Dive into the Chasm: Probing the Gap between In- and Cross-Topic Generalization
This work addresses the problem of cross-topic generalization for language model users, providing insights that are incremental but relevant for improving model robustness.
The study investigated the gap in generalization performance of pre-trained language models between in-topic and cross-topic scenarios, finding that factors like diverse pre-training objectives and architectural regularization reduce this gap and enhance model robustness.
Pre-trained language models (LMs) perform well in In-Topic setups, where training and testing data come from the same topics. However, they face challenges in Cross-Topic scenarios where testing data is derived from distinct topics -- such as Gun Control. This study analyzes various LMs with three probing-based experiments to shed light on the reasons behind the In- vs. Cross-Topic generalization gap. Thereby, we demonstrate, for the first time, that generalization gaps and the robustness of the embedding space vary significantly across LMs. Additionally, we assess larger LMs and underscore the relevance of our analysis for recent models. Overall, diverse pre-training objectives, architectural regularization, or data deduplication contribute to more robust LMs and diminish generalization gaps. Our research contributes to a deeper understanding and comparison of language models across different generalization scenarios.