When Informal Text Breaks NLI: Tokenization Failure, Distribution Shift, and Targeted Mitigations

1. Problem Statement

Natural Language Inference (NLI) models trained on standard benchmarks like SNLI and MultiNLI achieve high accuracy (>89%) on formal, edited text. However, these models often fail when exposed to informal surface forms common in real-world user-generated content (e.g., social media, chat).

The paper investigates how three specific types of informal transformations degrade NLI performance:

1. Slang Substitution: Replacing formal words with informal equivalents (e.g., going to $\to$ gonna, friend $\to$ homie).
2. Emoji Replacement: Replacing content words (nouns/verbs) with Unicode emoji characters.
3. Gen-Z Noise Tokens: Appending filler/affirmation tokens that carry no propositional content (e.g., no cap, deadass, tbh).

The core hypothesis is that accuracy drops in these scenarios reveal a model's dependence on specific surface patterns rather than robust semantic understanding.

2. Methodology

Datasets and Models

• Datasets: SNLI (550k train) and MultiNLI (393k train).
• Models:
  • ELECTRA-small: 14M parameters, WordPiece tokenizer (30k vocab).
  • RoBERTa-large: 355M parameters, BPE tokenizer (50k byte-level vocab).
• Baseline: Zero-shot GPT-4o-mini and GPT-3.5-turbo.

Transforms

Transforms were applied only to the evaluation sets to measure robustness.

• Slang: 100% probability substitution.
• Emoji: Content words mapped to semantically related emoji (many-to-one mapping).
• Noise: One noise token appended to the hypothesis.
• Combined: All three applied simultaneously.

Mitigation Strategies

The authors tested three approaches to recover accuracy:

1. Augmentation: Training on data where examples are randomly transformed (50% probability) to teach the model invariance to surface forms.
2. Preprocessing (Inference-time): Normalizing text before tokenization:
   • Expanding slang to formal equivalents.
   • Converting emoji back to text labels (imperfect due to many-to-one mapping).
   • Filtering out noise tokens.
3. Hybrid: Combining both augmentation (training) and preprocessing (inference).

Evaluation

• Fixed evaluation sets generated prior to training to ensure valid paired statistical tests (McNemar's test with Bonferroni correction).
• Comparison against zero-shot LLMs.

3. Key Findings & Results

Differential Failure Modes

The paper identifies that Emoji and Noise cause failures through fundamentally different mechanisms, while Slang has negligible impact.

Transform	Mechanism of Failure	Tokenization Impact	Accuracy Drop (ELECTRA/SNLI)
Slang	Minimal. Vocabulary is mostly in-coverage.	0% [UNK] tokens.	~1.1 pp (negligible)
Emoji	Tokenization Failure. Emoji maps to [UNK] in WordPiece, destroying input signal before the encoder sees it.	93.6% of examples contain [UNK].	8.1 pp
Noise	Distribution Shift. Tokens are in-vocabulary but absent from NLI training data. The model incorrectly assigns inferential weight to them.	0% [UNK] tokens.	9.0 pp

• Error Signatures:
  • Emoji errors collapse predictions toward Neutral (loss of signal).
  • Noise errors push predictions toward Contradiction (spurious correlation).

Effectiveness of Mitigations

The success of a mitigation strategy depends entirely on the failure mode:

• Preprocessing is superior for Emoji because it restores the text signal before tokenization. Augmentation cannot fix data lost to [UNK].
• Augmentation is superior for Noise because it teaches the model to ignore tokens it has never seen in an NLI context. Preprocessing filters cannot generalize to novel noise tokens.
• Hybrid Approach: Combining both yields the best results across all conditions.
  • ELECTRA-small on SNLI: Hybrid recovers accuracy from 75.88% (baseline) to 88.93% on the combined variant.
  • Clean Text: No statistically significant drop in accuracy on original text.

Model Scale Effects

• RoBERTa-large is more robust to Noise (drop of 3.4pp vs. 9.0pp for ELECTRA) likely due to exposure to informal text in its larger pretraining corpus.
• RoBERTa-large is not more robust to Emoji (drop of 8.0pp vs. 8.1pp for ELECTRA). Even a 25x larger model cannot overcome tokenization failure if the tokenizer maps input to [UNK] or fragments the signal excessively.
• Training Stability: Large models (RoBERTa) trained with augmentation required a warmup schedule to prevent training collapse (random guessing), whereas smaller models (ELECTRA) did not.

Comparison with LLMs

• Unmitigated: Fine-tuned ELECTRA is significantly worse than GPT-4o-mini on transformed variants ($p < 0.0001$), despite being better on clean text.
• Mitigated (Hybrid): The 14M-parameter Hybrid ELECTRA surpasses GPT-4o-mini on all SNLI variants and reaches statistical parity on MultiNLI. This demonstrates that robustness is a function of training strategy, not just model capacity.

4. Key Contributions

1. Mechanistic Diagnosis: The paper distinguishes between Tokenization Failure (Emoji $\to$ [UNK]) and Distribution Shift (Noise $\to$ spurious inference), showing they require different fixes.
2. Targeted Mitigation Framework: It demonstrates that a "one-size-fits-all" approach fails; the optimal strategy is a Hybrid of inference-time preprocessing (for tokenization issues) and training-time augmentation (for distribution shifts).
3. Efficiency vs. Capacity: It proves that a small model (14M) with the correct robustness training can outperform a frontier LLM (GPT-4o-mini) on informal text, challenging the assumption that larger models are inherently more robust to surface variation.
4. Practical Guidelines:
   • Use Preprocessing for emoji/Unicode issues.
   • Use Augmentation for slang/noise issues.
   • Use Warmup when applying augmentation to large models to ensure training stability.

5. Significance

This work highlights that the "messiness" of real-world text is not just a data quality issue but a structural pipeline failure. By identifying that emoji breaks the tokenizer and noise breaks the semantic prior, the authors provide a blueprint for deploying NLI systems in production environments (e.g., social media moderation, customer support chat) where informal language is the norm. The findings suggest that robustness is a training choice, achievable through targeted interventions rather than requiring massive scale or retraining from scratch.