1. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Summary of Example-based Machine Translation
Based on Syntactic Transfer with Statistical
Models
Hiroshi Matsumoto
Nagaoka University of Technology EEI Dept.
February 27, 2013
2. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Outline
1 About this paper
2 Introduction
3 Proposed method
4 Example of Tranfer Rules
5 Example of Syntactic Transfer
6 Statistical Generation
7 Bottom-up Generation
8 Evalution
9 Result
3. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
About this paper
About this paper:
Title: Example-based Machine Translation Based on Syntactic
Transfer with Statistical Models
Kenji Imamura, Hideo Okuma, Taro Watanaabe, and Eiichiro
Sumita
Book Title: Proceedings of the 20th international conference
on Computational Linguistics
Page 99
Year 2004
Organization Association for Computational Linguistics
4. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Introduction
Introduction
Automatic acquisition:
Example Based MT:
retrieves similar examples from parallel corpus to generate
translation
does not check correctness
Statistical MT:
translates by the combination of word transfer and word
re-ordering
high-quality if globally optimal solution is found
Proposed method
Uses two modules
1 EB syntactic transfer module: construction of TL's tree
structure
2 Statistical generation module: selection of best word sequence
Why this method
In addition to EBMT feature, this solves in which it lacks by
adding correct-ness in consideration of best word sequence
5. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Proposed method
Proposed method
Transfer Rules:
Revised ver. of Hierarchical Phrase Alignment-based Translator
by Imamura, 2002
Transfer rules are acquired from the HPA Translator
Rules are created from nodes correspondences
Syntactic Transfer Process
1 Input is parsed by transfer rule
2 Mapping the corresponding nodes
3 If non-terminals remain in the leaves, candidates are inserted
6. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Proposed method
Usage of Source Examples
Source Examples for disambiguation
Uses the semantic distance to calculate the ambiguitiy
Nearest example in semantic distance is used for tranfer rules
7. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Example of Tranfer Rules
Example of Tranfer Rules
8. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Example of Syntactic Transfer
Example of Syntactic Transfer
9. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Statistical Generation
The most appropriate
Use of LM & TM
The appropriateness is determined by product of Language
Model & Translation Model
For target F source E sequence, the following probability
E = argmaxE P (E |F ) = argmaxE P (E )P (F |E )
ˆ
For target f and source e words, the following probability
P (E |F ) = t (fj |ei )
j i
For LM, standard n-gram model
10. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Bottom-up Generation
Bottom-up Generation
11. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Evalution
Evaluation
Setting
Corpus Used: the Basic Travel Expression Corpus
Transfer Rules Acquired: 24,310
TM&LM: lexicon model of IBM Model 4 and word bigram and
trigram model
Compared Methods
Baseline: EB Tranfer only
Bottom-up: Best selection based on Bottom-up
All Search: Selection based on TM and LM probabilities
LM only: All Search without TM
Evaluation Metrics
BLEU, NIST, mWER, Subjective Evaluation
12. Summary of Example-based Machine Translation Based on Syntactic Transfer with Statistical Models
Result
Result