As Transformers become more widely incorporated into natural language processing tasks, there has been considerable interest in using surprisal from these models as predictors of human sentence processing difficulty. Recent work has observed a positive relationship between Transformer-based models' perplexity and the predictive power of their surprisal estimates on reading times, showing that language models with more parameters and trained on more data are less predictive of human reading times. However, these studies focus on predicting latency-based measures (i.e., self-paced reading times and eye-gaze durations) with surprisal estimates from Transformer-based language models. This trend has not been tested on brain imaging data. This study therefore evaluates the predictive power of surprisal estimates from 17 pre-trained Transformer-based models across three different language families on two functional magnetic resonance imaging datasets. Results show that the positive relationship between model perplexity and model fit still obtains, suggesting that this trend is not specific to latency-based measures and can be generalized to neural measures.

The effect of surprisal on processing difficulty has been a central topic of investigation in psycholinguistics. Here, we use eyetracking data to examine three language processing regimes that are common in daily life but have not been addressed with respect to this question: information seeking, repeated processing, and the combination of the two. Using standard regime-agnostic surprisal estimates we find that the prediction of surprisal theory regarding the presence of a linear effect of surprisal on processing times, extends to these regimes. However, when using surprisal estimates from regime-specific contexts that match the contexts and tasks given to humans, we find that in information seeking, such estimates do not improve the predictive power of processing times compared to standard surprisals. Further, regime-specific contexts yield near zero surprisal estimates with no predictive power for processing times in repeated reading. These findings point to misalignments of task and memory representations between humans and current language models, and question the extent to which such models can be used for estimating cognitively relevant quantities. We further discuss theoretical challenges posed by these results.

Recent psycholinguistic research has compared human reading times to surprisal estimates from language models to study the factors shaping human sentence processing difficulty. Previous studies have shown a strong fit between surprisal values from Transformers and reading times. However, standard Transformers work with a lossless representation of the entire previous linguistic context, unlike models of human language processing that include memory decay. To bridge this gap, this paper evaluates a modification of the Transformer model that uses ALiBi (Press et al., 2022), a recency bias added to attention scores. Surprisal estimates with ALiBi show an improved fit to human reading times compared to a standard Transformer baseline. A subsequent analysis of attention heads suggests that ALiBi's mixture of slopes -- which determine the rate of memory decay in each attention head -- may play a role in the improvement by helping models with ALiBi to track different kinds of linguistic dependencies.

Recent studies have shown that as Transformer-based language models become larger and are trained on very large amounts of data, the fit of their surprisal estimates to naturalistic human reading times degrades. The current work presents a series of analyses showing that word frequency is a key explanatory factor underlying these two trends. First, residual errors from four language model families on four corpora show that the inverse correlation between model size and fit to reading times is the strongest on the subset of least frequent words, which is driven by excessively accurate predictions of larger model variants. Additionally, training dynamics reveal that during later training steps, all model variants learn to predict rare words and that larger model variants do so more accurately, which explains the detrimental effect of both training data amount and model size on fit to reading times. Finally, a feature attribution analysis demonstrates that larger model variants are able to accurately predict rare words based on both an effectively longer context window size as well as stronger local associations compared to smaller model variants. Taken together, these results indicate that Transformer-based language models' surprisal estimates diverge from human-like expectations due to the superhumanly complex associations they learn for predicting rare words.

Numerous analyses of reading time (RT) data have been implemented -- all in an effort to better understand the cognitive processes driving reading comprehension. However, data measured on words at the end of a sentence -- or even at the end of a clause -- is often omitted due to the confounding factors introduced by so-called "wrap-up effects," which manifests as a skewed distribution of RTs for these words. Consequently, the understanding of the cognitive processes that might be involved in these wrap-up effects is limited. In this work, we attempt to learn more about these processes by examining the relationship between wrap-up effects and information-theoretic quantities, such as word and context surprisals. We find that the distribution of information in prior contexts is often predictive of sentence- and clause-final RTs (while not of sentence-medial RTs). This lends support to several prior hypotheses about the processes involved in wrap-up effects.

Zipf (1935) posited that wordforms are optimized to minimize utterances' communicative costs. Under the assumption that cost is given by an utterance's length, he supported this claim by showing that words' lengths are inversely correlated with their frequencies. Communicative cost, however, can be operationalized in different ways. Piantadosi et al. (2011) claim that cost should be measured as the distance between an utterance's information rate and channel capacity, which we dub the channel capacity hypothesis (CCH) here. Following this logic, they then proposed that a word's length should be proportional to the expected value of its surprisal (negative log-probability in context). In this work, we show that Piantadosi et al.'s derivation does not minimize CCH's cost, but rather a lower bound, which we term CCH-lower. We propose a novel derivation, suggesting an improved way to minimize CCH's cost. Under this method, we find that a language's word lengths should instead be proportional to the surprisal's expectation plus its variance-to-mean ratio. Experimentally, we compare these three communicative cost functions: Zipf's, CCH-lower , and CCH. Across 13 languages and several experimental settings, we find that length is better predicted by frequency than either of the other hypotheses. In fact, when surprisal's expectation, or expectation plus variance-to-mean ratio, is estimated using better language models, it leads to worse word length predictions. We take these results as evidence that Zipf's longstanding hypothesis holds.

The use of neural language models to model human behavior has met with mixed success. While some work has found that the surprisal estimates from these models can be used to predict a wide range of human neural and behavioral responses, other work studying more complex syntactic phenomena has found that these surprisal estimates generate incorrect behavioral predictions. This paper explores the extent to which the misalignment between empirical and model-predicted behavior can be minimized by training models on more developmentally plausible data, such as in the BabyLM Challenge. We trained teacher language models on the BabyLM "strict-small" dataset and used sentence level surprisal estimates from these teacher models to create a curriculum. We found tentative evidence that our curriculum made it easier for models to acquire linguistic knowledge from the training data: on the subset of tasks in the BabyLM challenge suite evaluating models' grammatical knowledge of English, models first trained on the BabyLM data curriculum and then on a few randomly ordered training epochs performed slightly better than models trained on randomly ordered epochs alone. This improved linguistic knowledge acquisition did not result in better alignment with human reading behavior, however: models trained on the BabyLM dataset (with or without a curriculum) generated predictions that were as misaligned with human behavior as models trained on larger less curated datasets. This suggests that training on developmentally plausible datasets alone is likely insufficient to generate language models capable of accurately predicting human language processing.

Recent psycholinguistic studies have drawn conflicting conclusions about the relationship between the quality of a language model and the ability of its surprisal estimates to predict human reading times, which has been speculated to be due to the large gap in both the amount of training data and model capacity across studies. The current work aims to consolidate these findings by evaluating surprisal estimates from Transformer-based language model variants that vary systematically in the amount of training data and model capacity on their ability to predict human reading times. The results show that surprisal estimates from most variants with contemporary model capacities provide the best fit after seeing about two billion training tokens, after which they begin to diverge from humanlike expectations. Additionally, newly-trained smaller model variants reveal a 'tipping point' at convergence, after which the decrease in language model perplexity begins to result in poorer fits to human reading times. These results suggest that the massive amount of training data is mainly responsible for the poorer fit achieved by surprisal from larger pre-trained language models, and that a certain degree of model capacity is necessary for Transformer-based language models to capture humanlike expectations.

We report a controlled study investigating the effect of visual information (i.e., seeing the speaker) on spoken language comprehension. We compare the ERP signature (N400) associated with each word in audio-only and audio-visual presentations of the same verbal stimuli. We assess the extent to which surprisal measures (which quantify the predictability of words in their lexical context) are generated on the basis of different types of language models (specifically n-gram and Transformer models) that predict N400 responses for each word. Our results indicate that cognitive effort differs significantly between multimodal and unimodal settings. In addition, our findings suggest that while Transformer-based models, which have access to a larger lexical context, provide a better fit in the audio-only setting, 2-gram language models are more effective in the multimodal setting. This highlights the significant impact of local lexical context on cognitive processing in a multimodal environment.

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.