Chelo Vargas-Sierra

Bilingual Terminology Extraction from TMXA State-of-the-Art Overview

Chelo Vargas-Sierra, PhDUniversity of Alicante,

Spain

2

Key wordsOverview of terms involved in the process

1st point 2nd point 3rd point 4th point

EvaluationBATE under evaluationMeasures for accuracyQuality in use model and tasks

Terminology and extractorsTerminology managementIts timelineBATE (approaches, state of the art)

ResultsPrecision & RecallParameters & Questionnaire

INDEXMain points of this presentation

Parallel corpusTMX

Alignment levelsParagraph, sentence and word

level

ATE & BATE Precision/RecallGetting only terms and all terms

Gold standardExhaustive, manually created

bilingual glossary

Validation* Term validation facility* Which TCs are real terms?

UsabilitySoftware used to achieve

user’s objectives with effectiveness, efficiency,

and satisfaction

Quality in use modelISO standard

KEY WORDSTerms involved in the process

2. Terminology & Extractors

5

IDENTIFY

FINDRETRIEVE

the terminology in the source text adequatelyIdentify and interpret

terminological data

Retrieve and store

proper documentation and information resources

Find and use

IMPORTANCE OF TERMINOLOGYTranslators were the first professionals to be aware of term-related issues

66

Time spent to solve terminological problems (Arntz 1993,Walker 1993).+40%

In specialized translation

TERMINOLOGY MANAGEMENT

77

Managing terminology (extracting, validating, importing, adding, editing, deleting,revising, updating, exporting, publishing) is a time-comsuming process.

Time spent to solve terminological problems (Arntz 1993, Walker 1993).+40%In specialized translation


88

Managing terminology (extracting, validating, importing, adding, editing, deleting, revising,updating, exporting, publishing) is a time-comsuming process.



Terminology work is “on backstage”, and customer oremployers may not be fully aware of their befefits for QA.

99

Managing terminology (extracting, validating, importing, adding, editing, deleting, revising,updating, exporting, publishing) is a time-comsuming process.



Return on Investment (ROI) on terminology managementreported by some corporate studies (Childress, 2007;Popiolek, 2015)

90%

Terminology work is “on backstage”, and customer or employers may not be fully aware oftheir befefits for QA.

1010


Extraction

• List of terms extracted from ST• List of terms to validate (accept or reject)

Translation

• List is added to a termbase• List is translated and additional data added

Approval

• List approved by a person in charge of terminology• When the client has requested there is an addtional

step for client approval

General model por project terminologycreation (Popiolek, 2015: 351)

Monolingualextraction & validation

Importing & looking forequivalents

11

Preparing the files and import them into the BATE

Preparation: TMX import

List of candidate term pairsextracted from TMX

Bilingual extraction

TIMELINE in Terminology Managementwith bilingual extraction

12

- List of pair of terms to validate (accept or reject terms and suggested equivalents)

- Term by term and additional data are added to a term base (Synchroterm)

Validation (& data entry)

- Export bilingual terms and additional data in an available file format (.xls, .txt, .TBX, …)

- Import output file to a TDB system (to be integrated into a MT System)

Export/Import

13

Person in charge of terminology or client

Approval

Ready to useFinish

14

Bilingual Automatic Term ExtractorsTwo approaches (Foo, 2012)

EXTRACT-ALIGN1ST step: monolingual terminology extractionin both languages.2nd step: cross-linguistic matching usingword-alignment or co-occurrence statistics tofind equivalents.

Commercial systems in this approach

15

ALIGN-FILTER1ST step: word-alignment on theparallel texts.2nd step: rank the aligned units tofinally select the most likely pair ofcandidates (statistics)

TExSIS (Macken et al, 2013)

Bilingual Automatic Term ExtractorsTwo approaches (Foo, 2012)

16

Bilingual Automatic Term ExtractorsAcademic / In-house

- English-French TERMIGHT (Dagan & Church, 1994)- English-French (Kupiek, 1993)- English-Dutch (Eijk, 1993)- English-French (Gaussier, 1995)- English and Swedish (Ahrenberg et al., 1998)

- French-Japanese (Morin et al 2010, from ACABIT, Daille, 2003): not bilingual, but multilingual- Slovene and English, Luiz (Vintar, 2010); - English and Swedish ITools suite (Foo & Merkel, 2010)- English and German (Gojun et al., 2012).- English, French, German, Spanish, TTC TermSuite (Daille, 2012)- English-Spanish TBXTools (Oliver & Vázquez, 2015) (under development)- Chinese, Czech, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish: Sketch Engine (Baisa et al 2015, Koval et al 2016)

- French-German (Blank, 2000)- Japanese-English, MNH (Nakagawa & Mori, 2003)- Spanish-Basque, Elexbi (Hernaiz et al., 2006), from a TMX; - Spanish-German, Autoterm (Haller, 2008); - English-Spanish, Mutual Bilingual Term Extractor (Ha et al, 2008)- French-English, French-Italian and French-Dutch (Lefever et al., 2009)

90s

2000-2009

2010 -2016

17

Bilingual Automatic Term ExtractorsOther BATE (free / comercial)

- TermExtractor (Shimohata et al 2001)

- MemoQ's built-in term extractor- Déjà Vu - Lexicon- TermoStat Web: http://termostat.ling.umontreal.ca/- Yate (IULA)- Okapi- TerMine:

http://www.nactem.ac.uk/software/termine/- TerminologyExtractor: https://goo.gl/yA2Cuf- PRoMT- FiveFilters (web-based): http://fivefilters.org/term-

extraction/- Concordace programs: WordSmith Tools,

AntConc (free), …

90s 2010 -2016MONOLINGUAL ATE

- Xerox Terminology Suite (2001)

- SDL Multiterm Extract- Synchroterm- CrossMining (Across)- MultiTrans Term Extractor- Similis™ (by Lingua et Machina™)- Anchovy (by Swordfish)- Araya Term Extractor

- Analysis software: Sketch Engine(terminology extraction from TMX)

BILINGUAL

http://termostat.ling.umontreal.ca/

http://www.nactem.ac.uk/software/termine/

https://goo.gl/yA2Cuf

3. Evaluation

19BATE UNDER EVALUATIONSketch Engine

SIMILIS

Multiterm Extract

Synchroterm

20

Multiterm Extract SynchroTerm Similis SkE Araya

Import TMX

Extraction config.

Extraction scores

Validation facility

Term base indexation

Export to TBX (xls, txt…)

Trados TMX

MAIN FEATURES

Others Others

21

TERMS

NO TERMS

EXTRACTED NON-EXTRACTED

A B

C DRECALL = 𝐴𝐴

𝐴𝐴+𝐵𝐵

PRECISION = 𝐴𝐴𝐴𝐴+𝐶𝐶

MEASURES FOR ACCURACY

Context coveragedegree to which the

product understands the complete context of its

usage. Flexibility

Effectivenessaccuracy and completeness

with which user achieves objectives

Satisfaction

Efficiencyresources expended in

relation to the accuracy and completeness

Freedom from riskno risk for the security of

users, software, context or the environment

degree to which user needs are satisfied when a software is

used in a specified context of use

QUALITY IN USE MODELCharacteristics (ISO-IEC 25010: 2011)

23

Setting up the extraction project

CONFIGURATION

Importing the source fileTMX IMPORT

Performing the extraction to get a

bilingual list

EXTRACTION

Selecting the real terms.VALIDATION

Creating and managing term entries

RECORD CREATION

Exporting the final result for later use in CAT Systems

EXPORTATION

6 TASKS TO EVALUATEwhen performing bilingual extraction

4. Results

25

28,30

43,33

10,66

14,85

21,29

62,33

45,42

51,61

0,00

10,00

20,00

30,00

40,00

50,00

60,00

70,00

PRECISION RECALL

PRECISION & RECALL IN %

Sketch MTE Synchr Similis

EXTRACTED NON-EXTRACTED

TERMS NO TERMS TERMS NO TERMS GOLD STANDARD TCs PRECISION RECALL

A C B D

Sketch 283 717 370 653 1,000 28,30 43,34

MTE 97 813 556 910 10,66 14,85

SynchroT. 407 1505 246 1,912 21,29 62,33

Similis 337 405 316 742 45,42 51,61

26

Characteristics and sub-characteristics to be measured METRICS

EFFECTIVENESS Value between 0 (minimum) and 5 (maximum) (EFE1+EFE2+EFE3)/3EFE1.- Degree of accuracy – precision of tasks & results (P1+P7+P13+P19+P25+P31)/6

EFE2.- Degree of completeness (tasks are accomplished and results are not missing) (P2+P8+P14+P20+P26+P32)/6

EFE3.- Frequency of errors (P3+P9+P15+P21+P27+P33)/6

EFFICIENCY Value between 0 (minimum) and 5 (maximum) (EFI2+EFI3+EFI4)/3EFI1.- Time spent in the accomplishment of the task. (TM1+TM2+TM3+TM4+TM5+TM6)

EFI2.- Need to use additional sources (material, software, etc.) for the task (P4+P10+P16+P22+P28+P34)/6

EFI3.- Productivity – effort exerted by the user to carry out the task (P5+P11+P17+P23+P29+P35)/6

EFI4.- Need to consult the software Help to perform the task (P6+P12+P18+P24+P30+P36)/6

SATISFACTION Value between 0 (minimum) and 5 (maximum)

(P37+P38+P39)/3SAT1.- UsefulnessSAT2.- TrustSAT3.- Pleasure

CONTEXT COVERAGE Value between 0 (minimum) and 5 (maximum)(P40+P41+P42)/3COB1.- Context of use

COB2.- Flexibility

PARAMETERS

27

QUESTIONNAIRE42 questions grouped by tasks

28

16

1314

25

20

26

21

24

0

5

10

15

20

25

30

EXTRACTION VALIDATION

RESULTS FOR EXTRACTION & VALIDATION


3,33 3,004,00

3,50

13,83

4,06 4,44

3,00

1,50

13,00

4,11 4,22 4,33

3,00

15,67

3,723,11 3,00 3,00

12,83

0,00

2,00

4,00

6,00

8,00

10,00

12,00

14,00

16,00

18,00

EFFECTIVENESS EFFICIENCY SATISFACTION CONTEXT COVERAGE TOTAL QIU

FINAL RESULTS FOR QUALITY IN USE


29

CONCLUSIONS

• Managing terminology still takes a lot of time and effort, even in this increasingly computerized profession.

• Research on automatic terminology extraction has been around for more than 20 years and significant enhancements concerning bilingual extraction and bilingual corporaexploitation have been introduced.

• I briefly described the BATE under evaluation and illustrated some results obtained for accuracy and with the QIU model.

• Results make it clear that much more work has to be done for BATE to be considered of real help to translators and terminologists, mainly due to poor accuracy results.

Some references• Baisa, Vit, Barbora Ulipová, and Michal Cukr. 2015. “Bilingual Terminology Extraction in Sketch Engine.” In 9th

Workshop on Recent Advances in Slavonic Natural Language Processing (RASLAN 2015), 61–67.

• Childress, Mark D. 2007. “Terminology Work Saves More Time than It Cost.” Multilingual, no. April/May: 43–46.

• Foo, Jody. 2012. Computational Terminology : Exploring Bilingual and Monolingual Term Extraction.

• Foo, Jody; Merkel, Magnus. 2010. “Computer Aided Term Bank Creation and Standardization. Building StardardizeTerm Banks through Automated Term Extraction and Advanced Editing Tools.” In Terminology in Everyday Life, edited by Marcel Thelen and Fireda Steurs, 163–80. John Benjamins Publishing Company. doi: 10.1075/tlrp.13.12foo.

• Kovář, Vojtěch, Vít Baisa, and Miloš Jakubíček. 2016. “Sketch Engine for Bilingual Lexicography.” International Journal of Lexicography 29 (3): 339–52. doi:10.1093/ijl/ecw029.

• Macken, Lieve, Els Lefever, and Veronique Hoste. 2013. “TExSIS: Bilingual Terminology Extraction from ParallelCorpora Using Chunk-Based Alignment.” Terminology 19 (2013): 1–30. doi:10.1075/term.19.1.01mac.

• Oliver, Antoni, and M. Vazquez. 2015. “TBXTools: A Free, Fast and Flexible Tool for Automatic Terminology Extraction.” In Proceedings of Recent Advances in Natural Language Processing, 473–79.

• Popiolek, Monika. 2015. “Terminology Management within a Translation Quality Assurance Process.” In Handbookof Terminology (Volume 1), edited by Hendrik J Kockaert and Frieda Steurs, 341–59. John Benjamins Publishing Company. doi:10.1075/hot.1.ter6.

• Sauron, Véronique. 2002. “Tearing out the Terms : Evaluating Terms Extractors.” In Translating and the Computer 24: Proceedings from the Aslib Conference, 21-22 November 2002.

• Vintar, Špela. 2010. “Bilingual Term Recognition revisited<BR> The Bag-of-Equivalents Term Alignment Approach and Its Evaluation.” Terminology 16 (2010): 141–58. doi:10.1075/term.16.2.01vin.

University of AlicanteIULMA

Campus de San VicenteApdo. 99

03080 Alicante

Phone & FaxDirect Line: +34 965903438

Fax: +34 [email protected]

Social Media@chelovargas

Many thanks for your attentionChelo Vargas-Sierra

Chelo Vargas-Sierra

Technology

Transcript of Chelo Vargas-Sierra