Identifying the Translations of Idiomatic Expressions using

August 20, 2011

Stéphane HUET and Philippe LANGLAIS

NLPCS 2011 - Copenhagen

Identifying the Translations of Idiomatic Expressions

using TransSearch

Identifying Translations of Idiomatic Expressions 2 S. HUET

Idiomatic Expressions

• Oxford Companion to the English Language– Idioms are expressions of a given language,

whose sense is not predictable from the meanings and arrangement of their elements

– To fight like cat and dog– It rains cats and dogs


The problem of idiomatic expressions

• Numerous in most languages• Have idiosyncratic meanings that disturb

– Non-native persons– NLP

• In Machine Translation (MT)– Group multi-word expressions before the

alignment process [Lambert and Bancs 05]– Add a new feature encoding the fact that a

phrase is a multi-word expression [Carpuat and Diab 10]


Idiomatic expressions and MT


Idiomatic expressions and MT


Objectives of the study

• The ability of the bilingual concordancer TSRali to retrieve the translations of idiomatic expressions

• Practical issues in querying such a system


Outline

• Introduction• TransSearch• Experimental setup• Evaluations• Conclusion


• Available on the Web since 1996• Developed by the Université de Montréal• Subscribed by many professional translators in

Canada• 7.2 M queries over 6 years• Exploits an English-French translation memory• Incorporates word alignment technology


User interface


User interface

1. Retrieve sentence pairs


User interface

2. Spot translations


User interface

3. Identify the list of translations


Alignment and translation

• Word-based alignment (IBM)

• Translation spotting

This is in keeping with that strategy .

La présente mesure est conforme à cette stratégie .

La présente mesure est conforme à cette stratégie .

This is in keeping with that strategy .

– Constrained to contiguous word alignment


Post-processing steps

• Objective: to have relevant and informative translations in the top list

• Bad translations filtering– Supervised classifier– Features: alignment probabilities, POS tags

• Similar translations merging– Inflectional forms of the same canonical words

conforme à / conforme aux

– Difference by grammatical words or punctuations

à l'encontre de / à l'encontre


Type of queries

• Verbatim queries: “normal” queries– is still in its infancy

• Ellipses: for discontinuous expressions– is .. in its infancy

• Dictionary queries: for morphological expansions– be+ still in its+ infancy

• Bilingual queries: to check translations– En: is still in its infancy

Fr: en est encore à ses premiers balbutiements


Outline



Resources

• Translation memory– Canadian Hansards (1986-2007)– 8.3 M sentence pairs

• Idiom lexicon– French-English phrase book– 1,467 expressions– Some entries with 2 or 3

translations


Type of idiomatic expressions

• 2% are expressed in an informal language– She's well-upholstered.

– Il roule des mécaniques.

• 99% are used in the context of a sentence– It's fantastic to bop till you drop.

• 80% are verbal phrases used in their inflected forms– I slept like a log.

• 20% are fixed expressions– When there's a will, there's a way


Manual preprocessing

• Annotation of words judged as extra information– They put the new salesman through his paces.

• Type of extra information words– Modal verbs: can, must– Semi-modal verbs: am going to, are likely

to– Catenative verbs: want to, keep– Adverbial phrases: in Italy, when he heard

the news– Noun phrases: this poet, his latest book


Number of queries found in the TM

BilingualBilingual ENEN FRFR

Verbatim queries 36 136 248

• EN: I have no axe to grind• FR: Je ne prêche pas pour ma paroisse





+ manual removal of extra words 91 302 410

• EN: I have .. axe to grind• FR: Je .. prêche .. pour ma paroisse





+ removal of extra pronoun 106 381 509


• EN: have .. axe to grind• FR: prêche .. pour ma paroisse



• EN: have+ .. axe to grind• FR: prêcher+ .. pour sa paroisse





+ verb lemmatization 210 624 650



• EN: have+ .. axe to grind• FR: prêcher+ .. pour sa+ paroisse





+ verb lemmatization 210 624 650

+ pronoun and determiner lemmatization 238 700 705


Outline



Evaluation using the phrase book

• 700 English queries found in the TM– 36 sentence pairs per query– 13 suggested translations

• 705 French queries found in the TM– 32 sentence pairs per query– 15 suggested translations

• Evaluation restrained to 238 entries with English and French sides in a same sentence pair


Recall measured using the phrase book

• For many queries, TransSearch displays relevant translations absent from the reference– est nébuleux displayed after the reference être

dans un état second for to be in a daze– 34 correct translations displayed for to be around

the corner

Rank 1 3 5 all

English queries 41.6 59.2 65.1 74.8

French queries 41.6 54.6 62.6 76.5


Manual evaluation

• 100 French queries• 5 annotators that judged 50 queries each• 3 labels: “correct”, “wrong”, “partial”• Low Fleiss inter-annotator agreement (0.25)

Q: manger à tous les rateliers J1 J2 J3

slurps at everyone's trough correct correct correct

double-dipper partial correct partial

them pot lickers and accusing them of being at the trough and pork barelling

wrong partial wrong


Manual evaluation

• Average rank of the 1st translation labeled as correct by 1 annotator: 1.4

• For 97/100 queries, a correct translation is displayed

correctpartialwrong

42%

22%

36%


Conclusion

• 50% of the idioms of a phrase book found in the TM of TransSearch

• Users should use morphological (+) and proximity (..) operators for idioms

• Only 36% of the displayed translations were clearly wrong


Thank you for your attention

Identifying the Translations of Idiomatic Expressions using

Documents