Dive into the Chasm: Probing the Gap between In- and Cross-Topic Generalization

Anonymous

Abstract: Pre-trained Language Models (PLMs) excel in downstream tasks but may struggle when faced with limited training data. Cross-Topic experiments simulate these scenarios by withholding data from distinct topics and often expose significant performance gaps compared to experiments, where instances are randomly chosen for training and evaluation (In-Topic). Transferring superior performance between these two setups is not always feasible. To better understand this generalization gap, we propose a set of probing-based experiments and analyze various PLMs.We show that this generalization gap already exists after pre-training and differs amongst PLMs and tasks. We found that pre-training objectives and architectural regularization are keys for more generalizable and robust PLMs. In addition, we consider a set of Large Language Models (LLMs) to bootstrap the analysis of other training paradigms like prompt-based learning. Ultimately, our research leads to a better understanding of PLMs and guides in selecting appropriate models or building more robust models.

Paper Type: long

Research Area: Interpretability and Analysis of Models for NLP