Cognitive Plausibility in Natural Language Processing

This book explores the cognitive plausibility of computational language models and why it’s an important factor in their development and evaluation. The authors present the idea that more can be learned about cognitive plausibility of computational language models by linking signals of cognitive processing load in humans to interpretability methods that allow for exploration of the hidden mechanisms of neural models. The book identifies limitations when applying the existing methodology for representational analyses to contextualized settings and critiques the current emphasis on form over more grounded approaches to modeling language. The authors discuss how novel techniques for transfer and curriculum learning could lead to cognitively more plausible generalization capabilities in models. The book also highlights the importance of instance-level evaluation and includes thorough discussion of the ethical considerations that may arise throughout the various stages of cognitive plausibility research. Acknowledgments 6 Contents 8 About the Authors 11 1 Introduction 12 [DELETE] 12 1.1 Does Cognitive Plausibility Matter? 14 1.1.1 Human-Centered Natural Language Processing 14 1.1.2 Understanding Language Versus Building Tools 15 1.1.3 Ethical Considerations 16 1.2 Dimensions of Cognitive Plausibility 17 1.2.1 Behavioral Patterns 17 1.2.2 Representational Structure 18 1.2.3 Procedural Strategies 18 1.3 Analyzing Cognitive Plausibility 19 2 Foundations of Language Modeling 22 [DELETE] 22 2.1 Methodological Concepts 22 2.1.1 Language Modeling with Recurrent Neural Networks 23 2.1.2 Evaluating Language Models 24 2.1.3 Language Modeling as Representation Learning 25 2.2 Modeling Decisions 26 2.2.1 Target Objective 27 2.2.2 Input Units 29 2.2.3 Processing Order 30 2.3 Ethical Aspects 32 3 Cognitive Signals of Language Processing 41 [DELETE] 41 3.1 Cognitive Signal Types 42 3.1.1 Offline Measures 42 3.1.2 Online Measures 45 3.1.3 Brain Activity Data 47 3.1.4 Combining Signal Types 50 3.2 Preprocessing Cognitive Signals for NLP 52 3.2.1 Participant Aggregation 52 3.2.2 Stimulus Alignment 53 3.2.3 Dimensionality Reduction 54 3.3 Available Datasets 55 3.3.1 Annotation Rationales Benchmark 56 3.3.2 Self-Paced Reading of Short Stories 56 3.3.3 A Multilingual Eye-Tracking Corpus 57 3.3.4 EEG Datasets of Reading and Listening 57 3.3.5 A Multilingual FMRI Dataset 58 3.4 Ethical Aspects 58 4 Behavioral Patterns 71 4.1 Analyzing Behavioral Patterns 72 4.1.1 Data and Error Analysis 72 4.1.2 Considering Difficulty 75 4.2 Testing Behavior 78 4.2.1 Testing Linguistic Phenomena 78 4.2.2 Robustness and Generalizability 81 4.3 Towards Cognitively Plausible Behavior 83 4.3.1 Finegrained Evaluation 84 4.3.2 Curriculum Learning 84 4.3.3 Multilingual Perspective 86 4.4 Ethical Aspects 87 5 Representational Structure 99 5.1 Analyzing Representational Structure 99 5.1.1 Representational Similarity 100 5.1.2 Comparing Representational Spaces 103 5.2 Testing Representational Characteristics 106 5.2.1 Probing Linguistic Knowledge 106 5.2.2 Probing Brain Activation Patterns 109 5.3 Towards Cognitively Plausible Representations 111 5.3.1 Multimodal Grounding 111 5.3.2 Cognitive Grounding 113 5.4 Ethical Aspects 115 6 Procedural Strategies 130 6.1 Analyzing Computational Processing Signals 131 6.1.1 Attention Values 132 6.1.2 Gradient-Based Saliency 133 6.2 Testing Processing Strategies 135 6.2.1 Relative Importance 135 6.2.2 Local Processing Effects 139 6.3 Towards Cognitively Plausible Processing 143 6.3.1 Multitask Learning 144 6.3.2 Transfer Learning 145 6.3.3 Integrating Linguistic Information 146 6.4 Ethical Aspects 147 7 Towards Cognitively More Plausible Models 161