Quantifying Adaptability in Pre-trained Language Models with 500 Tasks
This work addresses the need to systematically characterize adaptability in language models for NLP researchers, though it is incremental as it builds on existing generalization studies with a new benchmark.
The authors tackled the problem of understanding what predicts the performance of pre-trained language models when adapted to new tasks, by conducting a large-scale empirical study using a new benchmark of 500 procedurally generated tasks. They found that adaptation procedures vary in memorization ability, exhibit compositional adaptability in some task types, and that label distribution mismatches are explained by prediction difficulty, showing adaptability can be systematically described.
When a neural language model (LM) is adapted to perform a new task, what aspects of the task predict the eventual performance of the model? In NLP, systematic features of LM generalization to individual examples are well characterized, but systematic aspects of LM adaptability to new tasks are not nearly as well understood. We present a large-scale empirical study of the features and limits of LM adaptability using a new benchmark, TaskBench500, built from 500 procedurally generated sequence modeling tasks. These tasks combine core aspects of language processing, including lexical semantics, sequence processing, memorization, logical reasoning, and world knowledge. Using TaskBench500, we evaluate three facets of adaptability, finding that: (1) adaptation procedures differ dramatically in their ability to memorize small datasets; (2) within a subset of task types, adaptation procedures exhibit compositional adaptability to complex tasks; and (3) failure to match training label distributions is explained by mismatches in the intrinsic difficulty of predicting individual labels. Our experiments show that adaptability to new tasks, like generalization to new examples, can be systematically described and understood, and we conclude with a discussion of additional aspects of adaptability that could be studied using the new benchmark.