Intelligent Systems: 10th Brazilian Conference, BRACIS 2021, Virtual Event, November 29 – December 3, 2021, Proceedings, Part II (Lecture Notes in Computer Science)

The two-volume set LNAI 13073 and 13074 constitutes the proceedings of the 10th Brazilian Conference on Intelligent Systems, BRACIS 2021, held in São Paolo, Brazil, in November-December 2021. The total of 77 papers presented in these two volumes was carefully reviewed and selected from 192 submissions.The contributions are organized in the following topical sections: Part I: Agent and Multi-Agent Systems, Planning and Reinforcement Learning; Evolutionary Computation, Metaheuristics, Constrains and Search, Combinatorial and Numerical Optimization, Knowledge Representation, Logic and Fuzzy Systems; Machine Learning and Data Mining. Part II: Multidisciplinary Artificial and Computational Intelligence and Applications; Neural Networks, Deep Learning and Computer Vision; Text Mining and Natural Language Processing. Due to the COVID-2019 pandemic, BRACIS 2021 was held as a virtual event. Preface Organization Contents – Part II Contents – Part I Multidisciplinary Artificial and Computational Intelligence and Applications A Heterogeneous Network-Based Positive and Unlabeled Learning Approach to Detect Fake News 1 Introduction 2 Related Work 3 Positive and Unlabeled Learning Algorithms 4 Proposed Approach: PU-LP for Fake News Detection 4.1 News Collection and Representation Model 4.2 k-NN Matrix, Katz Index and Sets Extraction 4.3 Adding Features in the News Network 4.4 Label Propagation 5 Experimental Evaluation 5.1 News Datasets 5.2 Experimental Setup and Evaluation Criteria 5.3 Results and Discussions 6 Conclusion and Future Work References Anomaly Detection in Brazilian Federal Government Purchase Cards Through Unsupervised Learning Techniques 1 Introduction 2 Materials and Methods 2.1 K-Means Method 2.2 Agglomerative Clustering Method 2.3 Network-Based Approach 2.4 Hybrid Approach 3 Experimental Results 4 Final Remarks References De-Identification of Clinical Notes Using Contextualized Language Models and a Token Classifier 1 Introduction 2 Related Work 3 Materials and Methods 3.1 Data Source 3.2 Neural Network 3.3 Language Models 3.4 Design of the Experiments 4 Results 5 Conclusion References Detecting Early Signs of Insufficiency in COVID-19 Patients from CBC Tests Through a Supervised Learning Approach 1 Introduction 2 The Proposed MNBHL Technique 2.1 Description of the Training Phase 2.2 Description of the Testing Phase 3 Materials and Methods 4 Tests Performed on Benchmark Datasets 5 Experimental Results 6 Final Remarks References Encoding Physical Conditioning from Inertial Sensors for Multi-step Heart Rate Estimation 1 Introduction 2 Methodology 2.1 Datasets 2.2 Pre-processing 3 The Physical Conditional Embedding LSTM Model 3.1 Adaptations for PPG-Based HR Estimation 4 Reference Methods 5 Empirical Evaluation 5.1 Experimental Setup 5.2 IMU-Based Multi-step HR Estimation 5.3 Performance Impact of PCE-LSTM's Hidden State Initialization 5.4 PPG-Based HR Estimation 6 Related Work 7 Conclusions References Ensemble of Protein Stability upon Point Mutation Predictors 1 Introduction 2 Background 2.1 Point Mutations and Their Effects on Protein Structures 2.2 Gibbs Free Energy (G) 2.3 Supervised Machine Learning 2.4 Ensemble Learning 3 Individual Tools for Predicting the Effects of Point Mutations in Protein Stability 3.1 CUPSAT 3.2 SDM 3.3 mCSM 3.4 DUET 3.5 MAESTRO 3.6 PoPMuSic 4 Proposed Methodology 4.1 Input Data 4.2 Meta Data 4.3 Ensemble Learning - Stacking 4.4 Ensemble Learning - Bagging/Boosting 5 Results and Discussion 5.1 Experiment 1: Balanced Training Dataset 5.2 Experiment 2: Unbalanced Training Set 6 Conclusion References Ethics of AI: Do the Face Detection Models Act with Prejudice? 1 Introduction 2 Background 3 Methodology 3.1 Dataset 3.2 Face Detection Models 3.3 Validation Metrics 3.4 Hypothesis Test 4 Experimental Results 5 Conclusion References Evaluating Topic Models in Portuguese Political Comments About Bills from Brazil's Chamber of Deputies 1 Introduction 2 Related Work 3 Portuguese Political Comments 3.1 Corpora 4 Methodology 4.1 Sentence Embeddings 4.2 Topic Models 5 Experimental Evaluation 5.1 Quantity and Quality Evaluation 5.2 Setup 5.3 Results 6 Discussion References Evaluation of Convolutional Neural Networks for COVID-19 Classification on Chest X-Rays 1 Introduction 2 Related Work 3 Materials and Methods 3.1 Dataset 3.2 Preprocessing 3.3 Data Augmentation 3.4 Convolutional Architectures 4 Results and Discussion 4.1 Models Training and Validation 4.2 Comparison with Related Work 5 Conclusion References Experiments on Portuguese Clinical Question Answering 1 Introduction 2 Related Work 2.1 Corpora 2.2 Transfer Learning 2.3 Biomedical and Clinical QA 3 Materials and Methods 3.1 BioBERTpt-squad-v1.1-portuguese: A Biomedical QA Model for Portuguese 3.2 Evaluation Setup 4 Results 5 Discussion 6 Conclusion References Long-Term Map Maintenance in Complex Environments 1 Introduction 2 Related Work 3 Large-Scale Mapping System 3.1 Pre-processing 3.2 Hypergraph Building 3.3 Hypergraph Optimization 3.4 Large-Scale Environment Mapping 3.5 Large-Scale Map Merging 4 Experimental Methodology 4.1 Experiments 5 Results and Discussions 5.1 Odometry Bias Calibration 5.2 Map Building and Merging 6 Conclusions and Future Work References Supervised Training of a Simple Digital Assistant for a Free Crop Clinic 1 Introduction 2 Related Works 3 Proposed Approach 3.1 System Overview 3.2 Crop Clinic Digital Assistant 4 Hyperparameter Tuning and Training 4.1 Dataset Collection 4.2 Dataset Division 4.3 Model Training 4.4 Prediction 4.5 Performance Optimization 5 Case Study and Discussion 6 Conclusion References The Future of AI: Neat or Scruffy? 1 Introduction 2 The ``neats'' Vs. ``scruffies'' Debate in the History of AI 2.1 Marvin Minsky (1985–1995) 2.2 Nils Nilsson (2009) 2.3 Herbert Simon (1972) 2.4 John McCarthy (1958) 2.5 Russell and Norvig (1995–2020) 2.6 Yann LeCun (2018) 3 Is AI a Science of Intelligence or a Branch of Engineering? 4 Types of Neat and Scruffy's Attitudes in AI 4.1 Scruffy Type I: The Empirical Scientists 4.2 Scruffy Type II: The System Builders 4.3 Neats: The Computer Epistemologists 5 Implications for the Future of AI 6 Conclusion References Weapon Engagement Zone Maximum Launch Range Estimation Using a Deep Neural Network 1 Introduction 2 Background 2.1 Weapon Engagement Zone 2.2 Missile Model 2.3 Experimental Design 3 Methodology 3.1 Simulation 3.2 Preprocessing 3.3 Model Training 3.4 Model Evaluation 4 Results and Analysis 4.1 Exploratory Data Analysis 4.2 Model Predictions 4.3 Model Representation 5 Conclusions and Future Work References Neural Networks, Deep Learning and Computer Vision Code Autocomplete Using Transformers 1 Introduction 2 Related Work 3 Approach 3.1 Corpus 3.2 Model 4 Evaluation 4.1 DG Evaluation Metric 4.2 Methodology 4.3 Baseline 5 Results 6 Conclusion References Deep Convolutional Features for Fingerprint Indexing 1 Introduction 2 FVC Databases 3 Indexing, Evaluation, and Performance Metrics 4 Related Works 5 Proposed Indexing Method 5.1 Architecture 5.2 Deep Metric Learning 5.3 Approximate Nearest Neighbors 6 Experimental Evaluation 7 Results 8 Conclusion References How to Generate Synthetic Paintings to Improve Art Style Classification 1 Introduction 2 Related Work 2.1 Artwork Classification 2.2 Generative Adversarial Networks 3 Our Proposal 3.1 Image Augmentation 3.2 Generative Adversarial Network 3.3 Adversarial Loss Function 3.4 EfficientNet 4 Experimental Results 4.1 The Wikiart Dataset 4.2 GAN Training Configuration 4.3 EfficientNet B0 Training Configuration 4.4 Baseline Results 4.5 Sampling Low Quantity Classes 4.6 Sampling High Quantity Classes 4.7 Summary of Results 4.8 Generated Images 5 Conclusion References Iris-CV: Classifying Iris Flowers Is Not as Easy as You Thought 1 Introduction 2 Related Works 3 A New Iris Dataset 4 Benchmark Results 4.1 Deep Neural Net Results 5 Conclusion References Performance Analysis of YOLOv3 for Real-Time Detection of Pests in Soybeans 1 Introduction 2 Related Work 3 Materials and Methods 3.1 Image Acquisition 3.2 You only Look once (YOLOv3) 3.3 Experimental Design 3.4 Evaluation Metrics 4 Results and Discussion 5 Conclusion References Quaternion-Valued Convolutional Neural Network Applied for Acute Lymphoblastic Leukemia Diagnosis 1 Introduction 1.1 Complex and Quaternion-Valued Neural Networks 1.2 Contributions and the Organization of the Paper 2 Acute Lymphoblastic Leukemia (ALL) 2.1 Computer-Aided Diagnosis of Leukemia: Literature Review 3 Convolutional Neural Networks 3.1 Quaternion-Valued Convolutional Neural Networks 4 Computational Experiments 5 Concluding Remarks and Future Works References Sea State Estimation with Neural Networks Based on the Motion of a Moored FPSO Subjected to Campos Basin Metocean Conditions 1 Introduction 2 Background: the Description of Ocean Waves 3 Background: Neural Networks 4 An Estimation Procedure for Sea Parameters 4.1 Metocean Data 4.2 Data Processing 4.3 Network Architecture 5 Motion-Based Estimation Results and Discussion 6 Conclusions References Time-Dependent Item Embeddings for Collaborative Filtering 1 Introduction 2 Related Work 3 The Proposed Method 3.1 Item2Vec 3.2 Sequential Item2Vec 4 Experiments and Results 4.1 Datasets 4.2 Experimental Protocol 4.3 Results 5 Conclusion References Transfer Learning of Shapelets for Time Series Classification Using Convolutional Neural Network 1 Introduction 2 Related Work 3 Materials and Method 3.1 Shapelet Extraction 3.2 Neural Network Training 3.3 Training and Transfer Learning 4 Experimental Evaluation 5 Conclusion References Text Mining and Natural Language Processing A Deep Learning Approach for Aspect Sentiment Triplet Extraction in Portuguese 1 Introduction 2 Related Work 3 Proposed Framework 3.1 Problem Definition 3.2 Sentence Encoding 3.3 Aspect and Opinion Representation 3.4 Aspect and Opinion Tagging 3.5 Sentiment Dependency Parsing 3.6 Syntax-Aware Vector 4 Experiments 4.1 Datasets 4.2 Baselines 4.3 Experiment Settings 4.4 Evaluation 5 Results and Analysis 5.1 Comparison with Baselines 5.2 Ablation Study 6 Conclusions and Future Work References Aggressive Language Detection Using VGCN-BERT for Spanish Texts 1 Introduction 2 Related Work 2.1 Transfer Learning with BERT 2.2 Graph Convolutional Networks (GCN) 3 Methodology 3.1 VGCN 3.2 Integrating VGCN into BERT 4 Dataset 5 Experiments 5.1 Baselines and Task 5.2 Pre-processing and Model Setting 5.3 Evaluation Metrics 5.4 Experimental Results 6 Conclusion References An Empirical Study of Text Features for Identifying Subjective Sentences in Portuguese 1 Introduction 2 Related Work 3 Materials and Methods 3.1 Features 3.2 Models 3.3 Recursive Feature Elimination 4 Experiments 4.1 Setup 4.2 Metrics 4.3 Comparison of Classifier Performances 4.4 Feature Ablation Study 4.5 Feature Ranking 5 Conclusion References Comparing Contextual Embeddings forSemantic Textual Similarity in Portuguese 1 Introduction 2 Background 2.1 Transfer Learning 2.2 Contextual and Word Embeddings 2.3 BERT 2.4 Sentence-BERT (SBERT) 2.5 Knowledge Distillation 3 Semantic Textual Similarity in Portuguese 4 Experimental Assessment 4.1 Models Assessed 4.2 Experimental Setup 5 Results 5.1 Pre-trained Models Without Fine-Tuning for STS in Portuguese 5.2 Pre-trained Models with Fine-Tuning 6 Further Analysis 6.1 Fine-Tuning Effects 6.2 Language Variant Effects 7 Conclusions References Deep Active-Self Learning Applied to Named Entity Recognition 1 Introduction 2 Methodology 2.1 Deep Active-Self Learning Algorithm 2.2 Dynamic Update of Training Epochs - DUTE 2.3 Experiments 3 Results 3.1 Deep Active-Self Learning with DUTE Strategy 3.2 Ablation Study 4 Conclusion References DEEPAGÉ: Answering Questions in Portuguese About the Brazilian Environment 1 Introduction 2 Models and Architectures 2.1 BM25 as the Retriever 2.2 PTT5 as the Reader 3 Dataset Generation 3.1 Filtering Articles from Wikipedia in Portuguese 3.2 Scraping and Filtering News from the Biggest Brazilian Newspapers 3.3 Filtering and Translating the PAQ QA Dataset 4 Experiments 4.1 Experiment 1: Systems Without Fine-Tune 4.2 Experiment 2: Reader-Only, with Fine-Tune 4.3 Experiment 3: Retriever+Reader with Fine-Tune 5 Results and Discussion 5.1 Importance of Fine-Tune on the Specific Domain 5.2 Impact of Different KBs on Scores 5.3 Influence of Distractors on the Reader 6 Conclusion 7 Appendix A 8 Appendix B 9 Appendix C References Enriching Portuguese Word Embeddings with Visual Information 1 Introduction 2 Related Work 3 Resources and Methodology 3.1 Unimodal Embeddings 3.2 Dealing with the Information Gap 3.3 Fusion Techniques 3.4 Multimodal Embeddings 4 Tests and Results 4.1 Word Relatedness 4.2 Analogy Prediction 4.3 Semantic Similarity in Short Sentences 4.4 Named Entity Recognition 5 Conclusion and Future Work References Entity Relation Extraction from News Articles in Portuguese for Competitive Intelligence Based on BERT 1 Introduction 2 Background 2.1 Competitive Intelligence and News Processing 2.2 BERT 3 Related Work 4 Relationship Extractor Model 5 Experiments 5.1 Selection 5.2 Pre-processing 5.3 Transformation 5.4 Mining 5.5 Evaluation 6 Results and Discussion 7 Conclusion and Future Work References Experiments on Kaldi-Based Forced Phonetic Alignment for Brazilian Portuguese 1 Introduction 2 Methodology 2.1 Kaldi, Grapheme-to-Phoneme and Syllabification Tools 2.2 Training Speech Corpora and Lexicon 2.3 Acoustic Models 2.4 Kaldi Forced Phonetic Alignment 3 Evaluation Tests 3.1 Evaluation Speech Corpus 3.2 Simulation Overview 3.3 Many-to-Many (M2M) Phonetic Mapping 4 Results and Discussion 4.1 Discussion 5 Conclusion 5.1 Future Work References Incorporating Text Specificity into a Convolutional Neural Network for the Classification of Review Perceived Helpfulness 1 Introduction 2 Related Work 3 Specificity Prediction on Sentences from Reviews 4 Perceived Helpfulness Classification Model 5 Experimental Setup 5.1 Data and Preprocessing 5.2 Training 5.3 Model Variations 6 Results and Discussion 6.1 Perceived Helpfulness Classification 6.2 Sentence Specificity Prediction 7 Conclusion References Joint Event Extraction with Contextualized Word Embeddings for the Portuguese Language 1 Introduction 2 Related Work 3 Task Definition 4 Model 4.1 Trigger and Sentence Encoding 4.2 Sentence Representation 4.3 Trigger Representation 4.4 Event Detection and Argument Role Prediction 4.5 Training 5 Corpus 6 Experiments 6.1 Dataset, Parameters, and Resources 6.2 Data Augmentation via OIE 6.3 Word Embedding Layers Selection 6.4 Test Evaluation 7 Error Analysis 8 Conclusion References mRAT-SQL+GAP: A Portuguese Text-to-SQL Transformer 1 Introduction 2 Preliminary Tasks 2.1 Translating the Spider Dataset 3 Experiments 4 Analysis and Discussion 5 Conclusion and Future Work References Named Entity Recognition for Brazilian Portuguese Product Titles 1 Introduction 2 Background and Related Work 3 Attribute Set 4 Experiments and Results 4.1 Corpus Pre-processing and Annotation 4.2 Experimental Setup 4.3 Results 4.4 Comparison Between the Models 5 Final Remarks References Portuguese Neural Text Simplification Using Machine Translation 1 Introduction 2 Background 3 Related Works 4 Materials and Methods 4.1 Data Description 4.2 Automatic Text Simplification 4.3 Evaluation 5 Results 5.1 Simplification Quality 6 Conclusion References Rhetorical Role Identification for Portuguese Legal Documents 1 Introduction 2 Related Work 3 Corpus and Rhetorical Roles 4 Text Representation and Machine Learning Approaches 4.1 Baseline Methods 4.2 Deep Learning Methods 5 Experimental Evaluation 6 Conclusions and Future Work References Speech2Phone: A Novel and Efficient Method for Training Speaker Recognition Models 1 Introduction 2 Datasets and Pre-processing 2.1 Audio Datasets 2.2 Preprocessing of the Speech2Phone Dataset 2.3 Pre-processing of Speaker Verification Datasets 3 Speech2Phone 3.1 Proposed Method 3.2 Speech2Phone Results 4 Application: Speaker Verification 4.1 Speaker Verification Experiments 4.2 Speaker Verification Results 5 Conclusions and Future Work References Text Classification in Legal Documents Extracted from Lawsuits in Brazilian Courts 1 Introduction 2 Related Works 3 Corpus and Data Preparation 3.1 Corpus and Golden Collection 3.2 Integration with the Brazilian Legal Knowledge Graph 4 Lawsuit Classification 4.1 Experiment Scenarios 4.2 Models 4.3 Results and Discussion 5 Conclusions References Universal Dependencies-Based PoS Tagging Refinement Through Linguistic Resources 1 Introduction 2 Linguistic Analysis of Non-ambiguous Tokens in Portuguese 2.1 Non-ambiguous Single Tokens for Closed PoS Classes 2.2 Non-ambiguous Co-occurring Tokens 3 Evaluation 3.1 Evaluation of the Non-ambiguous Single Tokens from Closed PoS Classes Corrections 3.2 Evaluation of the Non-ambiguous Co-occurring Tokens Corrections 3.3 Evaluation of All Non-ambiguous Corrections 4 Concluding Remarks References When External Knowledge Does Not Aggregate in Named Entity Recognition 1 Introduction 2 Approach 2.1 Neural Models 2.2 Embeddings 2.3 External Knowledge 3 Experiments 3.1 Datasets 3.2 Experimental Settings 3.3 Experimental Results 4 Conclusion References Author Index