presenter: jinhua du ( 杜金华 ) xi’an university of technology 西安理工大学
DESCRIPTION
NLP&CC, Chongqing, Nov. 17-19, 2013. Discriminative Latent Variable Based Classifier for Translation Error Detection. Presenter: Jinhua Du ( 杜金华 ) Xi’an University of Technology 西安理工大学. Outline. Introduction. 2. DPLVM for Translation Error Detection. 3. Experiments and Analysis. - PowerPoint PPT PresentationTRANSCRIPT
Presenter: Jinhua Du (杜金华 )
Xi’an University of Technology西安理工大学
NLP&CC, Chongqing, Nov. 17-19, 2013
Discriminative Latent Variable Based Classifier for Translation Error
Detection
Outline
1. Introduction
2. DPLVM for Translation Error Detection
3. Experiments and Analysis
4. Conclusions and Future Work
1. IntroductionProblem
1. In localization industry, human is always involved in post-editing the MT results;
2. MT errors always increase human cost to obtain a reasonable translation;
3. Translation error detection or word confidence estimation can improve working efficiency of post-editors in some extent.
1. In localization industry, human is always involved in post-editing the MT results;
2. MT errors always increase human cost to obtain a reasonable translation;
3. Translation error detection or word confidence estimation can improve working efficiency of post-editors in some extent.
•Research Question: how to improve the detection accuracy of detecting translation errors?
Blatz et al. combined the neuralnetwork and a naive Bayes classifier
2004
Ueffing and Ney exhaustively explored various kinds of WPP features
2003/2007
Specia et al. worked on confidence estimation in CAT field2009/20
11
Xiong et al. used a MaxEnt-based classifier to predict translation errors
2010
1. IntroductionRelated Work
For same feature set,
different classifiers show
different performance,
thus how to select/design a
proper classifier is
important
Classifiers
For a classifier, different
features reflect different
characteristics of problem,
how to select/design a
feature set is crucial
Features
1. IntroductionKey Factors
Title in hereFeature set
Title in here
Comparison withSVM andMaxEnt
Title in hereDiscriminative
Latent Variableclassifier
1. IntroductionOur Work
2. DPLVM Algorithm
Conditions: a sequence of observations x = {x1, x2,…, xm}
a sequence of labels y = {y1, y2,…, ym}
Assumption: a sequence of latent variables h = {h1, h2,…, hm}
Goal: to learn a mapping between x and y
Definition:( | , ) ( | , , ) ( | , )P P P
h
y x y h x h x (1)
Simplified Algorithm
Assumptions: the model is restricted to have disjoint sets of latent
variables associated with each class label; Each hhjj is a member in a set HHyyjj of possible latent variables
for the class label yyj j ; so sequences which have any will by definition
have
Equation (1) can be re-written as:
where
jj yh H
( | , ) 0P y x
1...
( | , ) ( | , )y ym
P P
h H H
y x h x (2)
exp ( , )( | , )
exp ( , )h
P
f h x
h xf h x (3)
Parameter Estimation
Decoding for test set:
Decoding algorithm: Sun and Tsujii (2009): a latent-dynamic inference (LDI)
method based on A* search and dynamic programming;
DPLVM in Translation Error Detection Task
Prerequisites: Types of errors can be classified; Each class has a specific label; The classification task can be regarded as a labelling task;
2 Classes of word label C: correct Good words label: c I: incorrect Bad words label: i
Feature Set
Word Posterior
Probabilities
• Fixed position based WPP• Flexible position based WPP• Word alignment based WPP
Lexical
Features
• Part of speech (POS)
• word entity
Syntactic
Features
• word links from LG parser
Feature Representation
3. Experiments and Analysis
Experimental Settings – SMT system
•Language pair: Chinese-English•Training set: NIST data set,3.4m•Devset: NIST MT 2006 current set•Testset: NIST MT 2005,2008 sets
SMT Performance
Experimental Settings for Error Detection Task
• Devset: translations of NIST MT-08
• Testset: translations of NIST MT-05• Annotation: TER to determine the true labels for words, 37.99% ratio of correct words for MT-08, 41.59% RCW for MT-05
Data Set and Data Annotation
Evaluation MetricsEvaluation Metrics
(2) Classification Experiment on Combined Features
Observations
• The name entities are prone to be wrongly classified
• The prepositions, conjunctions, auxiliary verbs and articles are easier to be wrongly classified
• The proportion of the notional words that are wrongly classified is relatively small
4. Conclusions and Future Work
Conclusions
Presents a new classifier - DPLVM-based classifier -for translation error detection
Introduces three different kinds of WPP features,three linguistic features
Compares the MaxEnt classifier, SVM classifier and our DPLVM classifier
The proposed classifier performs best compared to two other individual classifiers in terms of CER
introducing paraphrases to annotate the hypothesesintroducing paraphrases to annotate the hypotheses
introducing new useful features to further improve the detection capability
performing experiments on more language pairs to verify our proposed method.
4. Conclusions and Future Work
Future Work
Thanks for your attention!