SlideShare a Scribd company logo
1 of 30
An Empirical Study of Segment Prioritization
for Incrementally Retrained Post-Editing-
Based SMT
Jinhua Du, Ankit Srivastava, Andy Way
ADAPT Centre,
Dublin City University, Ireland
Alfredo Maldonado-Guerra, David Lewis
ADAPT Centre,
Trinity College Dublin, Ireland
The ADAPT Centre is funded under the SFI Research Centres Programme (Grant 13/RC/2106) and is co-funded under the European Regional Development Fund.
MT Summit XV, 2015-11-02, Miami, FL
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Outline
Motivation
Incrementally Retrained PE-SMT
Experiments and Analysis
Conclusions and Future Work
Segment Prioritization Algorithms
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Motivation - Question
 Post-editing-based SMT has been widely deployed into the translator’s
workflow.
 The productivity of translators/post-editors can be increased due to the
the use of SMT systems.
 However, current SMT systems cannot generate high-quality translations,
so human effort is usually required.
 Our Question: How to significantly reduce the PE-time?
Two objective ways:
 Inbound: improving SMT system performance
 Outbound: post-editing settings, circumstances, etc.
Segment Prioritization plays an important role in reducing the
overall PE-time.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Motivation - Related work
Incrementally PE-SMT is essentially an active learning-based SMT,
including two key parts:
 Active learning framework
 Incremental retraining methods for model update
In 2009, Haffari et al. firstly proposed a practical active learning
framework for low-resource SMT where a number of high-quality
parallel data are acquired from the large-scale monolingual data.
In 2012, Gonzalez-Rubio et al. apply AL to the interactive MT to
select informative sentences to reduce human effort rather than PE
time.
The most relevant previous work: Dara et al. (2014) propose a Cross
Entropy Difference (CED) criterion to prioritize input segments in an
AL framework for PE-based incremental MT update applications.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Motivation - Related work
Ortiz-Martinez et al. (2010) incrementally update the feature values
of the phrase table based on the pre-stored statistics related to the
feature scores.
Hardt and Elming (2010) propose a sentence-level retraining
scheme in which a local phrase table is created and incrementally
updated as a file is translated and post-edited.
Denkowski et al. (2014) propose an online model adaptation for
PE-SMT in which three methods are used for incremental model
adaptation.
Germann (2014) proposes a dynamic phrase table strategy for an
interactive PE-SMT that computes phrase table entries on demand
by sampling a suffix array-indexed bitext.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Motivation
System Framework: batch-mode incremental retraining;
Goal: reducing the overall PE-time;
Our solution: source-side segments prioritization;
Contributions: an empirical study on different segment
prioritization algorithms.
Motivation
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Outline
Motivation
Incrementally Retrained PE-SMT
Experiments and Analysis
Conclusions and Future Work
Segment Prioritization Algorithms
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Incrementally Retrained PE-SMT
Workflow of active learning-based incrementally retrained PE-SMT
Where we are!
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Incrementally Retrained PE-SMT
Workflow of the batch-mode incrementally retraining
Where we are!
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Incrementally Retrained PE-SMT
Significant Features:
 Architecture:
 A combination of Moses and CDEC
 A combination of online and offline retraining
 Speed: about 20 minutes for each iteration to update
models at the scale of 4 million sentence pairs. However,
the time for Moses incremental retraining scheme is
more than 2 hours for the same-scale data.
 Performance: no significantly difference between the
incrementally retrained system and the offline retrained
system
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Outline
Motivation
Incrementally Retrained PE-SMT
Experiments and Analysis
Conclusions and Future Work
Segment Prioritization Algorithms
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
The initial parallel training corpus
a monolingual corpus (translation job)
F
Scoring metric: segment prioritization
algorithm
Segment Prioritization Algorithms
To prioritize the input segments, an importance score or uncertainty score for
the sentence s must be calculated under a scoring metric as:
Mathematical Formalization
),,(F:)( LUss 
)},{(:L ii ef
}{: jfU 
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Segment Prioritization Algorithms
Confidence of Translations (Confidence)
Definition: The probability p(e|f) of a translation output ê that is
calculated by the decoder.
 Geometry n-gram (Geom n-gram)
The Geom n-gram algorithm is defined as:
),|(
),|(
log
||
:)(
1 nLxP
nUxP
X
s n
sXx
N
n
n
s
n


 


|)|(log|:)( feps 
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Segment Prioritization Algorithms
Perplexity of Sentences (PPL)
The PPL algorithm is defined as:
 Cross Entropy Difference (CED)
The CED algorithm is defined as:
)()(:)( sHsHs LU 
N-OOVs
p(s)
s
log
10:)(


Ranking: The higher the score, the higher the rank.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Outline
Motivation
Incrementally Retrained PE-SMT
Experiments and Analysis
Conclusions and Future Work
Segment Prioritization Algorithms
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
•Corpora: Europarl
•Language pairs: EN-DE, EN-ES,
ES-EN, DE-EN
•Training set: 50K
•Incremental set: 10K
•Devset: Newswire 2012 test set
Settings for incrementally
retrained SMT system
•Corpora: DGT
•Language pair: EN-DE, EN-ES,
EN-FR, EN-PL
•Training set: 50K
•Incremental set: 10K
•Devset: 2K (extracted)
 Evaluation metric: TER to simulate the PE-time;
 Batch mode: 500 sentences for each batch;
 Simulated workflow: we use the references of the translated segments
instead of post-edited MT output per se;
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
Experimental results (DGT) -- TER Score
Pair Sequential
Geom
n-gram
PPL CED Confidence Gains
EN-DE 55.94 56.50 56.58 55.23 51.39 4.55
EN-ES 41.65 42.40 42.48 41.58 38.69 2.96
EN-FR 44.86 46.92 46.75 44.62 41.42 3.44
EN-PL 51.38 51.53 51.63 51.16 48.09 3.29
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
Experimental results (Europarl) -- TER Score
Pair Sequential
Geom
n-gram
PPL CED Confidence Gains
DE-EN 73.73 73.87 72.60 72.48 70.56 3.17
EN-DE 80.12 80.00 79.38 79.02 77.83 2.29
ES-EN 84.14 84.20 83.75 83.25 81.82 2.32
EN-ES 64.10 64.10 63.37 63.11 62.38 1.72
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
Iterative curves – DGT – TER Score
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
Iterative curves – Europarl – TER Score
 we can see that there is no obvious decrease (i.e. improvement) for the
baseline;
 the other four prioritization criteria have a trend of decreasing the TER
score;
 The decreasing trends show the effectiveness of these methods to
prioritize the input segments.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
 the sentence length distribution of the “Confidence” criterion strongly fluctuates
per iteration;
 the fluctuations of the length distribution of “Confidence” consistently
correspond to the TER score trend;
 The length distributions of the “Geom n-gram”, “PPL” and “CED” methods are
more smooth, but start from shorter sentences, end with longer sentences.
Sentence Length Distribution
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
 It is inferred that the prioritization score ϕ(s) of “Confidence” may correlate to
the sentence length of the input segment.
 Pearson correlation: EN-DE as an example
Correlation between sentence length and sentence score
EN-DE Geom n-gram PPL CED Confidence
DGT 0.21 0.009 0.02 0.29
Europarl 0.14 0.06 0.10 0.48
 Results show that the longer the sentence is, the more difficult it is to be
translated;
 Question: should we normalize the score by the sentence length or not in the
segment prioritization task?
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
Correlation: Normalize or not?
 The average TER score for the normalized “Confidence” is 56.09 which is much
higher (i.e. worse) than the baseline (55.94).
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
OOVs: Correlation with the sentence score
 In Moses, when an OOV occurs, the probability p(e|f) will be
significantly decreased, i.e. the confidence of the translation becomes
lower.
 Pearson correlation: EN-DE as an example
Pearson Correlation DGT Europarl
ρ 0.9993 0.9997
 As expected the more OOVs a sentence contains, the lower the
confidence score, and the greater the possibility that it is ranked at the top.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Experiments and Analysis
 Confidence: Pros and Cons
 Confidence performed best in all methods. HOWEVER,
 At each iteration, all the incremental segments need to be translated
beforehand, which might be a problem for sentence-level incremental
PE-SMT.
 The sentence length at the first several iterations is much longer,
which might significantly increase the amount of post-editing, which in
turn may adversely affect the perception of translators/post-editors as to
the utility of this approach.
Implication from the analysis: develop a method that can consider the
OOVs and sentence length distribution in the training data and the
incremental data rather than translating them.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Conclusions and Future Work
Conclusions
2
1
3
The “Confidence” method achieved the best results in all tasks that
reduced the TER score of 1.72-4.55 absolute points.
We conducted an empirical study on four different segment
prioritization algorithms, namely the Sequential, Geom n-gram, PPL,
CED and Confidence methods for incrementally retrained PE-SMT..
An investigation was carried out to examine the crucial factors that
make the “Confidence” effective.
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Updates: Continuous Work
Actual Post-editing Experiment
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Updates: Continuous Work
Actual Post-editing experiment
This work will be beneficial for CAT tools
(Web-based or Desktop-based)!
www.adaptcentre.ie
E-Mail: jdu@computing.dcu.ie
Conclusions and Future Work
 Future Work
The context problem: the sorted input segments lose the sequential
context that is helpful to the translators.Q
Considering the context when ranking the segments.
Carrying out actual PE experiments using our different segment
prioritization algorithms.
1
2
The ADAPT Centre is funded under the SFI Research Centres Programme (Grant 13/RC/2106) and is co-funded under the European Regional Development Fund.
This research is supported by Science Foundation
Ireland through the ADAPT Centre (Grant
13/RC/2106) (www.adaptcentre.ie) at Dublin City
University and Trinity College Dublin, and by Grant
610879 for the Falcon project funded by the
European Commission.

More Related Content

What's hot

BERT - Part 1 Learning Notes of Senthil Kumar
BERT - Part 1 Learning Notes of Senthil KumarBERT - Part 1 Learning Notes of Senthil Kumar
BERT - Part 1 Learning Notes of Senthil KumarSenthil Kumar M
 
Multi-Robot Task Allocation using Meta-heuristic Optimization
Multi-Robot Task Allocation using Meta-heuristic OptimizationMulti-Robot Task Allocation using Meta-heuristic Optimization
Multi-Robot Task Allocation using Meta-heuristic OptimizationMohamed Ghanem
 
Phonetic distance based accent
Phonetic distance based accentPhonetic distance based accent
Phonetic distance based accentsipij
 
Isolated word recognition using lpc & vector quantization
Isolated word recognition using lpc & vector quantizationIsolated word recognition using lpc & vector quantization
Isolated word recognition using lpc & vector quantizationeSAT Publishing House
 
An introduction to the Transformers architecture and BERT
An introduction to the Transformers architecture and BERTAn introduction to the Transformers architecture and BERT
An introduction to the Transformers architecture and BERTSuman Debnath
 
Understanding Natural Languange with Corpora-based Generation of Dependency G...
Understanding Natural Languange with Corpora-based Generation of Dependency G...Understanding Natural Languange with Corpora-based Generation of Dependency G...
Understanding Natural Languange with Corpora-based Generation of Dependency G...Edmond Lepedus
 
ENSEMBLE MODEL FOR CHUNKING
ENSEMBLE MODEL FOR CHUNKINGENSEMBLE MODEL FOR CHUNKING
ENSEMBLE MODEL FOR CHUNKINGijasuc
 
Neural machine translation of rare words with subword units
Neural machine translation of rare words with subword unitsNeural machine translation of rare words with subword units
Neural machine translation of rare words with subword unitsTae Hwan Jung
 
BERT: Bidirectional Encoder Representations from Transformers
BERT: Bidirectional Encoder Representations from TransformersBERT: Bidirectional Encoder Representations from Transformers
BERT: Bidirectional Encoder Representations from TransformersLiangqun Lu
 
Improving Neural Abstractive Text Summarization with Prior Knowledge
Improving Neural Abstractive Text Summarization with Prior KnowledgeImproving Neural Abstractive Text Summarization with Prior Knowledge
Improving Neural Abstractive Text Summarization with Prior KnowledgeGaetano Rossiello, PhD
 
PUNJABI SPEECH SYNTHESIS SYSTEM USING HTK
PUNJABI SPEECH SYNTHESIS SYSTEM USING HTKPUNJABI SPEECH SYNTHESIS SYSTEM USING HTK
PUNJABI SPEECH SYNTHESIS SYSTEM USING HTKijistjournal
 
Designing of an efficient algorithm for identifying Abbreviation definitions ...
Designing of an efficient algorithm for identifying Abbreviation definitions ...Designing of an efficient algorithm for identifying Abbreviation definitions ...
Designing of an efficient algorithm for identifying Abbreviation definitions ...ijcsit
 
T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...
T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...
T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...csandit
 
Bca2020 data structure and algorithm
Bca2020  data structure and algorithmBca2020  data structure and algorithm
Bca2020 data structure and algorithmsmumbahelp
 

What's hot (15)

BERT - Part 1 Learning Notes of Senthil Kumar
BERT - Part 1 Learning Notes of Senthil KumarBERT - Part 1 Learning Notes of Senthil Kumar
BERT - Part 1 Learning Notes of Senthil Kumar
 
Multi-Robot Task Allocation using Meta-heuristic Optimization
Multi-Robot Task Allocation using Meta-heuristic OptimizationMulti-Robot Task Allocation using Meta-heuristic Optimization
Multi-Robot Task Allocation using Meta-heuristic Optimization
 
Phonetic distance based accent
Phonetic distance based accentPhonetic distance based accent
Phonetic distance based accent
 
Isolated word recognition using lpc & vector quantization
Isolated word recognition using lpc & vector quantizationIsolated word recognition using lpc & vector quantization
Isolated word recognition using lpc & vector quantization
 
An introduction to the Transformers architecture and BERT
An introduction to the Transformers architecture and BERTAn introduction to the Transformers architecture and BERT
An introduction to the Transformers architecture and BERT
 
Understanding Natural Languange with Corpora-based Generation of Dependency G...
Understanding Natural Languange with Corpora-based Generation of Dependency G...Understanding Natural Languange with Corpora-based Generation of Dependency G...
Understanding Natural Languange with Corpora-based Generation of Dependency G...
 
ENSEMBLE MODEL FOR CHUNKING
ENSEMBLE MODEL FOR CHUNKINGENSEMBLE MODEL FOR CHUNKING
ENSEMBLE MODEL FOR CHUNKING
 
[Paper review] BERT
[Paper review] BERT[Paper review] BERT
[Paper review] BERT
 
Neural machine translation of rare words with subword units
Neural machine translation of rare words with subword unitsNeural machine translation of rare words with subword units
Neural machine translation of rare words with subword units
 
BERT: Bidirectional Encoder Representations from Transformers
BERT: Bidirectional Encoder Representations from TransformersBERT: Bidirectional Encoder Representations from Transformers
BERT: Bidirectional Encoder Representations from Transformers
 
Improving Neural Abstractive Text Summarization with Prior Knowledge
Improving Neural Abstractive Text Summarization with Prior KnowledgeImproving Neural Abstractive Text Summarization with Prior Knowledge
Improving Neural Abstractive Text Summarization with Prior Knowledge
 
PUNJABI SPEECH SYNTHESIS SYSTEM USING HTK
PUNJABI SPEECH SYNTHESIS SYSTEM USING HTKPUNJABI SPEECH SYNTHESIS SYSTEM USING HTK
PUNJABI SPEECH SYNTHESIS SYSTEM USING HTK
 
Designing of an efficient algorithm for identifying Abbreviation definitions ...
Designing of an efficient algorithm for identifying Abbreviation definitions ...Designing of an efficient algorithm for identifying Abbreviation definitions ...
Designing of an efficient algorithm for identifying Abbreviation definitions ...
 
T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...
T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...
T EXT M INING AND C LASSIFICATION OF P RODUCT R EVIEWS U SING S TRUCTURED S U...
 
Bca2020 data structure and algorithm
Bca2020  data structure and algorithmBca2020  data structure and algorithm
Bca2020 data structure and algorithm
 

Viewers also liked

Overdragelsesdokument (slutseddel) MagHus og DFE bilag 1
Overdragelsesdokument (slutseddel) MagHus  og DFE bilag 1Overdragelsesdokument (slutseddel) MagHus  og DFE bilag 1
Overdragelsesdokument (slutseddel) MagHus og DFE bilag 1Lis Walker
 
Journalism Headline Assigment
Journalism Headline AssigmentJournalism Headline Assigment
Journalism Headline AssigmentWarinya Kuanpinit
 
Navis Yachts #46 Perseus^3
Navis Yachts #46 Perseus^3Navis Yachts #46 Perseus^3
Navis Yachts #46 Perseus^3turnerkat
 
Nordplus project "Forest- for rest?" meeting in Latvia
Nordplus project "Forest- for rest?" meeting in LatviaNordplus project "Forest- for rest?" meeting in Latvia
Nordplus project "Forest- for rest?" meeting in LatviaIlze Stipra
 
україна на початку хх століття
україна на початку хх століттяукраїна на початку хх століття
україна на початку хх століття160289yfnfif
 
физика
физикафизика
физикаsokolovq
 
20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께
20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께
20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께Myeongnyun Mission Church
 
Презентация, ФПО МГУ, Домарева Евгения
Презентация, ФПО МГУ, Домарева ЕвгенияПрезентация, ФПО МГУ, Домарева Евгения
Презентация, ФПО МГУ, Домарева ЕвгенияDomareva
 
Nijat waladi al nabi minon naar
Nijat waladi al nabi minon naarNijat waladi al nabi minon naar
Nijat waladi al nabi minon naarMuhammad Tariq
 
Elite traffic and conversions
Elite traffic and conversionsElite traffic and conversions
Elite traffic and conversionsRichard Seaton
 
Python Training v2
Python Training v2Python Training v2
Python Training v2ibaydan
 
Теория автоматов и вычислений 1.3 знакомство, учебники, ссылки
Теория автоматов и вычислений 1.3 знакомство, учебники, ссылкиТеория автоматов и вычислений 1.3 знакомство, учебники, ссылки
Теория автоматов и вычислений 1.3 знакомство, учебники, ссылкиIgor Kleiner
 

Viewers also liked (20)

Agile Brazil 2015
Agile Brazil 2015Agile Brazil 2015
Agile Brazil 2015
 
Recycleoneline
RecycleonelineRecycleoneline
Recycleoneline
 
Overdragelsesdokument (slutseddel) MagHus og DFE bilag 1
Overdragelsesdokument (slutseddel) MagHus  og DFE bilag 1Overdragelsesdokument (slutseddel) MagHus  og DFE bilag 1
Overdragelsesdokument (slutseddel) MagHus og DFE bilag 1
 
Journalism Headline Assigment
Journalism Headline AssigmentJournalism Headline Assigment
Journalism Headline Assigment
 
Navis Yachts #46 Perseus^3
Navis Yachts #46 Perseus^3Navis Yachts #46 Perseus^3
Navis Yachts #46 Perseus^3
 
U1 test-b
U1 test-bU1 test-b
U1 test-b
 
Nordplus project "Forest- for rest?" meeting in Latvia
Nordplus project "Forest- for rest?" meeting in LatviaNordplus project "Forest- for rest?" meeting in Latvia
Nordplus project "Forest- for rest?" meeting in Latvia
 
україна на початку хх століття
україна на початку хх століттяукраїна на початку хх століття
україна на початку хх століття
 
физика
физикафизика
физика
 
20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께
20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께
20151101 주일예배, 막12장13절-17절, 하나님의 것은 하나님께
 
Photos chantier
Photos chantierPhotos chantier
Photos chantier
 
Презентация, ФПО МГУ, Домарева Евгения
Презентация, ФПО МГУ, Домарева ЕвгенияПрезентация, ФПО МГУ, Домарева Евгения
Презентация, ФПО МГУ, Домарева Евгения
 
Warmi silvia
Warmi silviaWarmi silvia
Warmi silvia
 
Nijat waladi al nabi minon naar
Nijat waladi al nabi minon naarNijat waladi al nabi minon naar
Nijat waladi al nabi minon naar
 
MODUL 3-PG
MODUL 3-PGMODUL 3-PG
MODUL 3-PG
 
Elite traffic and conversions
Elite traffic and conversionsElite traffic and conversions
Elite traffic and conversions
 
Group 8 qatar final
Group 8 qatar finalGroup 8 qatar final
Group 8 qatar final
 
Engnr Fazal CV
Engnr Fazal CVEngnr Fazal CV
Engnr Fazal CV
 
Python Training v2
Python Training v2Python Training v2
Python Training v2
 
Теория автоматов и вычислений 1.3 знакомство, учебники, ссылки
Теория автоматов и вычислений 1.3 знакомство, учебники, ссылкиТеория автоматов и вычислений 1.3 знакомство, учебники, ссылки
Теория автоматов и вычислений 1.3 знакомство, учебники, ссылки
 

Similar to Mt summit2015 jdu_v2

Cs 568 Spring 10 Lecture 5 Estimation
Cs 568 Spring 10  Lecture 5 EstimationCs 568 Spring 10  Lecture 5 Estimation
Cs 568 Spring 10 Lecture 5 EstimationLawrence Bernstein
 
vorl1.ppt
vorl1.pptvorl1.ppt
vorl1.pptbutest
 
HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...
HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...
HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...Lifeng (Aaron) Han
 
Cross-lingual Information Retrieval
Cross-lingual Information RetrievalCross-lingual Information Retrieval
Cross-lingual Information RetrievalShadi Saleh
 
IRJET - Automated Essay Grading System using Deep Learning
IRJET -  	  Automated Essay Grading System using Deep LearningIRJET -  	  Automated Essay Grading System using Deep Learning
IRJET - Automated Essay Grading System using Deep LearningIRJET Journal
 
Creating a dataset of peer review in computer science conferences published b...
Creating a dataset of peer review in computer science conferences published b...Creating a dataset of peer review in computer science conferences published b...
Creating a dataset of peer review in computer science conferences published b...Aliaksandr Birukou
 
IRJET- Deep Learning Model to Predict Hardware Performance
IRJET- Deep Learning Model to Predict Hardware PerformanceIRJET- Deep Learning Model to Predict Hardware Performance
IRJET- Deep Learning Model to Predict Hardware PerformanceIRJET Journal
 
IRJET- Analysis of PV Fed Vector Controlled Induction Motor Drive
IRJET- Analysis of PV Fed Vector Controlled Induction Motor DriveIRJET- Analysis of PV Fed Vector Controlled Induction Motor Drive
IRJET- Analysis of PV Fed Vector Controlled Induction Motor DriveIRJET Journal
 
IRJET - Response Analysis of Educational Videos
IRJET - Response Analysis of Educational VideosIRJET - Response Analysis of Educational Videos
IRJET - Response Analysis of Educational VideosIRJET Journal
 
Machine Learning Model Evaluation Methods
Machine Learning Model Evaluation MethodsMachine Learning Model Evaluation Methods
Machine Learning Model Evaluation MethodsPyingkodi Maran
 
On the application of SAT solvers for Search Based Software Testing
On the application of SAT solvers for Search Based Software TestingOn the application of SAT solvers for Search Based Software Testing
On the application of SAT solvers for Search Based Software Testingjfrchicanog
 
Multi-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learningMulti-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learningSanghamitra Deb
 
Part 1
Part 1Part 1
Part 1butest
 
IA3_presentation.pptx
IA3_presentation.pptxIA3_presentation.pptx
IA3_presentation.pptxKtonNguyn2
 
ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015
ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015
ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015RIILP
 
Best Data Science Online Training in Hyderabad
  Best Data Science Online Training in Hyderabad  Best Data Science Online Training in Hyderabad
Best Data Science Online Training in Hyderabadbharathtsofttech
 

Similar to Mt summit2015 jdu_v2 (20)

Cs 568 Spring 10 Lecture 5 Estimation
Cs 568 Spring 10  Lecture 5 EstimationCs 568 Spring 10  Lecture 5 Estimation
Cs 568 Spring 10 Lecture 5 Estimation
 
vorl1.ppt
vorl1.pptvorl1.ppt
vorl1.ppt
 
HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...
HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...
HOPE: A Task-Oriented and Human-Centric Evaluation Framework Using Profession...
 
Cross-lingual Information Retrieval
Cross-lingual Information RetrievalCross-lingual Information Retrieval
Cross-lingual Information Retrieval
 
IRJET - Automated Essay Grading System using Deep Learning
IRJET -  	  Automated Essay Grading System using Deep LearningIRJET -  	  Automated Essay Grading System using Deep Learning
IRJET - Automated Essay Grading System using Deep Learning
 
PerOpteryx
PerOpteryxPerOpteryx
PerOpteryx
 
Q P
Q PQ P
Q P
 
Creating a dataset of peer review in computer science conferences published b...
Creating a dataset of peer review in computer science conferences published b...Creating a dataset of peer review in computer science conferences published b...
Creating a dataset of peer review in computer science conferences published b...
 
IRJET- Deep Learning Model to Predict Hardware Performance
IRJET- Deep Learning Model to Predict Hardware PerformanceIRJET- Deep Learning Model to Predict Hardware Performance
IRJET- Deep Learning Model to Predict Hardware Performance
 
IRJET- Analysis of PV Fed Vector Controlled Induction Motor Drive
IRJET- Analysis of PV Fed Vector Controlled Induction Motor DriveIRJET- Analysis of PV Fed Vector Controlled Induction Motor Drive
IRJET- Analysis of PV Fed Vector Controlled Induction Motor Drive
 
IRJET - Response Analysis of Educational Videos
IRJET - Response Analysis of Educational VideosIRJET - Response Analysis of Educational Videos
IRJET - Response Analysis of Educational Videos
 
Machine Learning Model Evaluation Methods
Machine Learning Model Evaluation MethodsMachine Learning Model Evaluation Methods
Machine Learning Model Evaluation Methods
 
On the application of SAT solvers for Search Based Software Testing
On the application of SAT solvers for Search Based Software TestingOn the application of SAT solvers for Search Based Software Testing
On the application of SAT solvers for Search Based Software Testing
 
Machine learning
Machine learningMachine learning
Machine learning
 
Se exe 6
Se exe 6Se exe 6
Se exe 6
 
Multi-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learningMulti-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learning
 
Part 1
Part 1Part 1
Part 1
 
IA3_presentation.pptx
IA3_presentation.pptxIA3_presentation.pptx
IA3_presentation.pptx
 
ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015
ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015
ESR2 Santanu Pal - EXPERT Summer School - Malaga 2015
 
Best Data Science Online Training in Hyderabad
  Best Data Science Online Training in Hyderabad  Best Data Science Online Training in Hyderabad
Best Data Science Online Training in Hyderabad
 

Recently uploaded

Vector Search -An Introduction in Oracle Database 23ai.pptx
Vector Search -An Introduction in Oracle Database 23ai.pptxVector Search -An Introduction in Oracle Database 23ai.pptx
Vector Search -An Introduction in Oracle Database 23ai.pptxRemote DBA Services
 
Architecting Cloud Native Applications
Architecting Cloud Native ApplicationsArchitecting Cloud Native Applications
Architecting Cloud Native ApplicationsWSO2
 
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot TakeoffStrategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoffsammart93
 
Why Teams call analytics are critical to your entire business
Why Teams call analytics are critical to your entire businessWhy Teams call analytics are critical to your entire business
Why Teams call analytics are critical to your entire businesspanagenda
 
Simplifying Mobile A11y Presentation.pptx
Simplifying Mobile A11y Presentation.pptxSimplifying Mobile A11y Presentation.pptx
Simplifying Mobile A11y Presentation.pptxMarkSteadman7
 
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....rightmanforbloodline
 
Elevate Developer Efficiency & build GenAI Application with Amazon Q​
Elevate Developer Efficiency & build GenAI Application with Amazon Q​Elevate Developer Efficiency & build GenAI Application with Amazon Q​
Elevate Developer Efficiency & build GenAI Application with Amazon Q​Bhuvaneswari Subramani
 
Six Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal OntologySix Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal Ontologyjohnbeverley2021
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Orbitshub
 
MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MIND CTI
 
JohnPollard-hybrid-app-RailsConf2024.pptx
JohnPollard-hybrid-app-RailsConf2024.pptxJohnPollard-hybrid-app-RailsConf2024.pptx
JohnPollard-hybrid-app-RailsConf2024.pptxJohnPollard37
 
Corporate and higher education May webinar.pptx
Corporate and higher education May webinar.pptxCorporate and higher education May webinar.pptx
Corporate and higher education May webinar.pptxRustici Software
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingEdi Saputra
 
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc
 
Decarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational PerformanceDecarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational PerformanceIES VE
 
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfOrbitshub
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDropbox
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusZilliz
 
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Jeffrey Haguewood
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FMESafe Software
 

Recently uploaded (20)

Vector Search -An Introduction in Oracle Database 23ai.pptx
Vector Search -An Introduction in Oracle Database 23ai.pptxVector Search -An Introduction in Oracle Database 23ai.pptx
Vector Search -An Introduction in Oracle Database 23ai.pptx
 
Architecting Cloud Native Applications
Architecting Cloud Native ApplicationsArchitecting Cloud Native Applications
Architecting Cloud Native Applications
 
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot TakeoffStrategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
 
Why Teams call analytics are critical to your entire business
Why Teams call analytics are critical to your entire businessWhy Teams call analytics are critical to your entire business
Why Teams call analytics are critical to your entire business
 
Simplifying Mobile A11y Presentation.pptx
Simplifying Mobile A11y Presentation.pptxSimplifying Mobile A11y Presentation.pptx
Simplifying Mobile A11y Presentation.pptx
 
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
TEST BANK For Principles of Anatomy and Physiology, 16th Edition by Gerard J....
 
Elevate Developer Efficiency & build GenAI Application with Amazon Q​
Elevate Developer Efficiency & build GenAI Application with Amazon Q​Elevate Developer Efficiency & build GenAI Application with Amazon Q​
Elevate Developer Efficiency & build GenAI Application with Amazon Q​
 
Six Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal OntologySix Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal Ontology
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
 
MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024
 
JohnPollard-hybrid-app-RailsConf2024.pptx
JohnPollard-hybrid-app-RailsConf2024.pptxJohnPollard-hybrid-app-RailsConf2024.pptx
JohnPollard-hybrid-app-RailsConf2024.pptx
 
Corporate and higher education May webinar.pptx
Corporate and higher education May webinar.pptxCorporate and higher education May webinar.pptx
Corporate and higher education May webinar.pptx
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
 
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
 
Decarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational PerformanceDecarbonising Commercial Real Estate: The Role of Operational Performance
Decarbonising Commercial Real Estate: The Role of Operational Performance
 
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor Presentation
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with Milvus
 
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
 

Mt summit2015 jdu_v2

  • 1. An Empirical Study of Segment Prioritization for Incrementally Retrained Post-Editing- Based SMT Jinhua Du, Ankit Srivastava, Andy Way ADAPT Centre, Dublin City University, Ireland Alfredo Maldonado-Guerra, David Lewis ADAPT Centre, Trinity College Dublin, Ireland The ADAPT Centre is funded under the SFI Research Centres Programme (Grant 13/RC/2106) and is co-funded under the European Regional Development Fund. MT Summit XV, 2015-11-02, Miami, FL
  • 2. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Outline Motivation Incrementally Retrained PE-SMT Experiments and Analysis Conclusions and Future Work Segment Prioritization Algorithms
  • 3. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Motivation - Question  Post-editing-based SMT has been widely deployed into the translator’s workflow.  The productivity of translators/post-editors can be increased due to the the use of SMT systems.  However, current SMT systems cannot generate high-quality translations, so human effort is usually required.  Our Question: How to significantly reduce the PE-time? Two objective ways:  Inbound: improving SMT system performance  Outbound: post-editing settings, circumstances, etc. Segment Prioritization plays an important role in reducing the overall PE-time.
  • 4. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Motivation - Related work Incrementally PE-SMT is essentially an active learning-based SMT, including two key parts:  Active learning framework  Incremental retraining methods for model update In 2009, Haffari et al. firstly proposed a practical active learning framework for low-resource SMT where a number of high-quality parallel data are acquired from the large-scale monolingual data. In 2012, Gonzalez-Rubio et al. apply AL to the interactive MT to select informative sentences to reduce human effort rather than PE time. The most relevant previous work: Dara et al. (2014) propose a Cross Entropy Difference (CED) criterion to prioritize input segments in an AL framework for PE-based incremental MT update applications.
  • 5. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Motivation - Related work Ortiz-Martinez et al. (2010) incrementally update the feature values of the phrase table based on the pre-stored statistics related to the feature scores. Hardt and Elming (2010) propose a sentence-level retraining scheme in which a local phrase table is created and incrementally updated as a file is translated and post-edited. Denkowski et al. (2014) propose an online model adaptation for PE-SMT in which three methods are used for incremental model adaptation. Germann (2014) proposes a dynamic phrase table strategy for an interactive PE-SMT that computes phrase table entries on demand by sampling a suffix array-indexed bitext.
  • 6. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Motivation System Framework: batch-mode incremental retraining; Goal: reducing the overall PE-time; Our solution: source-side segments prioritization; Contributions: an empirical study on different segment prioritization algorithms. Motivation
  • 7. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Outline Motivation Incrementally Retrained PE-SMT Experiments and Analysis Conclusions and Future Work Segment Prioritization Algorithms
  • 8. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Incrementally Retrained PE-SMT Workflow of active learning-based incrementally retrained PE-SMT Where we are!
  • 9. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Incrementally Retrained PE-SMT Workflow of the batch-mode incrementally retraining Where we are!
  • 10. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Incrementally Retrained PE-SMT Significant Features:  Architecture:  A combination of Moses and CDEC  A combination of online and offline retraining  Speed: about 20 minutes for each iteration to update models at the scale of 4 million sentence pairs. However, the time for Moses incremental retraining scheme is more than 2 hours for the same-scale data.  Performance: no significantly difference between the incrementally retrained system and the offline retrained system
  • 11. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Outline Motivation Incrementally Retrained PE-SMT Experiments and Analysis Conclusions and Future Work Segment Prioritization Algorithms
  • 12. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie The initial parallel training corpus a monolingual corpus (translation job) F Scoring metric: segment prioritization algorithm Segment Prioritization Algorithms To prioritize the input segments, an importance score or uncertainty score for the sentence s must be calculated under a scoring metric as: Mathematical Formalization ),,(F:)( LUss  )},{(:L ii ef }{: jfU 
  • 13. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Segment Prioritization Algorithms Confidence of Translations (Confidence) Definition: The probability p(e|f) of a translation output ê that is calculated by the decoder.  Geometry n-gram (Geom n-gram) The Geom n-gram algorithm is defined as: ),|( ),|( log || :)( 1 nLxP nUxP X s n sXx N n n s n       |)|(log|:)( feps 
  • 14. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Segment Prioritization Algorithms Perplexity of Sentences (PPL) The PPL algorithm is defined as:  Cross Entropy Difference (CED) The CED algorithm is defined as: )()(:)( sHsHs LU  N-OOVs p(s) s log 10:)(   Ranking: The higher the score, the higher the rank.
  • 15. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Outline Motivation Incrementally Retrained PE-SMT Experiments and Analysis Conclusions and Future Work Segment Prioritization Algorithms
  • 16. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis •Corpora: Europarl •Language pairs: EN-DE, EN-ES, ES-EN, DE-EN •Training set: 50K •Incremental set: 10K •Devset: Newswire 2012 test set Settings for incrementally retrained SMT system •Corpora: DGT •Language pair: EN-DE, EN-ES, EN-FR, EN-PL •Training set: 50K •Incremental set: 10K •Devset: 2K (extracted)  Evaluation metric: TER to simulate the PE-time;  Batch mode: 500 sentences for each batch;  Simulated workflow: we use the references of the translated segments instead of post-edited MT output per se;
  • 17. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis Experimental results (DGT) -- TER Score Pair Sequential Geom n-gram PPL CED Confidence Gains EN-DE 55.94 56.50 56.58 55.23 51.39 4.55 EN-ES 41.65 42.40 42.48 41.58 38.69 2.96 EN-FR 44.86 46.92 46.75 44.62 41.42 3.44 EN-PL 51.38 51.53 51.63 51.16 48.09 3.29
  • 18. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis Experimental results (Europarl) -- TER Score Pair Sequential Geom n-gram PPL CED Confidence Gains DE-EN 73.73 73.87 72.60 72.48 70.56 3.17 EN-DE 80.12 80.00 79.38 79.02 77.83 2.29 ES-EN 84.14 84.20 83.75 83.25 81.82 2.32 EN-ES 64.10 64.10 63.37 63.11 62.38 1.72
  • 19. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis Iterative curves – DGT – TER Score
  • 20. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis Iterative curves – Europarl – TER Score  we can see that there is no obvious decrease (i.e. improvement) for the baseline;  the other four prioritization criteria have a trend of decreasing the TER score;  The decreasing trends show the effectiveness of these methods to prioritize the input segments.
  • 21. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis  the sentence length distribution of the “Confidence” criterion strongly fluctuates per iteration;  the fluctuations of the length distribution of “Confidence” consistently correspond to the TER score trend;  The length distributions of the “Geom n-gram”, “PPL” and “CED” methods are more smooth, but start from shorter sentences, end with longer sentences. Sentence Length Distribution
  • 22. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis  It is inferred that the prioritization score ϕ(s) of “Confidence” may correlate to the sentence length of the input segment.  Pearson correlation: EN-DE as an example Correlation between sentence length and sentence score EN-DE Geom n-gram PPL CED Confidence DGT 0.21 0.009 0.02 0.29 Europarl 0.14 0.06 0.10 0.48  Results show that the longer the sentence is, the more difficult it is to be translated;  Question: should we normalize the score by the sentence length or not in the segment prioritization task?
  • 23. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis Correlation: Normalize or not?  The average TER score for the normalized “Confidence” is 56.09 which is much higher (i.e. worse) than the baseline (55.94).
  • 24. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis OOVs: Correlation with the sentence score  In Moses, when an OOV occurs, the probability p(e|f) will be significantly decreased, i.e. the confidence of the translation becomes lower.  Pearson correlation: EN-DE as an example Pearson Correlation DGT Europarl ρ 0.9993 0.9997  As expected the more OOVs a sentence contains, the lower the confidence score, and the greater the possibility that it is ranked at the top.
  • 25. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Experiments and Analysis  Confidence: Pros and Cons  Confidence performed best in all methods. HOWEVER,  At each iteration, all the incremental segments need to be translated beforehand, which might be a problem for sentence-level incremental PE-SMT.  The sentence length at the first several iterations is much longer, which might significantly increase the amount of post-editing, which in turn may adversely affect the perception of translators/post-editors as to the utility of this approach. Implication from the analysis: develop a method that can consider the OOVs and sentence length distribution in the training data and the incremental data rather than translating them.
  • 26. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Conclusions and Future Work Conclusions 2 1 3 The “Confidence” method achieved the best results in all tasks that reduced the TER score of 1.72-4.55 absolute points. We conducted an empirical study on four different segment prioritization algorithms, namely the Sequential, Geom n-gram, PPL, CED and Confidence methods for incrementally retrained PE-SMT.. An investigation was carried out to examine the crucial factors that make the “Confidence” effective.
  • 28. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Updates: Continuous Work Actual Post-editing experiment This work will be beneficial for CAT tools (Web-based or Desktop-based)!
  • 29. www.adaptcentre.ie E-Mail: jdu@computing.dcu.ie Conclusions and Future Work  Future Work The context problem: the sorted input segments lose the sequential context that is helpful to the translators.Q Considering the context when ranking the segments. Carrying out actual PE experiments using our different segment prioritization algorithms. 1 2
  • 30. The ADAPT Centre is funded under the SFI Research Centres Programme (Grant 13/RC/2106) and is co-funded under the European Regional Development Fund. This research is supported by Science Foundation Ireland through the ADAPT Centre (Grant 13/RC/2106) (www.adaptcentre.ie) at Dublin City University and Trinity College Dublin, and by Grant 610879 for the Falcon project funded by the European Commission.

Editor's Notes

  1. Post-editing the output of a statistical machine translation (SMT) system to obtain high-quality translation has become an increasingly common application of SMT, which henceforth we refer to as post-editing-based SMT (PE-SMT). PE-SMT is often deployed as an incrementally retrained system that can learn knowledge from human post-editing outputs as early as possible to augment the SMT models to reduce PE time.
  2. In a nutshell, our motivation of this paper is: under a batch-mode incremental retraining SMT framework, with the goal of reducing the overall post-editing time, we use different source-side segment prioritization methods to perform an empirical study and to analyze their advantages and disadvantages.