A Neural Model of Adaptation in Reading

Marten van Schijndel, Tal Linzen

Abstract

It has been argued that humans rapidly adapt their lexical and syntactic expectations to match the statistics of the current linguistic context. We provide further support to this claim by showing that the addition of a simple adaptation mechanism to a neural language model improves our predictions of human reading times compared to a non-adaptive model. We analyze the performance of the model on controlled materials from psycholinguistic experiments and show that it adapts not only to lexical items but also to abstract syntactic structures.

Introduction

When humans read, they do not process text with static expectations. Instead, they continuously update their probabilistic models of language based on the current context. For example, if a text frequently employs a rare grammatical construction, readers eventually stop being surprised by it, leading to faster processing times. This phenomenon, known as adaptation, is a hallmark of human linguistic flexibility.

Traditional neural language models (LMs) are typically "frozen" after an initial training phase. While these models capture general linguistic patterns, they fail to adapt to the specific stylistic or grammatical idiosyncrasies of a particular document. This paper investigates whether a simple, error-driven learning mechanism can enable neural LMs to mimic human-like adaptation, thereby improving their ability to predict human reading effort.

Methods

The study employs a gradient-based adaptation mechanism applied to a standard Long Short-Term Memory (LSTM) network pre-trained on 90 million words of Wikipedia. As the model processes a text sentence by sentence, it calculates the cross-entropy loss (prediction error). After each sentence, the model performs a single gradient-descent update to its weights before proceeding to the next.

This approach treats "surprisal"—the negative log probability of a word—as a computational proxy for human reading effort. By updating weights in response to high surprisal, the model "learns" to better expect similar structures in subsequent sentences.

Results

Predicting Human Reading Times

The primary evaluation compared "adaptive surprisal" against standard "non-adaptive surprisal" in predicting human reading times from the Natural Stories Corpus. The results demonstrate that the adaptive model is a significantly better predictor of human behavior.

Table 1: Regression coefficients for reading times. When adaptive surprisal is included, the predictive value of non-adaptive surprisal becomes non-significant (t = -1.31), indicating that the adaptive model captures the relevant variance more effectively.

On the Natural Stories Corpus, this adaptation mechanism reduced the model's test perplexity from 141.49 to 86.99, a substantial improvement in predictive accuracy.

Syntactic Adaptation and the Garden Path Effect

To test whether the model adapts to abstract structures rather than just specific words, the researchers used "garden path" sentences (reduced relative clauses). In these experiments, the model's behavior closely mirrored human data.

Figure 1: The model's adaptation curve for ambiguous (garden path) vs. unambiguous sentences. As the model encounters more examples, the surprisal for the ambiguous condition decreases, reflecting rapid structural adaptation.

The model also showed success in "dative alternation" tasks. Even when test sentences shared no vocabulary with the adaptation sentences, the model achieved an 84% reduction in perplexity, confirming that it learns abstract syntactic patterns.

Discussion and Conclusion

The findings suggest that a simple error-driven update is sufficient to capture complex human adaptation behaviors. Notably, the model remained resilient to "catastrophic forgetting"; adapting to specific genres improved local performance without degrading the model's general linguistic knowledge. However, the study also revealed that syntactic adaptation is more "fragile" than lexical adaptation, requiring careful tuning of learning rates to avoid overfitting.

By incorporating a continuous learning mechanism, neural language models can more accurately reflect the dynamic nature of human language processing, bridging the gap between static computational models and the adaptive reality of human cognition.

Summary updated

Abstract

It has been argued that humans rapidly adapt their lexical and syntactic expectations to match the statistics of the current linguistic context. We provide further support to this claim by showing that the addition of a simple adaptation mechanism to a neural language model improves our predictions of human reading times compared to a non-adaptive model. We analyze the performance of the model on controlled materials from psycholinguistic experiments and show that it adapts not only to lexical items but also to abstract syntactic structures.
BibTeX
@inproceedings{schijndel2018neural,
  author = {Marten van Schijndel and Tal Linzen},
  title = {{A Neural Model of Adaptation in Reading}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
  url = {https://openreview.net/forum?id=3dUS-duwoYg},
}
Figure