We're Calling an Intervention: Exploring Fundamental Hurdles in Adapting Language Models to Nonstandard Text
Aarohi Srivastava, David Chiang
TL;DR
The paper investigates how language models adapt to nonstandard text by designing nine controlled interventions that disrupt tokenization and embeddings across character, subword, morphological, and lexical levels. Using BERT and LoRA with a mask-filling objective on data derived from Wikicorpus, the study systematically varies data size and composition and compares monolingual and multilingual models. Key findings show that character-level within-subword changes remain difficult to learn even with more data, while lexical and morphological variations can be learned with sufficient data, and multilingual models offer advantages for meaning-related shifts. The work highlights fundamental limitations in current tokenization and representation schemes and points toward the need for more flexible tokenization and robustness techniques to handle diverse user-generated text in practical NLP tasks.
Abstract
We present a suite of experiments that allow us to understand the underlying challenges of language model adaptation to nonstandard text. We do so by designing interventions that approximate core features of user-generated text and their interactions with existing biases of language models. Applying our interventions during language model adaptation to nonstandard text variations, we gain important insights into when such adaptation is successful, as well as the aspects of text variation and noise that are particularly difficult for language models to handle. For instance, on text with character-level variation, out-of-the-box performance improves even with a few additional training examples but approaches a plateau, suggesting that more data is not the solution. In contrast, on text with variation involving new words or meanings, far more data is needed, but it leads to a massive breakthrough in performance. Our findings reveal that existing models lack the necessary infrastructure to handle diverse forms of nonstandard text, guiding the development of more resilient language modeling techniques. We make the code for our interventions, which can be applied to any English text data, publicly available.