Machine Translation with Minimal Reliance on Parallel Resources

1 Preliminaries 1
1.1 Challenges in MT—Relevance to the European Environment 1
1.2 A Brief Review of MT Development History 3
1.3 Advantages and Disadvantages of Main MT Paradigms 4
1.4 The PRESEMT Methodology in a Nutshell 7
1.5 Closing Note on Implementation 9
References 9

2 Implementation 11
2.1 Introduction: Summary of the Approach 11
2.2 Linguistic Resources: Data and Existing Linguistic Tools 13
2.2.1 External Processing Tools 13
2.2.2 Lemma-Based Bilingual Dictionary 15
2.2.3 The Parallel Corpus 15
2.2.4 The TL Monolingual Corpus 17
2.3 Processing the Parallel Corpus 19
2.3.1 Phrase Aligner Module 19
2.3.2 Phrasing Model Generation 24
2.4 Creating a Language Model for the Target Language 26
References 27

3 Main Translation Process 29
3.1 Introduction 29
3.2 Translation Phase One: Structure Selection 30
3.2.1 The Dynamic Programming Algorithm 32
3.2.2 Example of How Structure Selection Works 34
3.3 Phase Two: Translation Equivalent Selection 35
3.3.1 Applying the Language Model to the Task 37
3.3.2 Example of How TES Works 39
References 40


4 Assessing PRESEMT 43
4.1 Evaluation Dataset 44
4.2 Objective Evaluation Metrics 44
4.3 System Evaluation 45
4.3.1 Evaluation Objectives 45
4.3.2 Evaluation Results 46
4.3.3 Expanding the Comparison 47
4.3.4 Experimenting with Further Data 47
4.4 Comparing PRESEMT to Other MT Systems 49
4.5 Conclusions 52
References 53

5 Expanding the System 55
5.1 Preparing the System for New Language Pairs 56
5.2 Examining Language-Pair-Specific Issues 57
5.2.1 Agreement Within a Nominal Phrase 58
5.2.2 Case Mismatches 58
5.2.3 The Null-Subject Parameter 58
5.2.4 Word Order 59
5.3 Notes on Implementation 60
5.4 Conclusions 60
References 61

6 Extensions to the PRESEMT Methodology 63
6.1 Splitting SL Sentences into Phrases More Accurately 63
6.1.1 Design and Implementation of TEM 64
6.1.2 Experimental Evaluation 67
6.1.3 Conclusions 68