📄 Cognitive Plausibility in Natural Language Processing

Summary

This item is a book review, not an original study. Yevgen Matusevych (University of Groningen) evaluates Cognitive Plausibility in Natural Language Processing by Lisa Beinborn and Nora Hollenstein (Springer, Synthesis Lectures on Human Language Technologies, 2024). The book asks how closely modern NLP systems resemble human language processing. This criterion is usually neglected in favour of accuracy, fairness or efficiency. The authors treat plausibility as graded and multidimensional and study three aspects: whether models and people make matching decisions, use comparable internal representations, and follow similar processing strategies. They draw on behavioral and neural datasets of language comprehension, though not production. They also survey interpretability tools such as test suites, probing classifiers, attention analysis and mapping to brain responses. They find no universal recipe but propose concrete directions: difficulty-aware evaluation, context-sensitive representation analysis, multimodal input and genuinely multilingual design. The reviewer praises the gradual, accessible structure and the ethics section closing each chapter. He highlights the lesson that human data is noisy and graded, and suggests NLP should learn to accommodate this.

Key points

• The book has seven chapters. Chapters 2–3 introduce language modeling basics and cognitive data sources (behavioral and brain signals). Chapters 4–6 cover behavioral patterns, representational structure and procedural strategies.
• Curriculum learning is presented as promising, but in the 2023 BabyLM challenge it brought only modest gains (Warstadt et al. 2023).
• Multi-head attention in transformers is judged cognitively implausible. However, attention weights can still be compared with human data such as reading gaze patterns.
• The authors call for evaluation that is finer-grained and less English-centric. They also propose integrating human problem-solving strategies through multi-task and transfer learning.
• The review appeared in Computational Linguistics 50(3), 2024 (doi:10.1162/coli_r_00517) under a CC BY-NC-ND 4.0 license. The book's DOI is 10.1007/978-3-031-43260-6.