Neural Natural Language Inference Models Partially Embed Theories of Lexical Entailment and Negation
This work addresses the problem of compositional reasoning in NLI models for NLP researchers, showing incremental improvements in handling negation and entailment through targeted fine-tuning.
The study investigated whether neural models for Natural Language Inference (NLI) can learn compositional interactions between lexical entailment and negation, finding that models trained on general-purpose datasets fail systematically on examples with negation, but fine-tuning on the new MoNLI dataset addresses this, with probes and interventions suggesting the BERT model partially embeds a theory of these concepts at an algorithmic level.
We address whether neural models for Natural Language Inference (NLI) can learn the compositional interactions between lexical entailment and negation, using four methods: the behavioral evaluation methods of (1) challenge test sets and (2) systematic generalization tasks, and the structural evaluation methods of (3) probes and (4) interventions. To facilitate this holistic evaluation, we present Monotonicity NLI (MoNLI), a new naturalistic dataset focused on lexical entailment and negation. In our behavioral evaluations, we find that models trained on general-purpose NLI datasets fail systematically on MoNLI examples containing negation, but that MoNLI fine-tuning addresses this failure. In our structural evaluations, we look for evidence that our top-performing BERT-based model has learned to implement the monotonicity algorithm behind MoNLI. Probes yield evidence consistent with this conclusion, and our intervention experiments bolster this, showing that the causal dynamics of the model mirror the causal dynamics of this algorithm on subsets of MoNLI. This suggests that the BERT model at least partially embeds a theory of lexical entailment and negation at an algorithmic level.