Text, Speech, and Dialogue: 25th International Conference, TSD 2022, Brno, Czech Republic, September 6–9, 2022, Proceedings (Lecture Notes in Artificial Intelligence)

This book constitutes the proceedings of the 25th International Conference on Text, Speech, and Dialogue, TSD 2022, held in Brno, Czech Republic, in September 2022. The 43 papers presented in this volume were carefully reviewed and selected from 94 submissions. The topical sections "Text", "Speech", and "Dialogue" deal with the following issues: speech recognition; corpora and language resources; speech and spoken language generation; tagging, classification and parsing of text and speech; semantic processing of text and speech; integrating applications of text and speech processing; automatic dialogue systems; multimodal techniques and modelling. Preface Organization Contents Text Evaluating Attribution Methods for Explainable NLP with Transformers 1 Introduction 2 Related Work 3 Evaluation Datasets 3.1 Sentiment Analysis 3.2 Multi-label Document Classification 4 Experiments 4.1 Attribution Methods 4.2 Models 4.3 Metrics 5 Results and Discussion 5.1 Gradients and Gradients I nputs 5.2 SmoothGrad 5.3 Integrated Gradients 5.4 Chefer et al. 5.5 Datasets 6 Conclusion References DaFNeGE: Dataset of French Newsletters with Graph Representation and Embedding 1 Introduction 2 Context and Data Origin 2.1 Existing Resources 2.2 Presentation of Our Dataset 3 Features Description 3.1 Features Extraction 3.2 Textual Features 3.3 Graphical Features 4 Graphical Connections 4.1 Linkage Rule 4.2 Graph Representation of a Newsletter 5 Model 5.1 Relational Graph Convolutional Network 5.2 Results 6 Conclusion References ANTILLES: An Open French Linguistically Enriched Part-of-Speech Corpus 1 Introduction 2 Existing Resources for French POS Tagging 2.1 POS Corpora 2.2 POS Taggers 3 Extended Corpus Proposal 4 Experiments 4.1 Proposed Approaches 4.2 Results 5 Conclusion and Perspectives References Statistical and Neural Methods for Cross-lingual Entity Label Mapping in Knowledge Graphs 1 Introduction 2 Related Work 3 Data Preparation 4 Entity Label Mapping 4.1 Label Cross-Lingual Similarity Scoring Methods 4.2 Best Match Algorithm 4.3 Ground Truth and Method Comparison 5 Evaluation and Discussion 6 Conclusion References Quality Assessment of Subtitles – Challenges and Strategies 1 Introduction 1.1 Quality Assessment of Machine Translated Subtitles 2 Data and Methodology 2.1 The Corpus 2.2 Error Annotation and Evaluation 3 Estimating Subtitling Quality and Post-Editing Effort 4 Analysis Results 4.1 Inter-annotator Agreement 4.2 Subtitling Quality 4.3 Post-editing Effort 4.4 Error Analysis 5 Conclusion References Computational Approaches for Understanding Semantic Constraints on Two-termed Coordination Structures 1 Introduction 2 Background and Related Work 3 Approach 3.1 Universal Dependencies Corpora 3.2 Coordination Extraction 3.3 Semantic Analysis 4 Results 4.1 WordNet Analysis 4.2 Word Embedding Analysis 5 Discussion 5.1 WordNet Analysis 5.2 Word Embedding Analysis 6 Conclusion References Review of Practices of Collecting and Annotating Texts in the Learner Corpus REALEC 1 Introduction 2 Learner Corpora Available for Research Purposes 3 Data Collection and Annotation Practices in REALEC 4 Corpus Statistics 5 Error Taxonomy in REALEC 6 Distribution of Learner Errors in REALEC 7 Corpus at Work 8 Conclusions and Future Research References New Language Identification and Sentiment Analysis Modules for Social Media Communication 1 Introduction 2 Language Identification 2.1 Related Works 2.2 On the Difficulty of Identifying Internet Language 2.3 Custom FastText Model 2.4 Dictionary-Based Language Identification 2.5 The Ensemble Method 2.6 Evaluation 3 Sentiment Analysis 3.1 Related Works 3.2 Sentiment Analysis for Czech 3.3 Evaluation 4 Conclusion References Identification of Metaphorical Collocations in Different Languages – Similarities and Differences 1 Introduction 2 Related Work 3 Methodology 3.1 Corpora 3.2 Grammatical Relations and Annotation 3.3 Experiment 4 Results 5 Conclusion References Automatic Grammar Correction of Commas in Czech Written Texts: Comparative Study 1 Introduction 2 Writing Commas in Czech Language 2.1 Typology of the Comma Insertion Place 3 Automatic Grammar Correction of Commas 3.1 Rule-based Approach 3.2 Transformer-Based Approach 4 Evaluation Data Sets 5 Experimental Results 6 Conclusion References Linear Transformations for Cross-lingual Sentiment Analysis 1 Introduction 2 Related Work 3 Experimental Setup 3.1 Data 3.2 Linear Transformations 3.3 Neural Network Models 3.4 Cross-lingual Sentiment Classification 4 Experiments and Results 4.1 Monolingual Results 4.2 Cross-lingual Results 4.3 Comparison with Existing Works 4.4 Discussion 5 Conclusion References TOKEN Is a MASK: Few-shot Named Entity Recognition with Pre-trained Language Models 1 Introduction 2 Related Work 3 Method 3.1 Zero-Shot Base Method 3.2 Few-Shot Hybrid Method 4 Data 5 Experiments 5.1 Comparing Language Models 5.2 Choice of Template 5.3 Domain Adaptation 6 Conclusion and Future Work References Use of Machine Learning Methods in the Assessment of Programming Assignments 1 Introduction 2 Related Work 3 Dataset 3.1 Preliminary Data Analysis 4 Model Exploration 4.1 Token Count Based Models (Feature Set 2) 4.2 Language Based Approach - Token Sequences (Feature Sets 3 and 4) 5 Discussion of Results 6 Conclusion and Further Work References Ontology-Aware Biomedical Relation Extraction 1 Introduction 2 Methods 2.1 Token Embedding 2.2 Type Embedding 2.3 Ontology Graph Embeddings 2.4 UMLS Graph Embeddings 2.5 Architecture 3 Implementation and Results 3.1 Ablation 4 Discussion 5 Conclusion References Balancing the Style-Content Trade-Off in Sentiment Transfer Using Polarity-Aware Denoising 1 Introduction 2 Related Work 3 Approach 3.1 Task Definition 3.2 Model Overview 4 Model Variants 4.1 Back-translation 4.2 Our Baseline Models 4.3 Polarity-Aware Denoising 5 Experiments 5.1 Datasets 5.2 Training Setup 5.3 Automatic Evaluation Metrics 5.4 Human Evaluation 5.5 Results 6 Conclusions and Future Work References A Self-Evaluating Architecture for Describing Data 1 Introduction 2 How Linguoplotter Works 2.1 Macro-level Processes in Linguoplotter 2.2 Developing the Program 3 Performance of Linguoplotter 3.1 Method for Evaluation by Human Judges 3.2 Results of Human Evaluation 3.3 Discussion 4 Future Improvements 5 Conclusion References A Novel Hybrid Framework to Enhance Zero-shot Classification 1 Introduction 2 Related Work 3 Problem Formulation 4 An Unified Framework 4.1 Label Name Expansion Using GPT2 4.2 Document Augmentation Using T5 4.3 Integration of Modules 5 Experiments 5.1 Datasets 5.2 Benchmark Solutions 5.3 Main Results and Analysis 5.4 Comparison with Alternative Methods 6 Conclusion and Future Work References Attention-Based Model for Accurate Stance Detection 1 Introduction 2 Related Work 3 Proposed Model 3.1 Stage 1: Features Representation 3.2 Stage 2: Multi-Head Attention 4 Experiment 4.1 Dataset 4.2 Baseline and State-of-the-Art Models 4.3 Training and Hyper-parameters 5 Results and Analysis 5.1 Error Analysis 6 Conclusion References OPTICS: Automatic MT Evaluation Based on Optimal Transport by Integration of Contextual Representations and Static Word Embeddings 1 Introduction 2 Related Work 2.1 Metrics Using Pre-trained Models 2.2 Metrics Using Fine-Tuned Models 2.3 Metrics Using Neural Models 3 New Automatic Metric: OPTICS 3.1 Learning of Encoder for Extracting Contextual Representations 3.2 Score Calculation Based on Optimal Transport Using Contextual Representations and Static Word Embeddings 4 Experiments 4.1 Experiment Data and Procedure 4.2 Experiment Results and Discussion 5 Conclusion References Exploration of Multi-corpus Learning for Hate Speech Classification in Low Resource Scenarios 1 Introduction 2 Proposed Methodology 2.1 Objective of Multi-Task Learning 2.2 Our Approach for Multi-Corpus Learning 2.3 Domain Adaptation Using Multi-Corpus Learning 3 Experimental Setup 3.1 Corpora 3.2 Dataset Split 3.3 Input Pre-processing 3.4 Multi-corpus Model and Training Description 4 Results and Discussion 4.1 Multi-Corpus Learning 4.2 Multi-Corpus Learning in Low-Resource Scenarios 4.3 Domain Adaptation Using Multi-Corpus Learning Approach 5 Conclusion References On the Importance of Word Embedding in Automated Harmful Information Detection 1 Introduction 2 Related Work 2.1 Fake News Detection 2.2 Hate Speech Detection 3 Data 3.1 Fake News Data 3.2 Hate Speech Data 4 Experiments 4.1 Pre-processing 4.2 Classification Models 5 Results 6 Conclusion and Future Works References BERT-based Classifiers for Fake News Detection on Short and Long Texts with Noisy Data: A Comparative Analysis 1 Introduction 2 Literature Review 3 Fake News Datasets 3.1 Description of Datasets 3.2 Text Preprocessing 3.3 Noisy Data 4 Models and Their Combination 4.1 BERT-based Models 4.2 Ensembling with an Additional Class 5 Experiments 5.1 Options and Platforms 5.2 Experiments with Different Models 5.3 Experiments with Noisy Data 5.4 Benchmark 6 Conclusion References Can a Machine Generate a Meta-Review? How Far Are We? 1 Introduction 2 Related Work 3 Dataset Description 4 Meta-Review Generation: A Text Summarization Struggle 4.1 Summarization Approaches 4.2 Evaluation Metrics 5 How Sentiment Plays a Part? 6 How Aspect Plays a Part? 7 Observation 8 Conclusion References Speech Autoblog 2021: The Importance of Language Models for Spontaneous Lecture Speech 1 Introduction 1.1 Related Works 2 Materials and Methods 2.1 Acoustic Model 2.2 Autoblog Data 2.3 Language Modelling 2.4 Datasets 3 Results 3.1 Intrinsic Language Model Evaluation 3.2 Extrinsic Language Model Evaluation 4 Discussion 5 Conclusion References Transformer-Based Automatic Speech Recognition of Formal and Colloquial Czech in MALACH Project 1 Introduction 2 MALACH Project 3 Formal vs. Colloquial Czech 4 Wav2Vec 2.0 5 Experimental Setup 5.1 Pretraining 5.2 Fine-tuning 5.3 Decoding 5.4 Evaluation 6 Results 7 Conclusion References Wakeword Detection Under Distribution Shifts 1 Introduction 2 Distribution Shifts in Wakeword Data 3 On-device Wakeword Detection Models 4 Proposed SSL Algorithm 5 Experimental Setup 6 Results 6.1 Far-field Distribution Shifts with CNN 6.2 Far and Near-Field Distribution Shifts with FCN 7 Conclusions References End-to-End Parkinson's Disease Detection Using a Deep Convolutional Recurrent Network 1 Introduction 2 Data 2.1 PC-GITA 2.2 Independent Test Set 3 Methods 3.1 1D Convolutional Layer 3.2 Temporal Max Pooling 3.3 Recurrent LSTM Layer 3.4 Network's Topology 4 Experiments and Results 4.1 Parameters Optimization 4.2 Bi-class and Multi-class Classification 4.3 Classification Using the Independent Test Set 5 Conclusions References Lexical Bundle Variation in Business Actors' Public Communications 1 Introduction 2 Literature Review 3 Lexical Bundle Methodology 3.1 Lexical Bundles 4 Lexical Bundle Experiments 5 Sentiment Lexical Bundle Experiments 6 Conclusion References 50 Shades of Gray: Effect of the Color Scale for the Assessment of Speech Disorders 1 Introduction 2 Methods 2.1 Data 2.2 Model Description 2.3 Optimization and Regression 3 Experiments and Results 3.1 Parkinson's Disease 3.2 Cleft Lip and Palate 4 Discussion and Conclusions References Sub 8-Bit Quantization of Streaming Keyword Spotting Models for Embedded Chipsets 1 Introduction 2 Small Footprint, Streaming, State-Free KWS Models 3 Relevant Embedded Chipsets 4 Proposed 2-Stage QAT Algorithm 4.1 Quantization Overview 4.2 First Stage: tanh(.) Quantization of Weights 4.3 Second Stage: Linear Quantization of Full Network 5 Experimental Setup 6 Results 6.1 Non-binary Sub-8 Bit Models: First Stage Training 6.2 Non-binary Sub-8 Bit Models: Second Stage Training 6.3 Binary Weight Models 7 Conclusions References Detection of Prosodic Boundaries in Speech Using Wav2Vec 2.0 1 Introduction 2 Data 3 Model for Text-Based Detection 4 Model for Audio-Based Detection 4.1 Influence of Intermediate Phrase Boundaries 4.2 Post-processing 5 Results 5.1 Evaluation Measures 5.2 Audio-Based Evaluation 5.3 Text-Based Evaluation 6 Discussion 7 Conclusion and Future Work References Text-to-Text Transfer Transformer Phrasing Model Using Enriched Text Input 1 Introduction 2 Data and Model Description 2.1 T5 Phrasing Model 2.2 Evaluation Measures 3 Experiments and Results 3.1 Former Experiment 3.2 Phrasing Models with Enriched Text Input 3.3 Results 4 Conclusion References Lexicon-based vs. Lexicon-free ASR for Norwegian Parliament Speech Transcription 1 Introduction 2 State-of-the-Art and Research Goals 3 Lexicon-based ASR System for Norwegian 4 Lexicon-free System 5 Experiments 5.1 Data 5.2 Tests and Results 5.3 Discussions 6 Conclusions and Future Work References On Comparison of Phonetic Representations for Czech Neural Speech Synthesis 1 Introduction 2 Research Questions 3 Speech Data 4 Phonetic Representations 4.1 Full Phonetic Alphabet 4.2 Reduced Phonetic Alphabet 4.3 Pause-free Phonetic Alphabet 5 Speech Synthesis Models 5.1 FastSpeech2 + Multi-band MelGAN 5.2 Glow-TTS + UnivNet 5.3 VITS 6 Results and Discussion 6.1 Listening Tests 6.2 Full vs Reduced Phonetic Alphabet (RQ1) 6.3 Full vs Pause-free Alphabet (RQ2) 7 Conclusions and Future Work References The Influence of Dataset Partitioning on Dysfluency Detection Systems 1 Introduction 2 Data 3 Methods 3.1 Classification Experiments 3.2 ECAPA-TDNN 3.3 Metadata Retrieval 3.4 Quality Criteria 4 SEP-28k-Extended 4.1 Speaker Imbalance 4.2 Data Partitioning Considering Metadata 5 Experiments 5.1 Results 5.2 Discussion 6 Conclusion References Going Beyond the Cookie Theft Picture Test: Detecting Cognitive Impairments Using Acoustic Features 1 Introduction 2 Data 2.1 Standardized Sub-Tests 2.2 Clinical Interview 3 Methods 3.1 openSMILE 3.2 wav2vec 2.0 4 Experiments 4.1 Results 5 Discussion 6 Conclusion References Dialogue Federated Learning in Heterogeneous Data Settings for Virtual Assistants – A Case Study 1 Introduction 2 Related Work 2.1 Federated Learning 2.2 Tackling Heterogeneity in Federated Learning 3 Materials and Methods 3.1 Our Case Study 3.2 Datasets Used 3.3 Emulating Heterogeneous Condition 3.4 Evaluation of Local Optimizers 4 Results 4.1 Results on Non-IID Data Distributions 4.2 Choosing the Best Local Optimizer 5 Conclusions References PoCaP Corpus: A Multimodal Dataset for Smart Operating Room Speech Assistant Using Interventional Radiology Workflow Analysis 1 Introduction 2 Related Works 3 Data Collection Procedure and Setup 3.1 Port-Catheter Placement 3.2 Data Collection Setup 4 Dataset 4.1 Data Structure 4.2 Alignment 4.3 Transcription 5 Discussion 6 Conclusion References Investigating Paraphrasing-Based Data Augmentation for Task-Oriented Dialogue Systems 1 Introduction 2 Related Work 3 Datasets 4 Methodology 4.1 Generating Paraphrases 4.2 Intrinsic Evaluation – Utterance Quality 4.3 Extrinsic Evaluation – NLU System 5 Results 5.1 NLU Performance with Augmented Data 5.2 Reduction of Manually Annotated Training Data 5.3 Comparison Between GPT-2 and CVAE Models 5.4 Limitations on Intent Preservation 5.5 Qualitative Analysis 6 Conclusions and Future Work References Transfer Learning of Transformers for Spoken Language Understanding 1 Introduction 2 Transfer Learning for Spoken Dialog Systems 2.1 Wav2Vec 2.0 Transformer 2.2 Text-to-Text Transfer Transformer 3 Speech Recognition 3.1 DNN-HMM Baseline 3.2 Wav2Vec 2.0 Recognizer 4 Spoken Language Understanding 4.1 CNN SLU Baseline 4.2 T5 SLU 5 Dataset Description 6 Experimental Evaluation 7 Conclusion References Evaluation of Wav2Vec Speech Recognition for Speakers with Cognitive Disorders*-12pt 1 Introduction 2 Spoken Dialog System 2.1 SpeechCloud 2.2 Spoken Dialog 2.3 Dialog Scenarios 3 Wav2Vec 2.0 Speech Recognition 4 Datasets (Participants and Their Inclusion Criteria) 4.1 Mobile Phone Recordings 4.2 Dialog System Recordings 5 Experimental Evaluation 6 Conclusion and Future Work References Fine-Tuning BERT for Generative Dialogue Domain Adaptation 1 Introduction 2 Generative Dialogue Domain Adaptation 2.1 Training Utterances for Fine-Tuning 2.2 Fine-Tuning: Choice of BERT 2.3 Slot-Values Generation 2.4 Slot-Value Substitution Parameters 3 Experimental Setting 3.1 Source and Target KBs 3.2 Datasets Implementation 3.3 Target Test Set 4 Evaluation 5 Results 6 Conclusion References Empathy and Persona of English vs. Arabic Chatbots: A Survey and Future Directions 1 Introduction 2 Research Challenges 2.1 Emotion vs Sentiment 2.2 Arabic Preprocessing Challenges 3 Emotion- and Persona-Based Chatbots 4 Learning Models for Chatbots 4.1 Deep Learning Models 4.2 Reinforcement Learning (RL) Models 5 Open Problems 6 Conclusion References Author Index