Abstract
Junghyun Min, R. Thomas McCoy, Dipanjan Das, Emily Pitler, Tal Linzen. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.
2020.
Introduction
State-of-the-art NLP models like BERT often achieve human-level accuracy on standard benchmarks such as Multi-Genre Natural Language Inference (MNLI). However, they frequently fail on simple logic tests, revealing a reliance on "inference heuristics"—lazy shortcuts rather than true linguistic understanding. For instance, the HANS (Heuristic Analysis for NLI Systems) challenge set shows that models often predict "entailment" simply because the premise and hypothesis contain the same words (lexical overlap). This leads to errors like concluding that "The lawyer saw the actor" entails "The actor saw the lawyer."
This research investigates whether this failure stems from Representational Inadequacy (BERT cannot represent necessary syntax) or a Missed Connection (BERT has the info but hasn't learned it's crucial for NLI because the training data is too "easy").
Methods: Syntactic Data Augmentation
The authors propose Syntactic Data Augmentation to force models to attend to syntax. They applied automated transformations to MNLI sentences to create new training pairs that "break" common heuristics:
- Inversion: Swapping subject and object (e.g., "The cat chased the dog" $\rightarrow$ "The dog chased the cat").
- Passivization: Converting active voice to passive (e.g., "The cat chased the dog" $\rightarrow$ "The dog was chased by the cat").
They tested various augmentation strategies, including keeping the original premise while transforming the hypothesis, and using the original hypothesis as a new premise. These were compared against a random shuffling control to ensure the model was learning syntax, not just word order awareness.
Key Results and Insights
The results strongly support the Missed Connection Hypothesis. Adding as few as 400 syntactically augmented examples (roughly 0.1% of the training data) led to a massive jump in robustness without any performance hit on the standard MNLI test set.
- The "Nudge" Effect: As shown in Figure 2, a small amount of data causes a significant jump in accuracy for cases where heuristics fail.
- Robustness Boost: On HANS cases where the lexical overlap heuristic fails (non-entailment), BERT's accuracy jumped from 0.28 to 0.73 with the large inversion augmentation.
- Abstract Generalization: Training on subject/object swaps improved performance on unrelated structures like prepositional phrases, suggesting the model learned a general principle to prioritize connectivity over simple word matching.
Discussion and Limitations
While augmentation was highly effective for some heuristics, simple data augmentation is not a silver bullet for all syntactic weaknesses.
- The Passive Wall: Interestingly, augmentation did not significantly help with passive sentences (see Figure 1, Passivization row). This suggests a potential case of Representational Inadequacy where BERT struggles to deeply encode the relationship between active and passive forms.
- Heuristic Specificity: While the model improved on lexical overlap, it still struggled with the "subsequence" heuristic (assuming a sentence entails any of its sub-parts), as seen in the middle panel of Figure 2.
- Future Directions: The authors suggest that instead of just augmenting data, we should look for models with stronger inductive biases—architectures that are "born" to prioritize syntax.
Summary updated