Uncovering Constraint-Based Behavior in Neural Models via Targeted Fine-Tuning
This work addresses the issue of non-human-like behavior in language models for NLP researchers, but it is incremental as it builds on existing literature about linguistic knowledge in models.
The study tackled the problem of competing linguistic processes obscuring underlying knowledge in pretrained language models, showing that targeted fine-tuning can re-weight constraints to uncover dormant linguistic knowledge and revealing cross-linguistic variation in model behavior.
A growing body of literature has focused on detailing the linguistic knowledge embedded in large, pretrained language models. Existing work has shown that non-linguistic biases in models can drive model behavior away from linguistic generalizations. We hypothesized that competing linguistic processes within a language, rather than just non-linguistic model biases, could obscure underlying linguistic knowledge. We tested this claim by exploring a single phenomenon in four languages: English, Chinese, Spanish, and Italian. While human behavior has been found to be similar across languages, we find cross-linguistic variation in model behavior. We show that competing processes in a language act as constraints on model behavior and demonstrate that targeted fine-tuning can re-weight the learned constraints, uncovering otherwise dormant linguistic knowledge in models. Our results suggest that models need to learn both the linguistic constraints in a language and their relative ranking, with mismatches in either producing non-human-like behavior.