Transformer Language Models Handle Word Frequency in Prediction Head

2023/05/29 by Goro Kobayashi, Tatsuki Kuribayashi, Kobayashi, Goro +5 · 1 citation
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Speech Recognition and Synthesis #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2305.18294

openalex publication_date 2023/05/29 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

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.

Cited by

• Next-Token Prediction Should be Ambiguity-Sensitive: A Meta-Learning Perspective

Related

• Explaining How Transformers Use Context to Build Predictions
• Transformers are Universal Predictors
• Frequency Explains the Inverse Correlation of Large Language Models' Size, Training Data Amount, and Surprisal's Fit to Reading Times
• Token-wise Decomposition of Autoregressive Language Model Hidden States for Analyzing Model Predictions
• Bias A-head? Analyzing Bias in Transformer-Based Language Model Attention Heads