Разделы презентаций
- Разное
- Английский язык
- Астрономия
- Алгебра
- Биология
- География
- Геометрия
- Детские презентации
- Информатика
- История
- Литература
- Математика
- Медицина
- Менеджмент
- Музыка
- МХК
- Немецкий язык
- ОБЖ
- Обществознание
- Окружающий мир
- Педагогика
- Русский язык
- Технология
- Физика
- Философия
- Химия
- Шаблоны, картинки для презентаций
- Экология
- Экономика
- Юриспруденция
Quality Estimation in Machine Translation
Содержание
- 1. Quality Estimation in Machine Translation
- 2. Plan Why do we need Neural Machine
- 3. 1990s – 2010s: Statistical Machine TranslationPictures from
- 4. 2014: Neural Machine Translation A way to
- 5. RNN conceptPredicted word probabilitiesSource sentencePredicted word probabilitiesAndrew Golman, DIHT18.06.2019Encoded wordsHidden statesPicture from Stanford CS224n, lecture 6
- 6. Sequence-to-sequence translationAndrew Golman, DIHT18.06.2019 Picture from Stanford CS224n, lecture 8
- 7. Evaluation Compare with baseline translations Pay assessors
- 8. Neural Quality EstimationAndrew Golman, DIHT18.06.2019к просмотру
- 9. Similar model for Quality Estimation First bidirectional
- 10. WMT task English-German, English-Spanish, English-Russian datasets 15,000
- 11. Our experimentsUse CRFs (lattice-structured RNNs) for final classificationTry transformer architectureResults from [3]Andrew Golman, DIHT18.06.2019
- 12. SummaryAndrew Golman, DIHT18.06.2019 NMT is more efficient
- 13. ReferencesGoogle’s Neural Machine Translation System: Bridging the
- 14. Thank You for Your attentionAndrew Golman, DIHT18.06.2019mailto:golman.as@phystech.edu
- 15. Скачать презентанцию
Plan Why do we need Neural Machine Translation? What are Seq2Seq models? Can we apply same techniques for Quality Estimation task? What have we achieved?Andrew Golman, DIHT18.06.2019
Слайды и текст этой презентации
Слайд 1Quality Estimation in Machine Translation
AndreY golman
Diht, 3rd year bachelor student
18.06.2019
Andrew
Golman, DIHT
Слайд 2Plan
Why do we need Neural Machine Translation?
What are
Seq2Seq models?
Can we apply same techniques for Quality Estimation
task?What have we achieved?
Andrew Golman, DIHT
18.06.2019
Слайд 31990s – 2010s: Statistical Machine Translation
Pictures from Stanford CS224n, lecture
8
Alignment issues
Huge word- and phrase-level dictionaries
For every pair of languages!
Andrew
Golman, DIHT18.06.2019
Слайд 42014: Neural Machine Translation
A way to do Machine Translation
with a single neural network
Most complex models have 200
million parametersMore fluent than SMT
Pick up the meaning first, then phrase it
Andrew Golman, DIHT
18.06.2019
Слайд 5RNN concept
Predicted word probabilities
Source sentence
Predicted word probabilities
Andrew Golman, DIHT
18.06.2019
Encoded words
Hidden
states
Picture from Stanford CS224n, lecture 6
Слайд 6Sequence-to-sequence translation
Andrew Golman, DIHT
18.06.2019
Picture from Stanford CS224n, lecture
Слайд 7Evaluation
Compare with baseline translations
Pay assessors for marking errors
Build a neural system for error detection
Andrew Golman, DIHT
18.06.2019
Слайд 9Similar model for Quality Estimation
First bidirectional LSTM
Second bidirectional
LSTM
OK/BAD labels
Error classification
Picture from [2]
Andrew Golman, DIHT
18.06.2019
![Similar model for Quality Estimation First bidirectional LSTM Second bidirectional LSTMOK/BAD labelsError classificationPicture from [2]Andrew Golman, DIHT18.06.2019](/img/thumbs/4b9058271ef1e1a1ad6c8d2c8cc87034-800x.jpg "Quality Estimation in Machine Translation Similar model for Quality Estimation First bidirectional LSTM Second bidirectional LSTMOK/BAD")
Слайд 10WMT task
English-German, English-Spanish, English-Russian datasets
15,000 labelled sentences
Pretrained
models are allowed
Results from [2]
Andrew Golman, DIHT
18.06.2019
![WMT task English-German, English-Spanish, English-Russian datasets 15,000 labelled sentences Pretrained models are allowedResults from [2]Andrew Golman, DIHT18.06.2019](/img/thumbs/273689a39b2b597b2fffa16fd4eb3605-800x.jpg "Quality Estimation in Machine Translation WMT task English-German, English-Spanish, English-Russian datasets 15,000 labelled sentences Pretrained models")
Слайд 11Our experiments
Use CRFs (lattice-structured RNNs) for final classification
Try transformer architecture
Results
from [3]
Andrew Golman, DIHT
18.06.2019
![Our experimentsUse CRFs (lattice-structured RNNs) for final classificationTry transformer architectureResults from [3]Andrew Golman, DIHT18.06.2019](/img/thumbs/ee5a748e16847fc7ba2adfdfc028910e-800x.jpg "Quality Estimation in Machine Translation Our experimentsUse CRFs (lattice-structured RNNs) for final classificationTry transformer architectureResults from [3]Andrew Golman, DIHT18.06.2019")
Слайд 12Summary
Andrew Golman, DIHT
18.06.2019
NMT is more efficient than SMT
RNNs
are used for text generation
First NMT systems were based
on two RNNsQuality estimation models can be used to compare translation systems
Results are improving every year
Слайд 13References
Google’s Neural Machine Translation System: Bridging the Gap between Human
and Machine Translation, Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V.
Le, Mohammad Norouzi, 2016Predictor-estimator using multilevel task learning with stack propagation for neural quality estimation, Hyun Kim, Jong-Hyeok Lee and Seung-Hoon Na, 2017
Comparison of Various Architectures for World-Level Quality Estimation Mikhail Mosyagin, Amir Yagudin, Andrey Golman, 2019
Andrew Golman, DIHT
18.06.2019
