Cognitive Plausibility in Natural Language Processing

Acknowledgments
Contents
About the Authors
1 Introduction
	[DELETE]
	1.1 Does Cognitive Plausibility Matter?
		1.1.1 Human-Centered Natural Language Processing
		1.1.2 Understanding Language Versus Building Tools
		1.1.3 Ethical Considerations
	1.2 Dimensions of Cognitive Plausibility
		1.2.1 Behavioral Patterns
		1.2.2 Representational Structure
		1.2.3 Procedural Strategies
	1.3 Analyzing Cognitive Plausibility
2 Foundations of Language Modeling
	[DELETE]
		2.1 Methodological Concepts
			2.1.1 Language Modeling with Recurrent Neural Networks
			2.1.2 Evaluating Language Models
			2.1.3 Language Modeling as Representation Learning
		2.2 Modeling Decisions
			2.2.1 Target Objective
			2.2.2 Input Units
			2.2.3 Processing Order
		2.3 Ethical Aspects
3 Cognitive Signals of Language Processing
	[DELETE]
	3.1 Cognitive Signal Types
		3.1.1 Offline Measures
		3.1.2 Online Measures
		3.1.3 Brain Activity Data
		3.1.4 Combining Signal Types
	3.2 Preprocessing Cognitive Signals for NLP
		3.2.1 Participant Aggregation
		3.2.2 Stimulus Alignment
		3.2.3 Dimensionality Reduction
	3.3 Available Datasets
		3.3.1 Annotation Rationales Benchmark
		3.3.2 Self-Paced Reading of Short Stories
		3.3.3 A Multilingual Eye-Tracking Corpus
		3.3.4 EEG Datasets of Reading and Listening
		3.3.5 A Multilingual FMRI Dataset
	3.4 Ethical Aspects
4 Behavioral Patterns
	4.1 Analyzing Behavioral Patterns
		4.1.1 Data and Error Analysis
		4.1.2 Considering Difficulty
	4.2 Testing Behavior
		4.2.1 Testing Linguistic Phenomena
		4.2.2 Robustness and Generalizability
	4.3 Towards Cognitively Plausible Behavior
		4.3.1 Finegrained Evaluation
		4.3.2 Curriculum Learning
		4.3.3 Multilingual Perspective
	4.4 Ethical Aspects
5 Representational Structure
	5.1 Analyzing Representational Structure
		5.1.1 Representational Similarity
		5.1.2 Comparing Representational Spaces
	5.2 Testing Representational Characteristics
		5.2.1 Probing Linguistic Knowledge
		5.2.2 Probing Brain Activation Patterns
	5.3 Towards Cognitively Plausible Representations
		5.3.1 Multimodal Grounding
		5.3.2 Cognitive Grounding
	5.4 Ethical Aspects
6 Procedural Strategies
	6.1 Analyzing Computational Processing Signals
		6.1.1 Attention Values
		6.1.2 Gradient-Based Saliency
	6.2 Testing Processing Strategies
		6.2.1 Relative Importance
		6.2.2 Local Processing Effects
	6.3 Towards Cognitively Plausible Processing
		6.3.1 Multitask Learning
		6.3.2 Transfer Learning
		6.3.3 Integrating Linguistic Information
	6.4 Ethical Aspects
7 Towards Cognitively More Plausible Models