Post on 09-Feb-2016
description
Language & K nowledge Engineering Lab
A Probabilistic Framework for Structure-based
AlignmentKurohashi-lab M2
56430 Toshiaki Nakazawa
Language & Knowledge Engineering Lab
OutlineI. Introduction of Machine Translation
i. What is Alignment?ii. Statistical Machine Translation (SMT)iii. Example-based Machine Translation (EBMT)
II. Baseline alignment methodIII. A probabilistic framework for alignment
i. Corresponding Pattern score (CP-score)ii. Integration of Maximum Entropy (ME)
IV. Experiments and resultsV. Discussion and conclusion
Language & Knowledge Engineering Lab
OutlineI. Introduction of Machine Translation
i. What is Alignment?ii. Statistical Machine Translation (SMT)iii. Example-based Machine Translation (EBMT)
II. Baseline alignment methodIII. A probabilistic framework for alignment
i. Corresponding Pattern score (CP-score)ii. Integration of Maximum Entropy (ME)
IV. Experiments and resultsV. Discussion and conclusion
Language & Knowledge Engineering LabStandard Way of Machine Translation
ParallelCorpus Alignment Resource
Output
Translation
Input
Parallel Corpus: Text which is written in two different languages but the content is almost same.Alignment: To find the correspondence between two parallel sentences. (word level, phrase level, etc…)
The performance of alignment affects the accuracy of
translation.
Language & Knowledge Engineering LabStatistical Machine Translation (SMT)
Learn models for translation from parallel corpus statistically
Not use any linguistic resources Small translation unit (= “word”)
– Recently, the number of studies handling bigger unit (= “couple of words” or “phrase”) is increasing
Require large parallel corpus for highly-accurate translation
Language & Knowledge Engineering Lab
Basic Method for SMT Translate by maximizing the probability:
)|()(maxarg
)|(maxarg
EJPEP
JEPE
E
E
Language Model Translation Model
Learn from a parallel corpus(usually with unsupervised learning algorithm)
Ex) IBM Model [Brown et al., 93]
Language & Knowledge Engineering Lab
Overview of EBMT
ParallelCorpus Alignment TMDB
Output
Translation
Input
Advanced NLP technologies
Translation Memory Data
Base
Language & Knowledge Engineering LabExample-based Machine Translation (EBMT)
Divide the input sentence into a few parts Find a similar expressions (examples)
from parallel corpus for each parts Combine the examples to generate output
translation Use any linguistic resources as much as
possible Larger translation unit (larger example) is
better
Language & Knowledge Engineering Lab
Flow of EBMT
Language & Knowledge Engineering Lab
SMT vs. EBMT
SMT EBMT
GoodPoint
- Works enough for languages which don’t have sufficient NLP resources.
- Active to utilize any kinds of NLP resources.- High performance.
BadPoint
- Not easy to achieve high performance.- Weak for the wide difference between the languages.
- Algorithm is usually heuristic.- Modification is necessary for each language pair.
We introduce a probabilistic framework for structure-
based alignment.
Language & Knowledge Engineering Lab
OutlineI. Introduction of Machine Translation
i. What is Alignment?ii. Statistical Machine Translation (SMT)iii. Example-based Machine Translation (EBMT)
II. Baseline alignment methodIII. A probabilistic framework for alignment
i. Corresponding Pattern score (CP-score)ii. Integration of Maximum Entropy (ME)
IV. Experiments and resultsV. Discussion and conclusion
Language & Knowledge Engineering Lab
Alignment
交差点 で 、突然
あの車 が
飛び出して 来た のです
the carcame
at mefrom the side
at the intersection
1. Transformation into dependency structure
J: 交差点で、突然あの車が 飛び出して来たのです。
E : The car came at me from the side at the intersection.
J: JUMAN/KNPE: Charniak’s nlparser → Dependency tree
Language & Knowledge Engineering Lab
Alignment
交差点 で 、突然
あの車 が
飛び出して 来た のです
the carcame
at mefrom the side
at the intersection
1. Transformation into dependency structure2. Detection of word(s) correspondences
• Bilingual dictionaries• Transliteration detection
ローズワイン → rosuwain ⇔ rose wine (similarity:0.78)新宿 → shinjuku ⇔ shinjuku (similarity:1.0)
Language & Knowledge Engineering Lab
Alignment
交差点 で 、突然
あの車 が
飛び出して 来た のです
the carcame
at mefrom the side
at the intersection
1. Transformation into dependency structure2. Detection of word(s) correspondences3. Disambiguation of correspondences
Language & Knowledge Engineering Lab
Disambiguation
日本 で保険
会社 に対して
保険請求の
申し立て が可能です よ
you
will haveto file
insurance
an claim
insurance
with the office
in Japan
Cunamb → Camb : 1/(Distance in J tree) + 1/(Distance in E tree)
1/2 + 1/1
Language & Knowledge Engineering Lab
Alignment
交差点 で 、突然
あの車 が
飛び出して 来た のです
the carcame
at mefrom the side
at the intersection
1. Transformation into dependency structure2. Detection of word(s) correspondences3. Disambiguation of correspondences4. Handling of remaining phrases
Language & Knowledge Engineering Lab
Alignment
交差点 で 、突然
あの車 が
飛び出して 来た のです
the carcame
at mefrom the side
at the intersection
1. Transformation into dependency structure2. Detection of word(s) correspondences3. Disambiguation of correspondences4. Handling of remaining phrases5. Registration to translation example database
Language & Knowledge Engineering Lab
OutlineI. Introduction of Machine Translation
i. What is Alignment?ii. Statistical Machine Translation (SMT)iii. Example-based Machine Translation (EBMT)
II. Baseline alignment methodIII. A probabilistic framework for alignment
i. Corresponding Pattern score (CP-score)ii. Integration of Maximum Entropy (ME)
IV. Experiments and resultsV. Discussion and conclusion
Language & Knowledge Engineering Lab
Corresponding Pattern (CP)
Language & Knowledge Engineering Lab
Corresponding Pattern (CP)
Language & Knowledge Engineering Lab
Corresponding Pattern (CP)
(1, 2, 1, 1)
(0, 1, 0, 1)(0, 2, 0, 1) (0, 2) (0, 1) (0, 1) (0, 1)
(1, 2) (1, 1)
Language & Knowledge Engineering Lab
CP-score Assign a score to each CP = CP-score Calculation of CP-score
– Count the frequency of each CP Using the aligned parallel corpus by the baseline align
ment method
– Divide the frequency by the total frequency of all CPs (CP-score is a probability of occurrence)
Alignment Score (AS) by CP-score
1
1 1,
M
i
M
ijjiscoreCPAS
Language & Knowledge Engineering Lab
Alignment Disambiguation by AS
Adopt the alignment with highest AS
Language & Knowledge Engineering Lab
OutlineI. Introduction of Machine Translation
i. What is Alignment?ii. Statistical Machine Translation (SMT)iii. Example-based Machine Translation (EBMT)
II. Baseline alignment methodIII. A probabilistic framework for alignment
i. Corresponding Pattern score (CP-score)ii. Integration of Maximum Entropy (ME)
IV. Experiments and resultsV. Discussion and conclusion
Language & Knowledge Engineering Lab
Maximum Entropy (ME) The principle of maximum entropy:
– a method for analyzing the available information in order to determine a unique epistemic probability distribution. (by WIKIPEDIA)
Language & Knowledge Engineering Lab
Maximum Entropy (ME) The principle of maximum entropy:
– a method for analyzing the available information in order to determine a unique epistemic probability distribution. (by WIKIPEDIA)
Alignment probability with ME [Och et al,. 02]
),,( TSAmhS: Source sentenceT: Target sentenceA: Alignment
ATSA
TSATSA
]),,(exp[
]),,(exp[),|Pr(
1
1M
m mm
M
m mm
h
h
m
:Feature function
: Model parameter
Language & Knowledge Engineering Lab
Feature Functions
1. Alignment Score (AS)2. Parse score (Jap. and Eng.)3. Depth pattern score (DP-score)4. Probability of lexicon (Jap. and Eng.)5. Coverage of the correspondences (Jap.
and Eng.)6. Average size of the correspondences
(Jap. and Eng.)
Language & Knowledge Engineering Lab
OutlineI. Introduction of Machine Translation
i. What is Alignment?ii. Statistical Machine Translation (SMT)iii. Example-based Machine Translation (EBMT)
II. Baseline alignment methodIII. A probabilistic framework for alignment
i. Corresponding Pattern score (CP-score)ii. Integration of Maximum Entropy (ME)
IV. Experiments and resultsV. Discussion and conclusion
Language & Knowledge Engineering Lab
Experiments Select 500 moderately long sentences
from BTEC corpus of IWSLT2005 training data set
Manually annotate phrase-to-phrase alignment
Conducted 5-fold cross validation– 400 for training and 100 for testing
Calculated the F-measure
RPPRF
2 P: Precision
R: Recall
Language & Knowledge Engineering Lab
Results
Method
All sentences w/ function words
All sentences w/o function words
Ambiguous sentences w/ function words
Baseline 63.86 65.14 60.43
+CP-score 64.21 65.54 61.60
+ME 64.58 66.03 63.00
GIZA++ 22.14 52.85 23.78
Language & Knowledge Engineering Lab
Discussion Not considering clause
– Correspondences in the same clause of source sentence are likely to be in the same clause of target sentence
Sentence complexity– Proposed method works effectively for
long and complex sentences Preciseness of dictionary
– Erroneous correspondence by the dictionary makes bad effects on alignment
Language & Knowledge Engineering Lab
Conclusion Proposed a probabilistic framework to
improve structure-based alignment Proposed a new criteria CP-score for
evaluating alignment Integrate the ME model into alignment
approach
Language & Knowledge Engineering Lab
Future Work Sophisticate the CP and CP-score
– Consider clauses Select the feature functions Test our method on other corpora
– Longer and more complex sentences