Early Detection of Mental Health Disorders by Social Media Monitoring: The First Five Years of the eRisk Project (Studies in Computational Intelligence, 1018)

eRisk stands for Early Risk Prediction on the Internet. It is concerned with the exploration of techniques for the early detection of mental health disorders which manifest in the way people write and communicate on the internet, in particular in user generated content (e.g. Facebook, Twitter, or other social media). Early detection technologies can be employed in several different areas but particularly in those related to health and safety. For instance, early alerts could be sent when the writing of a teenager starts showing increasing signs of depression, or when a social media user starts showing suicidal inclinations, or again when a potential offender starts publishing antisocial threats on a blog, forum or social network. eRisk has been the pioneer of a new interdisciplinary area of research that is potentially applicable to a wide variety of situations, problems and personal profiles. This book presents the best results of the first five years of the eRisk project which started in 2017 and developed into one of the most successful track of CLEF, the Conference and Lab of the Evaluation Forum. Foreword 6 Preface 10 Contents 11 Early Risk Prediction of Mental Health Disorders 13 1 Introduction 13 2 Book Content 15 The eRisk Initiative 19 The Challenge of Early Risk Prediction on the Internet 20 1 Introduction 20 2 Text as a Source of Risk Evidence 22 3 Datasets 23 3.1 Positive Group 24 3.2 Control Group 25 3.3 Extraction of the User's Submissions 25 3.4 eRisk Collections 26 4 Evaluation Design 29 4.1 Decision-Based Metrics 30 4.2 Ranking-based Evaluation 34 5 eRisk Task on Automatically Filling a Depression Questionnaire 34 5.1 Evaluation Metrics 37 6 Conclusions 38 References 39 A Survey of the First Five Years of eRisk: Findings and Conclusions 41 1 Introduction 42 2 Early Risk Detection Systems Design 43 2.1 Feature Extraction 43 2.2 Assessment Technologies 47 2.3 Decision Policy 50 3 Measuring the Severity of the Signs of Depression 51 4 Findings and Lessons 53 4.1 A Challenging Trade-Off 54 4.2 Experimental Design and Evaluation Framework 57 4.3 Signals Crossover 58 5 Conclusions 61 References 62 The Best of eRisk 68 From Bag-of-Words to Transformers: A Deep Dive into the Participation in the eRisk Early Risk Detection of Depression Tasks with Classical and New Approaches 69 1 Introduction 70 2 eRisk 2017/2018 Task for Early Detection of Depression 71 2.1 Overview 71 2.2 Datasets 72 3 Evaluation Criteria and Concerns About ERDE Score 73 4 Features and Architectures Used for Depression Detection 76 4.1 Hand-Crafted User Features 77 4.2 Readability 81 4.3 Word and Grammar Usage 81 4.4 Metadata Feature Summary 82 4.5 Linguistic Metadata 83 4.6 Emotions and Sentiment Lexica 86 4.7 Word Embeddings 86 4.8 Neural Network Architecture 90 4.9 Transformers 91 5 Chosen Models for the eRisk Tasks, Related Work and Ongoing Work 91 5.1 Bag-of-Words Ensemble - BCSGA - 2017/2018 92 5.2 Paragraph Vector - BCSGB - 2017 94 5.3 Bag-of-Words Ensemble - BCSGB - 2018 96 5.4 LSTM with LSA Vectors - BCSGC - 2017 96 5.5 CNN with GloVe Embeddings - BCSGC - 2018 97 5.6 LSTM with Paragraph Vectors - BCSGD - 2017 98 5.7 CNN with fastText Embeddings - BCSGD - 2018 98 5.8 Late LSTM with Paragraph Vectors - BCSGE - 2017 98 5.9 CNN and Bag-of-Words Metadata Ensemble - BCSGE - 2018 99 5.10 CNN's and metadata features 100 5.11 Setup for Transformer Experiments 100 6 Results 102 6.1 Results eRisk 2017 102 6.2 Results eRisk 2018 104 6.3 Related results 106 6.4 Results of Ongoing Work 108 6.5 Results with Transformers 110 7 Discussion 111 8 Conclusion 112 References 113 Comparison of Machine Learning Models for Early Depression Detection from Users' Posts 118 1 Introduction 119 2 Related Work 121 3 Information Modeling 124 3.1 Feature-Based Representation 125 4 Machine Learning Models 131 4.1 Well-Established Machine Learning Models 131 4.2 BERT-Based Model 132 5 Experimental Framework 133 5.1 Collections 133 6 Results and Discussion 134 6.1 Depression Detection 134 6.2 Early Depression Detection 137 6.3 Simplified Models 137 6.4 Ablation Analysis 139 7 Visualization of Early Detection 142 8 Conclusion 142 References 143 Quick and (Maybe Not So) Easy Detection of Anorexia in Social Media: To Explainability and Beyond 147 1 Introduction 147 2 Related Work 148 3 System Overview 150 3.1 Sub-models 150 3.2 Ensemble Model 153 4 Experimental Setup 153 4.1 Sub-models Implementation 153 4.2 Ensemble Classifiers 155 4.3 Submitted Runs 155 5 Shared Task Results 155 6 Explainability Analysis 157 6.1 Experiments 157 6.2 Results and Discussion 157 6.3 Further Analysis 160 7 Conclusion 162 References 163 Two Simple and Domain-independent Approaches for Early Detection of Anorexia 165 1 Introduction 166 2 Approaches 167 2.1 Flexible Temporal Variation of Terms (FTVT) 167 2.2 SS3 Text Classifier 169 3 Participation and Results 173 3.1 Early Detection of Anorexia—2018 Edition 173 3.2 Early Detection of Anorexia—2019 Edition 177 4 Conclusion and Future Work 185 References 187 Early Risk Detection of Self-Harm Using BERT-Based Transformers 189 1 Introduction 190 2 Related Work 191 3 Approach 191 4 Experiments 193 4.1 Provided Training Data 193 4.2 Created Training Data 194 4.3 Method 197 4.4 Results 199 5 Additional Experiments 200 5.1 Provided Training Data (Anorexia and Depression) 200 5.2 Created Training Data (Anorexia and Depression) 203 5.3 Method (Anorexia and Depression) 203 5.4 Results (Anorexia and Depression) 203 6 Summary 206 References 211 Detecting Traces of Self-harm on Reddit Through Emotional Patterns 213 1 Introduction 214 2 Related Work 215 3 From Sentiments to Emotions 216 4 Can Emotions Have Shades? A Sub-emotion Based Representation 217 4.1 Generating Sub-emotions 217 4.2 Analysis of the Novel Sub-emotions 219 4.3 Converting Text to Sub-emotions Sequences 221 5 Using BoSE to Identify Self-harm 223 5.1 BoSE Definition 223 5.2 BoSE at eRisk 2020 224 5.3 BoSE for Whole Post's History 228 6 Learning Sequential Information from Sub-emotions 229 6.1 normal upper DeltaΔ-BoSE 229 7 Deep Learning for Extracting Sequential Emotion Patterns 231 7.1 Convolutional Neural Network and BoSE 232 7.2 Recurrent Neural Network and BoSE 233 7.3 Adding Attention to the Sub-emotions 234 8 Conclusion 238 References 238 On the Estimation of Depression Through Social Mining 241 1 Introduction 241 2 Background 242 3 Measuring Depression 244 3.1 Dataset Description 245 3.2 Metrics 245 3.3 Methods 246 3.4 Results 249 4 Discussion 249 5 Conclusions 250 References 251 Automatically Estimating the Severity of Multiple Symptoms Associated with Depression 253 1 Introduction 254 2 Task and Data 254 3 Evaluation Metrics 256 4 Related Work 257 5 Approaching the Task as One of Authorship Attribution 258 5.1 Topic Models 260 5.2 Contextualizer 261 5.3 Stylometry 262 6 Results and Discussion 263 7 Conclusion 264 References 265 Beyond eRisk 268 Beyond Risk: Individual Mental Health Trajectories from Large-Scale Social Media Data 269 1 Introduction 269 2 Studying Behavioral Traces of Sleep and Emotions 274 2.1 Sampling Approach 274 2.2 Behavioral Analysis 276 2.3 Emotional Analysis 278 3 Detecting Cognitive Markers of Mental Health Disorders 279 3.1 Constructing Cognitive Markers 280 3.2 Sampling Approach 281 3.3 Lexical Analysis of Cognitive Distortion Markers 282 4 Conclusion, Discussion, and Future Research 284 References 286 Explainability of Depression Detection on Social Media: From Deep Learning Models to Psychological Interpretations and Multimodality 292 1 Introduction 293 2 Previous Work 295 3 Datasets 297 4 Models 299 4.1 Features 299 4.2 Hierarchical Attention Networks 300 4.3 Baseline Classifiers 302 4.4 Classification Results 303 5 Explainability of Depression Detection 303 5.1 Error Analysis 304 5.2 Hidden Layer Analysis 308 5.3 Ablation Experiments 308 6 Multimodal Depression Detection 310 6.1 multiRedditDep Dataset Collection 310 6.2 Classification Experiments 311 6.3 Image Analyses and Psychological Interpretations 313 7 Conclusions and Future Directions 318 References 319 The Future 324 The Future of eRisk 325 1 eRisk so Far 325 2 Future Outlines for Early Risk Prediction on the Internet 326 2.1 Topic Coverage, Domain Expansion and New Metrics for Early Risk 327 2.2 A Continuous Improvement of the eRisk Cycle 330 2.3 New Challenges: Estimating Risk from Standard Questionnaires 331 3 Conclusions 334 References 335