Machine Translation: 18th China Conference, CCMT 2022, Lhasa, China, August 6–10, 2022, Revised Selected Papers

Description

This book constitutes the refereed proceedings of the 18th China Conference on Machine Translation, CCMT 2022, held in Lhasa, China, during August 6–10, 2022. The 16 full papers were included in this book were carefully reviewed and selected from 73 submissions. Preface Organization Contents PEACook: Post-editing Advancement Cookbook 1 Introduction 2 Related Work 2.1 APE Problem and APE Metrics 2.2 APE Baselines 3 PEACook Corpus 3.1 PEACook Corpus Details 4 Baseline Model Experiments 4.1 Pre-training AR-APE Model 4.2 Fine-Tuning AR-APE Model 4.3 Pre-training NAR-APE Model 4.4 Fine-Tuning NAR-APE Model 5 Conclusion References Hot-Start Transfer Learning Combined with Approximate Distillation for Mongolian-Chinese Neural Machine Translation 1 Introduction 2 Background 2.1 NMT 2.2 Transfer Learning 2.3 Pre-train Techniques 3 Methods 3.1 Word Alignment Under Hot-Start 3.2 Approximate Distillation 4 Experiment 4.1 Settings 4.2 Results and Analysis 4.3 Ablation Test 4.4 Case Analysis 5 Conclusion References Review-Based Curriculum Learning for Neural Machine Translation 1 Introduction 2 Related Work 3 Review-Based Curriculum Learning 3.1 Time-Based Review Method 3.2 Master-Based Review Method 3.3 General Domain Enhanced Training 4 Experiment 4.1 Data and Setup 4.2 Main Results 5 Analysis 5.1 Effect of Mixed Fine Tuning 5.2 Low-Resource Scenario 5.3 Data Sharding 5.4 Training Efficiency 6 Conclusion References Multi-strategy Enhanced Neural Machine Translation for Chinese Minority Languages 1 Introduction 2 Dataset 3 System Overview 3.1 Back-Translation 3.2 Alternated Training 3.3 Ensemble 4 Experiments 4.1 Mongolian Chinese 4.2 TibetanChinese 4.3 UyghurChinese 5 Analysis 5.1 The Effect of Different Back-Translation Methods 5.2 The Impact of Sentence Segmentation on the Translation Quality of Machine Translation 5.3 Analysis of BLEU Scores of MongolianChinese Machine Translation on the Development Set 6 Conclusion References Target-Side Language Model for Reference-Free Machine Translation Evaluation 1 Introduction 2 Target-Side Language Model Metrics 3 Experiments 3.1 Datasets and Baselines 3.2 Results 3.3 Discussion 4 Conclusion References Life Is Short, Train It Less: Neural Machine Tibetan-Chinese Translation Based on mRASP and Dataset Enhancement 1 Introduction 2 Prerequisite 2.1 Neural Machine Translation with mRASP 2.2 Diversification Method 2.3 Curvature 3 Methodology 3.1 Overall Structure 3.2 Curvature Based Checkpoint Hijack 4 Experiments 4.1 Dataset Description and Finetune Parameters 4.2 Experiment Result 5 Conclusion References Improving the Robustness of Low-Resource Neural Machine Translation with Adversarial Examples 1 Introduction 2 Background and Related Work 2.1 Neural Machine Translation 2.2 Adversarial Example, Adversarial Attack and Adversarial Training in NLP 2.3 Genetic Algorithm-Based Adversarial Attack 2.4 Gradient-Based Adversarial Attack 3 Adversarial Examples Based on Reinforcement Learning 3.1 Reinforcement Learning 3.2 Environment 3.3 Agent 4 Experiment 4.1 Data Preprocessing 4.2 NMT Model 4.3 Evaluating Indicator 4.4 Adversarial Attack Results and Analysis 4.5 Adversarial Training Results and Analysis 4.6 Ablation Study 5 Conclusion References Dynamic Mask Curriculum Learning for Non-Autoregressive Neural Machine Translation 1 Introduction 2 Background 2.1 Non-autoregressive Neural Machine Translation 2.2 Curriculum Learning 3 Method 3.1 Model 3.2 Dynamic Mask Curriculum Learning 3.3 Train and Inference 4 Experiment 4.1 Data Preparation 4.2 Configuration 4.3 Baseline 4.4 Results 5 Analysis 5.1 Mask Strategy 5.2 Method Generality 6 Conclusion References Dynamic Fusion Nearest Neighbor Machine Translation via Dempster-Shafer Theory 1 Introduction 2 Background 3 Method 3.1 Dempster-Shafer Theory 3.2 Label Smoothing 4 Experiment 4.1 Experimental Setup 4.2 Result and Analysis 4.3 Robustness 4.4 Case Study 5 Conclusion References A Multi-tasking and Multi-stage Chinese Minority Pre-trained Language Model 1 Introduction 2 Related Work 2.1 Pre-trained Language Model 2.2 Multilingual Model 2.3 Chinese Minority Languages 3 Main Methods 3.1 Model Architecture 3.2 Multi-tasking Multi-stage Pre-training 3.3 Model Parameter Details 3.4 Model Setting Details 4 Experiments 4.1 Main Results 4.2 Case Study 5 Conclusion References An Improved Multi-task Approach to Pre-trained Model Based MT Quality Estimation 1 Introduction 2 Related Works 3 PE Based Multi-task Learning for Sentence Level QE 3.1 Multi-task Learning Framework for QE 3.2 PE Based Multi-task Learning QE 3.3 Multi-model Ensemble 4 Experiments 4.1 Dataset 4.2 Model Training and Evaluation Metric 4.3 Experimental Results and Analysis 4.4 Ablation Study 5 Conclusion References Optimizing Deep Transformers for Chinese-Thai Low-Resource Translation 1 Introduction 2 Background 2.1 Transformer 2.2 Low-Resource NMT 2.3 Parameter Initialization for Deep Transformers 2.4 Deep Transformers for Low-Resource Tasks 3 Our Work 3.1 Data Processing 3.2 Exploration of Training Settings 3.3 Deep Transformers for Low-Resource Machine Translation 4 Related Work 5 Conclusion References CCMT 2022 Translation Quality Estimation Task 1 Introduction 2 Estimation System 3 Data 4 Method 4.1 System Training 4.2 System Test 5 Experiment 5.1 System Environment 5.2 Experiment Settings 5.3 Experiment Result 6 Conclusion References Effective Data Augmentation Methods for CCMT 2022 1 Introduction 2 System Architecture 3 Methods 3.1 Data Augmentation 3.2 CE Task and EC Task 3.3 CThai Task and ThaiC Task 4 Experiments 4.1 System Settings 4.2 Data Pre-processing 4.3 Experimental Results 5 Conclusion References NJUNLP's Submission for CCMT 2022 Quality Estimation Task 1 Introduction 2 Methods 2.1 Existing Methods 2.2 Proposed Methods 3 Experiments 3.1 Dataset 3.2 Settings 3.3 Single Model Results 3.4 Ensemble 3.5 Analysis 4 Conclusion References ISTIC's Thai-to-Chinese Neural Machine Translation System for CCMT' 2022 1 Introduction 2 System Architecture 2.1 Baseline System 2.2 Our System 3 Methods 3.1 Back Translation 3.2 Add External Data 3.3 Model Averaging 3.4 Model Ensemble Strategy 4 Experiments 4.1 System Settings 4.2 Data Preprocessing 4.3 Experimental Results 4.4 Conclusion References Author Index