Source-linked AI summary
Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?
Byung-Doh Oh, William Schuler
TL;DR
Larger pre-trained Transformer language models can achieve lower perplexity while producing surprisal estimates that fit human reading times less well. This paper tests the pattern across GPT-Neo and OPT variants, then analyzes residuals to identify linguistic sources of the mismatch and relates them to memorization.
Problem
Larger pre-trained language models with more parameters and lower perplexity have been observed to provide less predictive surprisal estimates for human reading times, challenging the assumption that larger models are better.
Method
The authors evaluate five GPT-Neo and eight OPT variants on reading-time datasets using regression analyses, then analyze residual errors across linguistic phenomena.
Results
A strictly monotonic, positive log-linear relationship links perplexity to fit for GPT-Neo and OPT models; larger models especially underpredict reading times for nouns and adjectives, including named entities.
Takeaways & Limitations
Researchers should not select the largest available pre-trained language model by default and should include smaller variants as baselines when studying naturalistic sentence processing.
Takeaways & Limitations
The comparison with earlier trained-from-scratch models may involve distinct regimes because of the especially large discrepancy in model size between those models and the pre-trained models studied here.
Abstract
from arXiv · showhide
This work presents a detailed linguistic analysis into why larger Transformer-based pre-trained language models with more parameters and lower perplexity nonetheless yield surprisal estimates that are less predictive of human reading times. First, regression analyses show a strictly monotonic, positive log-linear relationship between perplexity and fit to reading times for the more recently released five GPT-Neo variants and eight OPT variants on two separate datasets, replicating earlier results limited to just GPT-2 (Oh et al., 2022). Subsequently, analysis of residual errors reveals a systematic deviation of the larger variants, such as underpredicting reading times of named entities and making compensatory overpredictions for reading times of function words such as modals and conjunctions. These results suggest that the propensity of larger Transformer-based models to 'memorize' sequences during training makes their surprisal estimates diverge from humanlike expectations, which warrants caution in using pre-trained language models to study human language processing.
1 Introduction
Although surprisal has long been used to model processing difficulty, larger pre-trained Transformer LMs can produce lower-perplexity estimates that fit human reading times worse. This work tests that pattern across model families and analyzes the linguistic sources of the mismatch.
- Surprisal operationalizes contextual predictability and has been shown to predict behavioral and neural measures of processing difficulty.
- Larger GPT-2 variants achieve lower perplexity but provide less predictive surprisal estimates for self-paced reading times and eye-gaze durations.
- The study evaluates GPT-Neo and OPT variants to test whether the larger-model pattern generalizes beyond GPT-2.
- Regression analyses find a strictly monotonic, positive log-linear relationship between perplexity and reading-time fit across five GPT-Neo variants and eight OPT variants on two datasets.
- Residual analysis shows larger models generally underpredict reading times at nouns and adjectives, with underprediction increasing alongside model size.
2 Related Work
Earlier psycholinguistic modeling generally linked lower perplexity with better prediction of human reading times, but pre-trained GPT-2 produced the opposite pattern. Related work also documents systematic underprediction of several human reading-time effects.
- Studies of n-gram, recurrent, and Transformer language models generally report that lower perplexity corresponds to better prediction of comprehension difficulty.
- Earlier work found negative relationships between perplexity and regression fit across n-gram, LSTM, Transformer, RNNG, GRU, and multilayer models.
- Oh et al. reported the contradictory result that the largest pre-trained GPT-2 variant made the smallest contribution to regression fit on Natural Stories and Dundee reading data.
- Neural language models have been reported to underpredict garden-path effects, deeply embedded-sentence delays, and other human reading-time increases.
- Some counterexamples compared models trained on different data or evaluated Japanese text with different syntactic head-directionality, limiting direct comparison with the established trend.
3 Main Experiment: Predictive Power of Language Model Surprisal Estimates
The experiment compares surprisal from GPT-2, GPT-Neo, and OPT variants against reading data from Natural Stories and Dundee using mixed-effects regression. It evaluates model perplexity and surprisal-based improvements in fit across families and capacities.
- 3 Main Experiment: Predictive Power of Language Model Surprisal Estimates: Surprisal predictors from GPT-2, five GPT-Neo, and eight OPT decoder-only Transformer variants were evaluated on Natural Stories self-paced reading and Dundee go-past durations.
- 3 Main Experiment: Predictive Power of Language Model Surprisal Estimates: The Natural Stories data include 181 subjects and 770,102 filtered observations, while Dundee includes 10 subjects and 195,507 filtered observations.
- 3 Main Experiment: Predictive Power of Language Model Surprisal Estimates: Surprisal was computed from autoregressive token probabilities, summing negative log probabilities when a word was split into multiple subword tokens.
- 3 Main Experiment: Predictive Power of Language Model Surprisal Estimates: Baseline and full linear mixed-effects models estimated each surprisal predictor’s contribution to fit using ∆LL, alongside word-level and eye-tracking controls.
- 3.4 Results: The smallest variant in each family made the biggest contribution to regression fit, while larger variants yielded strictly poorer fits to reading times.
- 3.4 Results: The three families showed a strong log-linear relationship between perplexity and ∆LL, with all slopes significantly above zero except GPT-2 on Natural Stories at p = 0.07.
- 3.4 Results: At comparable capacities, OPT generally achieved lower perplexity and more predictive surprisal than GPT-2, whereas GPT-Neo showed no improvement over OPT despite more training data.
4 Post-hoc Analysis: Linguistic Phenomena Underlying the Trend
Residual-error analyses identify linguistic subsets and prediction directions underlying the poorer fit of larger language models. Larger models especially underpredict open-class words, while coefficient scaling can produce overpredictions for some function-word subsets.
- Residual-error analysis: Residual errors were computed from seventeen surprisal-containing linear mixed-effects models after predictions were subtracted from target reading times.The post-hoc analysis used the exploratory set because the held-out set was reserved for significance testing.
- Residual-error analysis: Removing by-word random intercepts brought model likelihoods and mean squared errors closer before residuals were analyzed.The discrepancy was linked primarily to by-word intercepts in the Dundee models.
- Subset identification: Top subsets included named entities, nouns before relativizers, adjectives, modals, and several noun-phrase and clause-boundary categories across the two corpora.Natural Stories and Dundee shared some syntactic-category effects but also yielded corpus-specific structural subsets.
- Overprediction: Larger models overpredicted some function-word subsets because their smaller surprisal values led regression models to assign higher coefficients.This mechanism explains overprediction for words preceding sentential clauses, which mainly consisted of conjunctions; Natural Stories modals were another counterexample.
- Underprediction: Larger models increasingly underpredicted reading times for nouns and adjectives, including named entities and words completing long-distance dependencies.Syntactically complex words, such as high-DLT-cost items and ends of long center-embedded constituents, showed related underprediction patterns, especially in Natural Stories.
- Interpretation: The poorer fit is primarily associated with lower surprisal for open-class words, potentially reflecting domain knowledge from training examples unavailable to humans.Named entities illustrate the mismatch: humans incur increased processing times from retrieved semantic associations, whereas models can exploit high mutual information across multi-token names.
5 Discussion and Conclusion
Across multiple pre-trained Transformer families, smaller variants provide better fits to human reading times despite larger variants achieving better next-word prediction. Linguistic residual analyses suggest memorization and stronger local associations contribute to this divergence, while the lower size boundary remains unresolved.
- Better next-word prediction and lower perplexity in larger pre-trained variants nevertheless produce less predictive surprisal estimates for human reading times.This pattern was observed across multiple large pre-trained language models and contrasts with the assumption that larger models are always better cognitive predictors.
- A strictly monotonic, positive log-linear relationship between perplexity and reading-time fit was found across five GPT-Neo and eight OPT variants.The trend was evaluated on self-paced reading times and eye-gaze durations, providing support beyond earlier GPT-2-only results.
- Residual analyses found especially large model differences on nouns and adjectives, including named entities and predicative adjectives.Larger-model regressions increasingly underpredicted reading times for these word classes.
- Larger Transformer variants may learn stronger local token associations because self-attention can function like a lookup table and larger GPT-Neo models reproduce more training sequences verbatim.This memorization-based account may explain why nouns and adjectives show the strongest residual effects.
- The relationship between model performance and reading-time fit may differ across pre-trained and trained-from-scratch models, and the lower size boundary for pre-trained models remains unknown.Future work could examine how far model size can be reduced before surprisal predictive power begins to decline.
- The findings motivate caution when selecting pre-trained language models as surprisal-based predictors of human processing difficulty.The discussion specifically rejects choosing the largest available pre-trained model solely under a ‘larger is better’ assumption.