improving text generation diversity, though it is incremental in nature.
The study investigated the role of bias parameters in Transformer language models' prediction heads, finding that they help reflect word frequency in corpora, similar to logit adjustment methods, and demonstrated that controlling these biases can increase text diversity without quality loss in generation tasks.
Prediction head is a crucial component of Transformer language models. Despite its direct impact on prediction, this component has often been overlooked in analyzing Transformers. In this study, we investigate the inner workings of the prediction head, specifically focusing on bias parameters. Our experiments with BERT and GPT-2 models reveal that the biases in their word prediction heads play a significant role in the models' ability to reflect word frequency in a corpus, aligning with the logit adjustment method commonly used in long-tailed learning. We also quantify the effect of controlling the biases in practical auto-regressive text generation scenarios; under a particular setting, more diverse text can be generated without compromising text quality.