machine translation and the translator - termcoord · machine translation and the translator...

Machine Translationand the Translator

Philipp Koehn

8 April 2015

Philipp Koehn Machine Translation 8 April 2015

1About me

• Professor at Johns Hopkins University (US), University of Edinburgh (Scotland)

• Author of textbook on statistical machine translation

• Leading development of open source Moses toolkit

– developed since 2006

– reference implementationof state-of-the art methods

– used in academiaas benchmark and testbed

– extensive commercial deployment(20% of MT market)


2Recent Projects

• Speech translation

• Computer aided translation

Development of an open source toolkit tightly integrated with machine translation

Novel types of assistance for translators Adaptation of machine translation to user needs

• Open source infrastructure

MOSES CORE


3

how good is machine translation?


4Machine Translation: Chinese

5Machine Translation: French


6Quality

HTER assessment

0%publishable

10%editable

20%

30% gistable

40% triagable

50%

(scale developed in preparation of DARPA GALE programme)


7Applications

HTER assessment application examples

0% Seamless bridging of language dividepublishable Automatic publication of official announcements

10%editable Increased productivity of human translators

20% Access to official publicationsMulti-lingual communication (chat, social networks)

30% gistable Information gatheringTrend spotting

40% triagable Identifying relevant documents

50%


8Current State of the Art

HTER assessment language pairs and domains

0%publishable French-English restricted domain

10% French-English news storieseditable German-English news stories

20% Chinese-English news storiesEnglish-Czech open domain

30% gistable English-Japanese open domain

40% triagable

50%

(informal rough estimates by presenter)


9

big picture


10A Clear Plan

Source Target

Lexical Transfer

Interlingua


11A Clear Plan

Source Target

Lexical Transfer

Syntactic Transfer

InterlinguaAna

lysis

Generation


12A Clear Plan

Source Target

Lexical Transfer

Syntactic Transfer

Semantic Transfer

Interlingua

Analys

isGeneration


13A Clear Plan

Source Target

Lexical Transfer

Syntactic Transfer

Semantic Transfer

Interlingua

Analys

isGeneration


14Learning from Data

statistical analysis statistical analysis

foreign/Englishparallel text

Englishtext

TranslationModel

LanguageModel

Decoding Algorithm


15Finding the Best Translation

eBEST = argmaxe p(e|f)


16

why is that a good plan?


17Word Translation Problems

• Words are ambiguous

He deposited money in a bank accountwith a high interest rate.

Sitting on the bank of the Mississippi,a passing ship piqued his interest.

• How do we find the right meaning, and thus translation?

• Context should be helpful


18Phrase Translation Problems

• Idiomatic phrases are not compositional

It’s raining cats and dogs.

Es schuttet aus Eimern.(it pours from buckets.)

• How can we translate such larger units?


19Syntactic Translation Problems

• Languages have different sentence structure

das behaupten sie wenigstensthis claim they at leastthe she

19Syntactic Translation Problems

• Languages have different sentence structure

das behaupten sie wenigstensthis claim they at leastthe she

• Convert from object-verb-subject (OVS) to subject-verb-object (SVO)

• Ambiguities can be resolved through syntactic analysis

– the meaning the of das not possible (not a noun phrase)– the meaning she of sie not possible (subject-verb agreement)


20Semantic Translation Problems

• Pronominal anaphora

I saw the movie and it is good.

• How to translate it into German (or French)?





– it refers to movie– movie translates to Film– Film has masculine gender– ergo: it must be translated into masculine pronoun er

• We are not handling this very well[Le Nagard and Koehn, 2010]


• Coreference

Whenever I visit my uncle and his daughters,I can’t decide who is my favorite cousin.

• How to translate cousin into German? Male or female?

• Complex inference required


22No Single Right Answer

Israeli officials are responsible for airport security.Israel is in charge of the security at this airport.The security work for this airport is the responsibility of the Israel government.Israeli side was in charge of the security of this airport.Israel is responsible for the airport’s security.Israel is responsible for safety work at this airport.Israel presides over the security of the airport.Israel took charge of the airport security.The safety of this airport is taken charge of by Israel.This airport’s security is the responsibility of the Israeli security officials.



• What is the best translation?

Sicherheit→ security 14,516Sicherheit→ safety 10,015Sicherheit→ certainty 334





• Counts in European Parliament corpus





• Phrasal rulesSicherheitspolitik→ security policy 1580

Sicherheitspolitik→ safety policy 13Sicherheitspolitik→ certainty policy 0

Lebensmittelsicherheit→ food security 51Lebensmittelsicherheit→ food safety 1084Lebensmittelsicherheit→ food certainty 0

Rechtssicherheit→ legal security 156Rechtssicherheit→ legal safety 5

Rechtssicherheit→ legal certainty 723


26

better models


27Phrase-Based Model

natürlich hat John Spaß am Spiel

of course John has fun with the game

• Foreign input is segmented in phrases

• Each phrase is translated into English

• Phrases are reordered

• Workhorse of today’s statistical machine translation


28Synchronous Grammar Rules

• Nonterminal rules

NP→ DET1 NN2 JJ3 | DET1 JJ3 NN2

• Terminal rules

N→maison | house

NP→ la maison bleue | the blue house

• Mixed rules

NP→ la maison JJ1 | the JJ1 house


29Learning Rules

I shall be passing on to you some comments

PRP MD VB VBG RP TO PRP DT NNS

NPPP

VP

VP

VP

S

Ich werde Ihnen die entsprechenden Anmerkungen aushändigen

Extracted rule: VP→ X1 X2 aushandigen | passing on PP1 NP2


31New State of the Art

• Good results for German–English [WMT 2014]

language pair syntax preferredGerman–English 57%English–German 55%

• Mixed for other language pairs

language pair syntax preferredCzech–English 44%Russian–English 44%Hindi–English 54%

• Also very successful for Chinese–English


32

better machine learning


rank

frequency

33Sparse Data

• Statistical estimation often suffers from sparse data

• Zipf’s law

– most words are extremely rare

– frequency× rank = constant

• Statistics from Europarl

– the occurs 1,929,379 times

– large tail of words that occur once:33,447 words, for instancecornflakes, mathematicians, or Tazhikhistan


34Brown Clusters

• Main idea: share evidence with similar words

• Cluster words to reduce sparsity

presented the laconic messagepursued these pompous lesson

aired that melancholic lettercommissioned this bouncy counterfactuals

published incompletable stunner

• For instance: use in language modeling

p(cluster(message)|cluster(presented), cluster(the), class(laconic)


35Word Embeddings


36Word Embeddings


37Deep Learning

• Autoencoders

– first: learn embeddings unsupervised– then: supervised learning of task

• Neural network language models

– several implementations– some integrated in Moses

• Neural networks everywhere

– translation model– reordering model– operation sequence model


38

data


39Big Data

For many language pairs, lots of text available.

Text you read in your lifetime

Translated textavailable

English textavailable

300 million words

billions of words

trillions of words


40Mining the Web

• Largest source for test: the World Wide Web

• Common Crawl

– publicly available crawl of the web

– hosted by Amazon Web Services, but can be downloaded

– regularly updated (semi-annual)

– 2-4 billion web pages per crawl

• Currently filling up our hard drives


41Monolingual Data

• Starting point: 35TB of text

• Processing pipeline [Buck et al., 2014]

– language detection– reduplication– normalization of Unicode characters– sentence splitting

• Obtained corpora

Language Lines (B) Tokens (B) Bytes BLEU (WMT)English 59.13 975.63 5.14 TB -German 3.87 51.93 317.46 GB +0.5Spanish 3.50 62.21 337.16 GB -French 3.04 49.31 273.96 GB +0.6Russian 1.79 21.41 220.62 GB +1.2Czech 0.47 5.79 34.67 GB +0.6


42Parallel Data

• Basic processing pipeline [Smith et al., 2013]– find parallel web pages (based on URL only)– align document by HTML structure– sentence splitting and tokenization– sentence alignment– filtering (remove boilerplate)

• Obtained corporaFrench German Spanish Russian Japanese Chinese

Segments 10.2M 7.50M 5.67M 3.58M 1.70M 1.42MForeign Tokens 128M 79.9M 71.5M 34.7M 9.91M 8.14MEnglish Tokens 118M 87.5M 67.6M 36.7M 19.1M 14.8M

Bengali Farsi Telugu Somali Kannada PashtoSegments 59.9K 44.2K 50.6K 52.6K 34.5K 28.0KForeign Tokens 573K 477K 336K 318K 305K 208KEnglish Tokens 537K 459K 358K 325K 297K 218K

• Much more work needed!


43

computer aided translation


44Post-Editing Machine Translation

(source: Autodesk)


45Interactivity

• Traditional professional translation approaches

– translation from scratch– post-editing translation memory match– post-editing machine translation output

• More interactive collaboration between machine and professional?


46Interactive Machine Translation

Input Sentence

Er hat seit Monaten geplant, im April einen Vortrag in New York zu halten.

Professional Translator

|



Input Sentence



| He



Input Sentence



He | has



Input Sentence



He has | for months



Input Sentence



He planned |



Input Sentence



He planned | for months


52Word Alignment Visualization

Input Sentence



He planned for months to give a lecture in New York | in


53Word Alignment Visualization

Input Sentence





54Shading off Translated Material

Input Sentence

Er hat seit Monaten geplant, im April einen Vortrag in New York zu halten .




55Choices

• Trigger the passive vocabulary

• Display multiple translations for words and phrases

er hat seit Monaten geplant , im April einen Vortrag ...

he has for months the plan in April a lecture ...it has for months now planned , in April a presentation ...

he was for several months planned to in the April a speech ...he has made since months the pipeline in April of a statement ...

he did for many months scheduled the April a general ...

• Rank and color-highlight by probability of each translation

• Prefer diversity


56Instant Feedback Loop

source text

MTtranslation

human translation

MTengine

post-editre-train

translate


57CASMACAT Home Edition

• Available as open source software

• Features

– installation on any desktop machine

– allows training of MT engines

– all new types of assistance

– incremental updating of models

• Warning: still in development stage(help welcome!)


58

summary


59Summary

• Machine translation is not perfect, but useful

• Better models(esp. syntax)

• Better machine learning(esp. neural networks)

• More data

• Closer integration with target application(e.g., computer aided translation)


60Thank You

questions?


machine translation and the translator - termcoord · machine translation and the translator...

Documents