Download presentation
Presentation is loading. Please wait.
Published byJuliana Miles Modified over 9 years ago
1
Evaluation of the Statistical Machine Translation Service for Croatian-English Marija Brkić Department of Informatics, University of Rijeka mbrkic@uniri.hr Tomislav Vičić Freelance teacher of economics and translator ssimonsays@gmail.com Sanja Seljan Department of Information Sciences Faculty of Humanities and Social Sciences, University of Zagreb sanja.seljan@ffzg.hr
2
I.Machine translation Syntactic transfer Example-based translation Statistically-based translation II.Evaluation Manual Automatic III.Experimental study Google Translate Service (Croatian → English) Comparison and analysis Manual evaluation IV.Conclusion 2/18 Outline
3
3/18 I. Machine translation basics Speeding up translation processSpeeding up translation process Limited human componentLimited human component Multilingual access to written materialMultilingual access to written material Limited capabilitiesLimited capabilities Help in discovering general idea behind textHelp in discovering general idea behind text For limited use onlyFor limited use only
4
4/18 Approaches Word-for-wordWord-for-word Syntactic transfer*Syntactic transfer* InterlinguaInterlingua Controlled languageControlled language Example based*Example based* Statistically based*Statistically based* Various combinationsVarious combinations
5
5/18 Syntactic transfer Involves some linguistic rulesInvolves some linguistic rules Analyzes sources and translates using intermediary linguistic representationsAnalyzes sources and translates using intermediary linguistic representations Usage still limited for particular purposes (i.e. scientific, marketing, etc.)Usage still limited for particular purposes (i.e. scientific, marketing, etc.) Examples: Systran, Eurotra, MetalExamples: Systran, Eurotra, Metal
6
6/18 Example-based Uses blocks of words (example sentences)Uses blocks of words (example sentences) Utilizes analogy principleUtilizes analogy principle Needs to be fed with infoNeeds to be fed with info System “learns” during augmenting stageSystem “learns” during augmenting stage Suitable for structurally completely different languagesSuitable for structurally completely different languages Example: translation memoriesExample: translation memories
7
7/18 Statistically-based (a.k.a. SMT) Utilizes statistical modelsUtilizes statistical models Parameters derived from bilingual corpora Phrases as n-grams (n is number of terms in a phrase)Phrases as n-grams (n is number of terms in a phrase) Requires vast quantities of matched bilingual textsRequires vast quantities of matched bilingual texts Outputs most likely match inputsOutputs most likely match inputs Does not apply linguistic rulesDoes not apply linguistic rules Attempts to match language patternsAttempts to match language patterns
8
ProblemsProblems Modeling / Learning / Decoding ApproachesApproaches Word-based / Phrase-based / Syntax-based ExampleExample Google Translate Service 8/18 Statistically based ( cont. )
9
ManualManual Human bilingual or monolingual evaluators score outputs according to fluency (grammar) and adequacy (preservation of information) Time-consuming, expensive and very subjective Automatic (BLEU, METEOR, etc.)Automatic (BLEU, METEOR, etc.) Reference translations Goal higher degree of correlation with human judgements 9/18 II. Evaluation
10
10/18 III. Experimental study Croatian – EnglishCroatian – English “Very odd couple” A lot of systematic, idiosyncratic and lexical differences Three types of texts:Three types of texts: Corpus linguistics, annotation and research methods Enterprises and Government's reform plan Washing machine manual
11
Google translate web UI
12
12/18 SMT Offers Croatian as source and target languageOffers Croatian as source and target language Statistically basedStatistically based Monolingual target language texts Aligned texts (human translations) Fluency and adequacy highly depend on available corporaFluency and adequacy highly depend on available corpora
13
Reference translations vs. candidate translationsReference translations vs. candidate translations Levels of analysis:Levels of analysis: Lexical (misuse of words, zerotones) Morphological (wrong word forms) Syntactic (word order) Semantic (preservation of original message) Usage of punctuation marks 13/18 Task
14
14/18 Manual Evaluation Procedure 6 bilingual evaluators and 21 sentences (machine translation and reference translation)6 bilingual evaluators and 21 sentences (machine translation and reference translation) 1 – 5 scale1 – 5 scale Fluency Adequacy 1 incomprehensible none 2 disfluent English little meaning 3 non- native English much meaning 4 good English most meaning 5 flawless English all meaning
15
HypothesesHypotheses There is no significant difference in assigning score 3 according to both criteria (fluency and adequacy). There are no significant differences in assigning score 3 to fluency and adequacy per each evaluator. 15/18 Chi-square Test ( 2 )
16
T here is no significant difference in assigning score 3 according to both criteria. There are no significant differences in assigning score 3 to fluency and adequacy per each evaluator. There are significant differences in assigning score 3 to fluency and adequacy for half of the evaluators. 16/18 Results 0 2 4 1st2nd3rd4th5th6th Fluency Adequacy Fluency and adequacy per average judgements
17
Usage:Usage: Basic information transfer Personal use only Improvements:Improvements: Integration with language dependent modules Human post-editing Greater number of evaluators needed 17/18 IV. Conclusion
18
Thank you on your attention!
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.