When Informal Text Breaks NLI: Tokenization Failure, Distribution Shift, and Targeted Mitigations
Authors: Avinash Goutham Aluguvelly
Organizations: University of Texas at Austin · Austin, TX, USA
Abstract
We study how informal surface forms degrade NLI accuracy in ELECTRA-small (14M) and RoBERTa-large (355M) across four transforms applied to SNLI and MultiNLI: slang substitution, emoji replacement, Gen-Z filler tokens, and their combination. Slang substitution (replacing formal words with informal equivalents, e.g., "going to" -> "gonna", "friend" -> "homie") causes minimal degradation (at most 1.1pp): slang vocabulary falls largely within WordPiece coverage, so the tokenizer handles it without signal loss. Emoji replaces content words with Unicode characters that ELECTRA's WordPiece tokenizer maps to [UNK], destroying the input signal before any learned parameters see it (93.6% of emoji examples contain at least one [UNK], mean 2.91 per example). Noise tokens (no cap, deadass, tbh) are fully in-vocabulary but absent from NLI training data, consistent with the model assigning them inferential weight they do not carry. The two failure modes respond to different interventions: preprocessing recovers emoji accuracy by normalizing text before tokenization; augmentation handles noise by exposing the model to noise-bearing examples during training. A hybrid of both achieves 88.93% on the combined variant for ELECTRA on SNLI (up from 75.88%), with no statistically significant drop on clean text. Against GPT-4o-mini zero-shot, unmitigated ELECTRA is significantly worse on transformed variants (p < 0.0001); hybrid ELECTRA surpasses it across all SNLI variants and reaches statistical parity on MultiNLI.
Neural NLI models overfit dataset artifacts instead of truly reasoning. A hypothesis-only model gets 57.7% in SNLI, showing strong spurious correlations, and 38.6% of the baseline errors are the result of these artifacts. We propose Product-of-Experts (PoE) training, which downweights examples where biased models are overconfident. PoE nearly preserves accuracy (89.10% vs. 89.30%) while cutting bias reliance by 4.71% (bias agreement 49.85% to 45%). An ablation finds lambda = 1.5 that best balances debiasing and accuracy. Behavioral tests still reveal issues with negation and numerical reasoning.
We introduce Telegraph English (TE), a prompt-compression protocol that rewrites natural language into a symbol-rich, formally-structured dialect. Where token-deletion methods such as LLMLingua-2 train a classifier to delete low-importance tokens at a fixed ratio, TE performs a full semantic rewrite: it decomposes the input into atomic fact lines, substitutes verbose phrases with ∼40 logical and relational symbols, and lets the compression ratio adapt to each document's information density. A consequence of the line-structure rule is that compression and semantic chunking become the same operation -- each output line is an independently addressable fact, so the compressed representation is simultaneously a semantic index. We evaluate TE on 4{,}081 question-answer pairs from LongBench-v2 across five OpenAI models and two difficulty levels. At roughly 50% token reduction, TE preserves 99.1% accuracy on key facts with GPT-4.1 and outperforms LLMLingua-2 at matched compression ratios on every model and task tested. The gap widens on smaller models -- up to 11 percentage points on fine-detail tasks -- suggesting that explicit relational structure compensates for limited model capacity. We release the grammar specification, compression prompt, benchmark data, and reference implementation.
Large language models are increasingly used to annotate datasets for training smaller, task-specialized models such as named entity recognition. While this method yields effective models, it assumes that the synthetic dataset is correctly annotated. In this work, we find that (i) current fine-tuning processes simply ignore LLM-introduced annotation noise, resulting in degraded performance and (ii) existing noise-robust losses are not transferable to sequence labeling because annotation noise in named entity recognition is heterogeneous: for example, missing mentions and type errors affect the training signal in different ways. Treating all noisy tokens equally in noise-robust losses and applying a single reweighing criterion for all may therefore remove useful supervision or reinforce incorrect labels. To address this limitation, we propose error-type-aware loss reweighting for NER, which introduces separate reweighing rules for different types of potentially erroneous tokens. Our approach is simple and efficient, does not require additional training resources, and improves F1 by 0.8 - 2.0 percentage points on dataset-level average for noise levels between 15% and 40%, with a maximum improvement of 4.6 percentage points with 24.1% noise on Wikigold.