Breaking Down the Language Barrier with Statistical Machine Translation

Advanced Research Seminar I/III
Graduate School of Information Science
Nara Institute of Science and Technology
January/february 2014

Instructor:

Course Description

Machine translation is a technology to automatically translate from one language to another, and is becoming more and more widely used as the accuracy is reaching more usable levels. In this course I will introduce the basics behind machine-translation in a hands-on manner. At the end of each class there will be an exercise where you try out what you learned on real data.

Prerequisites: basic probability, programming skills.
Optional Textbooks:

Course Schedule

Jan 28, 30. Feb 4, 6. (9:20-10:50am) @ IS Building Room L2

• Lecture 1 (Jan 28): Language Modeling
• Slides (pdf, updated after class), Video (forgot to turn on the mic... sorry!)
• Assignment Data/Code: Download
• Recommended reading: "Language Model" chapters of both textbooks. LM Smoothing, Class LM, Neural Network LM, Discriminative LM

Results
#	Username	LM Entropy	Method
1	clusthermal	7.476017	Kneser Ney Bigram
2	Spivak	7.481626	Linear Interpolation Trigram
3	Colorful black ideas (3p. team)	7.562	Modified Absolute Discounting Bigram
4	tmrevolutionx/delacroix/abhi	7.710441	Witten-Bell Bigram
5	yemantu	7.733526	Witten-Bell Bigram
6	s10018	7.747171	Witten-Bell Bigram
7	gacon	7.750392	Witten-Bell Bigram
8	pongsakorn-u	7.818588	Linear Interpolation Bigram (adjusted weights)
9	1351213	7.82259	Linear Interpolation Bigram (adjusted weights)
9	Lenessia-Erhart-Cowen	7.8226	Linear Interpolation Bigram (adjusted weights)
11	Baseline	7.883958	Linear Interpolation Bigram



• Lecture 2 (Jan 30): Alignment/Translation Modeling
• Slides (pdf, updated after class), Video
• Assignment Data/Code: Download
• Recommended reading: "Word-based Models" chapter of SMT, "Alignment" chapter of 機械翻訳. Both chapters on phrase-table/reordering-model construction. Systematic Comparison of Alignment Models, Inversion Transduction Grammars, Supervised Alignment, Bilingual Class Induction

Results
#	Username	F-Measure	Precision	Recall	Method
1	Colorful black ideas	0.453471	0.747608	0.325434	Dice/Max-Score/Grow-Diag
2	s10018	0.444331	0.772623	0.311832	Dice/Max-Score/Diag
3	Lenessia-Erhart-Cowen	0.429068	0.870318	0.284717	Dice+Absolute Discounting/Max-Score/Intersect
4	tmrevolutionx/delacroix/abhi	0.425522	0.491179	0.375348	Dice/Competitive Linking
5	gacon	0.424353	0.70838	0.302903	Dice/Competitive Linking
6	pongsakorn-u	0.420217	0.848739	0.279234	Dice/Max-Score/Diag
7	Spivak	0.420003	0.838293	0.280193	Dice/Max-Score+Threshold/Intersect
8	Baseline	0.419585	0.857011	0.277796	Dice/Max-Score/Intersect
9	clusthermal/1351213	0.414811	0.779329	0.28262	Model 1/Max-Score/Intersect
10	yemantu	0.394243	0.467954	0.340594	Dice/Competitive Linking



• Lecture 3 (Feb 4): Phrase-based MT
• Slides (pdf), Video
• Assignment Data/Code: Download
• Recommended reading: "Phrase-based Models" chapter of SMT and 機械翻訳. Phrase Based MT


• Lecture 4 (Feb 6): MT Evaluation/Optimization for MT
• Slides (pdf), Video
• Recommended reading: "Optimization" and "tree-based models" chapters of SMT and 機械翻訳. MERT, Hiero, Tree-to-String, Japanese-English Preordering

Final Results
#	Username	BLEU	Method
1	clusthermal/gacon/1351213	0.059281	Grow-Diag Alignment/Trigram LM/Beam Width 200/Distortion Limit 10/Parameter Tuning
2	s10018	0.056561	Big Alignment Data/Witten Bell LM/Phrase Length 6/Parameter Tuning/Beam Width 110/Distortion Limit 10/Forward/Backward Phrase Probabilities
3	Colorful black ideas	0.055404	4-gram LM/Distortion Limit 12/Parameter Tuning
4	Lenessia-Erhart-Cowen	0.051152	Katz-Brown Preordering
5	tmrevolutionx/delacroix/abhi	0.050997	Grow-Diag Alignment/Translation Parameters
6	yemantu	0.0501	Distortion Limit 10
7	pongaskorn-u	0.047973	Distortion Limit 7/LM Parameters
8	Baseline	0.046933
9	Spivak	0	Trigram LM

Links

• Moses: A package for phrase-based machine translation
• Travatar: A package for syntax-based machine translation