Text, Speech, and Dialogue: 24th International Conference, TSD 2021, Olomouc, Czech Republic, September 6–9, 2021, Proceedings (Lecture Notes in Computer Science)

This book constitutes the proceedings of the 24th International Conference on Text, Speech, and Dialogue, TSD 2021, held in Olomouc, Czech Republic, in September 2021.* The 2 keynote speeches and 46 papers presented in this volume were carefully reviewed and selected from 101 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, and others. * Due to the COVID-19 pandemic the conference was held in a "hybrid" mode. Preface Organization Abstracts of Keynotes Towards User-Centric Text-to-Text Generation: A Survey Wasserstein Autoencoders with Mixture of Gaussian Priors for Stylized Text Generation Use of Deep Learning in Free Speaking Non-native English Assessment Cross-Lingual Knowledge Transfer and Adaptation in Low-Data Regimes: Achievements, Trends, and Challenges Contents Keynote Talks Towards User-Centric Text-to-Text Generation: A Survey 1 Motivation 2 Related Surveys 3 User-Centric NLG 3.1 When Is NLG User-Centric? 3.2 User and Application Domain 3.3 User Information 4 Data for User-Centric NLG 5 User-Centric Generation Models 6 Evaluations 7 Challenges and Opportunities 8 Conclusion References Wasserstein Autoencoders with Mixture of Gaussian Priors for Stylized Text Generation 1 Introduction 2 Related Work 3 Approach 3.1 Wasserstein Autoencoder 3.2 WAE with Gaussian Mixture Model Prior (GMM-WAE) 3.3 Training 3.4 Sentence Generation with GMM-WAE 4 Experiments 4.1 Results 5 Conclusion References Text Evaluating Semantic Similarity Methods to Build Semantic Predictability Norms of Reading Data 1 Introduction 2 Review on Sentence Completion Task and Datasets 3 The Initial Version of the RastrOS Corpus 4 Building the Sentence Completion Dataset via Cloze Data 4.1 Automatically Generating Alternates 4.2 Human Judgment 5 Evaluation Results and the Hybrid Method Proposed 6 Conclusions References SYN2020: A New Corpus of Czech with an Innovated Annotation 1 Introduction 2 New Features 2.1 Sublemma 2.2 Verbtag 2.3 Multiword Tokens 3 Annotation 3.1 Segmentation, Tokenization and Morphological Analysis 3.2 Segmentation Adjustment 3.3 Tokenization Adjustment 3.4 Applying the Verbtags 3.5 Rule-Based Disambiguation 3.6 Completing the Disambiguation with a Tagger 3.7 Parsing 3.8 Sublemmas and Multiword Tokens 4 Annotation Accuracy 4.1 Results of the Automatic Testing: Etalon Data 4.2 Manual Comparison of the Hybrid Disambiguation and the Tagger only Method 4.3 Parsing Accuracy 4.4 Auxiliaries: Verbtag and Parsing 5 Conclusions References Deep Bag-of-Sub-Emotions for Depression Detection in Social Media 1 Introduction 2 Related Work 3 Deep Bag-of-Sub-Emotions 3.1 Model Description 3.2 Probabilistic Intepretation of the Differentiable Bag-of-Features 3.3 Learning Sub-Groups of Emotions 4 Experimental Settings 4.1 Datasets Description 4.2 Learning Approaches 4.3 Hyperparameter Selection 5 Results and Analysis 6 Conclusions and Future Work References Rewriting Fictional Texts Using Pivot Paraphrase Generation and Character Modification 1 Introduction 2 Related Work 3 Approach 3.1 Paraphrase Generation 3.2 Character Modification 4 Evaluation 4.1 Dataset 4.2 Performance of NER and CR 4.3 Text Readability 4.4 User Study 4.5 Detecting Spelling and Grammar Issues 5 Conclusion References Transformer-Based Automatic Punctuation Prediction and Word Casing Reconstruction of the ASR Output 1 Introduction 2 Proposed Models 2.1 SentencePiece Model 2.2 BERT Pre-training 2.3 T5 Pre-training 3 Experimental Evaluation 4 Results 5 Conclusions References A Database and Visualization of the Similarity of Contemporary Lexicons 1 Introduction 2 State of the Art 3 Automated Similarity Computation 4 Comparison to Results from Lexicostatistics 5 Automated Visualization 6 Conclusions and Future Work References The Detection of Actors for German 1 Introduction 2 Actor Noun Identification 3 Improving the Actor List 4 Polar Actor Noun Phrases 5 Related Work 6 Conclusions References Verbal Autopsy: First Steps Towards Questionnaire Reduction 1 Introduction 2 Materials: Verbal Autopsy GS 3 Methods 3.1 Semantic Similarity 3.2 Automatic Selection of Closed Questions 3.3 Cause of Death Extraction Approach 4 Experimental Results 4.1 Identification of Candidate Questions 4.2 CoD Prediction Reducing CQs 4.3 Promoting Reproducible Research: Software Release 5 Conclusions References Effective FAQ Retrieval and Question Matching Tasks with Unsupervised Knowledge Injection 1 Introduction 2 Approach to FAQ Retrieval and Question Matching 2.1 The q-Q Similarity Measure 2.2 The q-A Relevance Measure 2.3 Voting Mechanism 2.4 Supervised and Unsupervised Knowledge Injections for the q-A Relevance Measure 2.5 Unsupervised Knowledge Injections for the q-A Relevance Measure 3 Experimental Datasets and Setup 4 Evaluation Metrics 5 Experimental Results 6 Conclusion and Future Work References Exploring Conditional Language Model Based Data Augmentation Approaches for Hate Speech Classification 1 Introduction 2 Data Augmentation 2.1 Conditional Language Modeling 2.2 Proposed Methodology 3 Experimental Setup 3.1 Data Description 3.2 Data Preprocessing 3.3 Model Parameters 4 Results and Discussion 4.1 Improvements with Data Augmentation 4.2 Quality of Augmented Data 4.3 Influence of Filtering the Samples 5 Conclusion References Generating Empathetic Responses with a Pre-trained Conversational Model 1 Introduction 2 Related Work 3 Method 3.1 Dataset 3.2 Fine-Tuning DialoGPT Model 3.3 Evaluation 4 Results and Discussion 4.1 Automatic Evaluation 4.2 Human Evaluation 4.3 Empathetic Qualities 5 Conclusion References Adaptation of Classic Readability Metrics to Czech 1 Introduction 2 Related Work 2.1 Readability Metrics 2.2 Language-Specific Adaptations of Readability Metrics 3 Data 3.1 Pre-processing 4 Method 4.1 Determining Language-Specific Function Parameters 5 Results 6 Discussion 7 Conclusion References New Parallel Corpora of Baltic and Slavic Languages — Assumptions of Corpus Construction 1 Introduction 2 State of the Art 3 Methodology 3.1 Selection of Texts 3.2 Pre-processing of Texts 3.3 Segmentation 3.4 Automatic Lemmatisation and Tagging the Inflectional Layer of Tokens 3.5 Manual Lemmatisation and Tagging of the Forms Not Detected by the Automatic Lemmatisation and Tagging Process 3.6 Metadata 3.7 Converting Resources into a Format Compatible with the KonText and Publishing Corpora 4 Parallel Corpora of Baltic and Slavic Languages on the CLARIN-PL Website (2021 Version) 4.1 Corpora Including the Polish Language 4.2 New Corpus Resources for Slavic and Baltic Languages Without Polish 5 Conclusions and Future Work References Use of Augmentation and Distant Supervision for Sentiment Analysis in Russian 1 Introduction 2 Related Work 3 Russian Sentiment Benchmark Datasets 4 Automatic Generation of Annotated Data 4.1 Augmentation 4.2 Distant Supervision Approach 5 Text Preprocessing 6 BERT Architectures 7 Experiments and Results 7.1 Use of Augmentation 7.2 Use of Distant Supervision 7.3 Difficult Examples 8 Conclusion References RobeCzech: Czech RoBERTa, a Monolingual Contextualized Language Representation Model 1 Introduction 2 Related Work 3 Training the Czech RoBERTa 4 Evaluation Tasks 4.1 Morphological Tagging and Lemmatization on PDT 3.5 4.2 Dependency Parsing on PDT 3.5 4.3 Morphosyntactic Analysis on Universal Dependencies 4.4 Named Entity Recognition 4.5 Semantic Parsing on Prague Tectogrammatical Graphs 4.6 Sentiment Analysis 5 Results 6 Conclusion References Labelled EPIE: A Dataset for Idiom Sense Disambiguation 1 Introduction 2 Related Work 3 Data 4 Model 5 Experiments 6 Results 7 Future Work 8 Conclusion References Introducing NYTK-NerKor, A Gold Standard Hungarian Named Entity Annotated Corpus 1 Introduction 2 Related Work 3 The Size of the Corpus 4 Genres and Sources 5 The Format of the Annotation 6 Named Entity Annotation 6.1 Annotation Principles 7 Morphological Annotation 7.1 The emMorph Tagset 7.2 The UD Tagset 8 The Workflow of Annotation 8.1 The Workflow of NE Annotation 8.2 The Workflow of Morphological Annotation References Semantic Templates for Generating Long-Form Technical Questions 1 Introduction 2 Related Work 3 Question Templates 4 Using Templates for Distant Supervision 5 Empirical Evaluation 6 Conclusion References Rethinking Adversarial Training for Language Adaptation 1 Introduction 2 Related Work 3 Adversarial Training for Cross-Lingual NLI 3.1 Experimental Setup 3.2 Results 3.3 Analysis 4 Towards Robust Auxiliary Tasks for Language Adaptation 4.1 Masked Language Modeling 4.2 Translation Language Modeling 4.3 Combining Adversarial Training and Language Modeling 5 Conclusions References A Corpus with Wavesurfer and TEI: Speech and Video in TEITOK 1 Introduction 2 Visualization 2.1 Document View 2.2 Wavesurfer View 3 Searching 3.1 Map Based Searches 4 Editing 4.1 Converting 4.2 Annotating 4.3 Transcribing 5 Conclusion References Exploiting Subjectivity Knowledge Transfer for End-to-End Aspect-Based Sentiment Analysis 1 Introduction 2 Related Work 2.1 Separate Methods 2.2 Unified Methods 3 Opinion Typology 3.1 Tasks Definitions 3.2 Dataset 4 Evaluation Methodology 4.1 Baselines 4.2 Architecture 5 Experiments 5.1 Data 5.2 Evaluation Metrics 5.3 Results 6 Conclusion References Towards Personal Data Anonymization for Social Messaging 1 Introduction 2 Related Work 3 Dataset 4 Methods 4.1 Annotation Schema 4.2 Personal Data Detection 4.3 Anonymization 5 Results 5.1 Detection Efficiency 5.2 Data Utility 5.3 Performance Discussion 6 Limitations and Further Work 7 Conclusions References ParCzech 3.0: A Large Czech Speech Corpus with Rich Metadata 1 Introduction 2 Related Work 3 Methodology 3.1 Data Gathering 3.2 Data Processing and Annotation 3.3 Sentence- and Word-Level Alignment with Speech 4 Corpus Description 4.1 Source HTML and Audio Files 4.2 Version for ASR 4.3 TEI Encoding 4.4 Division into Training, Development and Test Sets 4.5 Corpus Statistics 5 Conclusion References Using Zero-Shot Transfer to Initialize azWikiNER, a Gold Standard Named Entity Corpus for the Azerbaijani Language 1 Introduction 2 Literature Review 3 Materials and Methods 4 Existing Solutions 5 Corpus Annotation Method 6 Evaluation 7 Limitations 8 Conclusion References Using BERT Encoding and Sentence-Level Language Model for Sentence Ordering 1 Introduction 2 Related Work 3 Methodology 3.1 Task Formulation 3.2 Sentence-Level Language Model for Sentence Ordering (SLM) 4 Experiment 4.1 Dataset 4.2 Baselines and Competitors 4.3 Metrics 4.4 Results and Analysis 5 Conclusion References Using Presentation Slides and Adjacent Utterances for Post-editing of Speech Recognition Results for Meeting Recordings 1 Introduction 2 Related Work 3 Proposed Method 3.1 System Architecture 4 Experiments 4.1 Data Preparation 4.2 Methods for Comparison 4.3 Evaluation Metrics 4.4 Results 5 Summary and Future Work References Leveraging Inter-step Dependencies for Information Extraction from Procedural Task Instructions 1 Introduction 2 Related Work 3 Model 3.1 Token-Level Representation 3.2 Step-Level Representation 3.3 Leveraging Step-Level Information 3.4 Classifying the Final Representations 4 Experiments and Results 4.1 Datasets 4.2 Experiments 4.3 Analysis of Sequential Dependencies 5 Conclusion References Speech DNN-Based Semantic Rescoring Models for Speech Recognition 1 Introduction 2 Proposed Methodology 2.1 Introduction 2.2 DNN-Based Rescoring Models 3 Experimental Conditions 3.1 Corpus Description 3.2 Recognition System 3.3 Rescoring Models 4 Experimental Results 4.1 Impact of Hyperparameters 4.2 Global Results 5 Conclusion References Identification of Scandinavian Languages from Speech Using Bottleneck Features and X-Vectors 1 Introduction 1.1 Motivation for This Work 2 State-of-the-Art Approaches to LID 3 Datasets and Metrics 4 Baseline i-Vector System and Direct Methods 4.1 Feed-Forward FC DNN 4.2 TDNN Architecture 4.3 FSMN Architecture 4.4 Results 5 BTN-Based Approaches 5.1 Monolingual Feed-Forward FC BTNs 5.2 Multilingual FSMN-Based BTNs 5.3 Results 6 The Use of X-Vectors 6.1 Results 7 Conclusions References LSTM-XL: Attention Enhanced Long-Term Memory for LSTM Cells 1 Introduction 2 Methods 2.1 LSTM 2.2 Global and Local Attention 3 LSTM-XL 4 Training Corpora 5 Experimental Setup 5.1 Subword Language Models 6 Results 6.1 Cost Analysis 6.2 Finnish 6.3 Hungarian 7 Conclusions References Improving RNN-T ASR Performance with Date-Time and Location Awareness 1 Introduction 2 Prior Work 3 Context Representation with RNN-T 3.1 Date Time as Context 3.2 Location as Context 3.3 Combination of Context 4 Data and Experimental Setup 4.1 Datasets 4.2 Experimental Setup 5 Results and Discussion 5.1 Overall WER Comparison 5.2 Domain-Wise WER Comparison 5.3 Combined Context 5.4 Low Resource Simulation 5.5 Month and State-Wise WER Comparison 5.6 Baseline and Contextual Model Outputs 5.7 t-SNE Plot Analysis 6 Conclusions References BrAgriSpeech: A Corpus of Brazilian-Portuguese Agricultural Reported Speech 1 Introduction 2 Related Work 3 Extraction Methodology 4 Evaluation 5 BrAgriSpeech's Linguistic Characteristics 6 Resource Overview 7 Conclusion References Exploiting Large-Scale Teacher-Student Training for On-Device Acoustic Models 1 Introduction 2 System Description 3 SSL for Small Footprint AM 3.1 Learning Curves on Large Unsupervised Data 3.2 Learning Saturation for On-Device SSL System 3.3 Improving the Efficiency of Knowledge Distillation 4 SSL in Low Resource Settings 4.1 Accuracy Gains with Unsupervised Data 4.2 An Analysis of Empirical Risk with Student-Teacher Learning 5 Discussion 6 Conclusions and Future Work References An AI-Based Detection System for Mudrabharati: A Novel Unified Fingerspelling System for Indic Scripts 1 Introduction 2 Fingerspelling Convention for Mudrabharati 3 Detection System for Mudrabharati 4 Working of the System 4.1 Calibration Step 4.2 Character Prediction 4.3 Experimental Results 5 Conclusion References Is There Any Additional Information in a Neural Network Trained for Pathological Speech Classification? 1 Introduction 2 Data 3 Methods 3.1 Pre-processing 3.2 Training 3.3 Embedding Extraction 3.4 Classification 4 Experiments and Results 4.1 Base Models 4.2 Classification of Age, Gender and Task 4.3 Classification of Language 5 Conclusion References On Comparison of XGBoost and Convolutional Neural Networks for Glottal Closure Instant Detection 1 Introduction 2 Experimental Data 3 Feature Extraction 4 Methods of Evaluation 5 Experiments and Results 6 Conclusions References Emotional State Modeling for the Assessment of Depression in Parkinson's Disease 1 Introduction 2 Datasets 2.1 Parkinson's Disease Assessment 2.2 Assessment of Depression in Parkinson's Disease 2.3 Auxiliary Dataset: IEMOCAP 3 Methods 3.1 PAD Emotional Model 3.2 Deep Learning Based Model 3.3 Baseline Features 3.4 Optimization and Classification 4 Experiments and Results 4.1 Experiment 1: PD Vs. HC Subjects 4.2 Experiment: Depressive Vs. Non-Depressive PD Patients 5 Discussion and Conclusions References Attention-Based End-to-End Named Entity Recognition from Speech 1 Introduction 2 Data 3 Methods 3.1 Pipeline NER Systems 3.2 Baseline ASR System 3.3 Augmented Labels Approach 3.4 Multi-task Approach 4 Experiments 5 Results 5.1 Finnish Results 5.2 Swedish Results 5.3 English Results 6 Analysis of the Results 7 Conclusion References Incorporation of Iterative Self-supervised Pre-training in the Creation of the ASR System for the Tatar Language 1 Introduction 2 System Description 3 Data Collection 4 Experiments 5 Conclusion References Speakers Talking Foreign Languages in a Multi-lingual TTS System 1 Introduction 2 TTS Implementation 2.1 Phonetic Alphabet 3 Single-Voice System with Phonetic Mapping 4 Experiments and Results 4.1 Experimental Speech Data 4.2 SUS Listening Test 4.3 Preference Listening Test 5 Conclusion References Voice Activity Detection for Ultrasound-Based Silent Speech Interfaces Using Convolutional Neural Networks 1 Introduction 2 The Ultrasound-Based SSI Framework 3 Voice Activity Detection from Speech and from Ultrasound 4 Experimental Setup 4.1 The Ultrasound Dataset 4.2 CNNs for the SSI and for the VAD Task 5 Results and Discussion 5.1 The Impact of VAD on the MCD Metric and on Speech Synthesis 5.2 The Impact of VAD on the SSI 5.3 Classification of Speech and Silence from Ultrasound 5.4 Replacing Speech-VAD by UTI-VAD 6 Conclusion References How Much End-to-End is Tacotron 2 End-to-End TTS System 1 Introduction 2 Tacotron 2 Details 3 Tests Design 4 Evaluation 5 Conclusion References CNN-TDNN-Based Architecture for Speech Recognition Using Grapheme Models in Bilingual Czech-Slovak Task 1 Introduction 2 Differences and Similarities Between Czech and Slovak 2.1 Graphemes 2.2 Phonemes 3 Training and Test Data 4 Experimental Setup 4.1 Acoustic Feature Extraction 4.2 Acoustic Modeling 4.3 Language Modeling 4.4 Decoding 5 Experiments and Results 6 Conclusion References Dialogue A Multimodal Model for Predicting Conversational Feedbacks 1 Introduction 2 Related Works 3 Feedbacks: Typology, Cues 4 The Dataset 5 Data Analysis 6 The Model 7 Results 8 Conclusion References Estimating Social Distance Between Interlocutors with MFCC-Based Acoustic Models for Vowels 1 Introduction 2 Material 3 Method 3.1 Acoustic Modelling 3.2 Distance Calculation 3.3 Statistical Analysis 4 Results and Discussion 5 Conclusions References Remote Learning of Speaking in Syntactic Forms with Robot-Avatar-Assisted Language Learning System 1 Introduction 2 Robot-Avatar-Assisted Language Learning System 3 Collection of Gaze Data 4 Experiment 4.1 Experimental Methods 4.2 Evaluation of Learning Effect 4.3 Analysis of Gaze Data 5 Discussion 6 Conclusion References Author Index