Upload
dylan-gibson
View
214
Download
0
Embed Size (px)
Citation preview
Profile
The METIS Approach
Future Work
Evaluation
METIS II Architecture
METIS II, the continuation of the successful assessment project METIS I, is an IST Programme, with a 3-year duration (01/10/2004 – 30/09/2007).
The METIS II consortium comprises the following partners:
Institute for Language & Speech Processing [ILSP] (co-ordinator) Katholieke Universiteit Leuven [KUL] Gesellschaft zur Förderung der Angewandten Informationsforschung [GFAI] Universitat Pompeu Fabra [UPF]
METIS II is a hybridhybrid system, combining various approaches to machine translation (rule-based, statistical, pattern-matching techniques). It makes use of readily readily availableavailable resourcesresources, such as bilingual dictionaries or basic NLP tools, and it can be easily customised to handle different source (SL) and target language (TL) tags.
Most importantly, however, METIS II is innovativeinnovative because it does not need bilingual corpora for the translation process, but exclusivelyexclusively relies on monolingual TL monolingual TL corporacorpora.
METIS II handles sequences both at sentence and sub-sentential level, achieving thus to exploit the recursiverecursive property of natural language.
METIS II employs a series of weightsweights, i.e. system parameters, in various phases of the translation process. Weights are associated with system resources and employed by the pattern-matching algorithm; they can be automaticallyautomatically adjusted to customise system performance.
Four (4) language pairs have been developed as yet, namely GreekGreek, DutchDutch, GermanGerman & SpanishSpanish EnglishEnglish.
METIS IIMETIS II: : Statistical Machine Translation Statistical Machine Translation using Monolingual Corporausing Monolingual Corpora (FP6-IST-003768)
Database Database ServerServer
Lexicon
BNC Clauses
BNC Chunks
Token Generation
Rules Final TranslationFinal Translation
NLPNLP NLP tools handle the SL input yielding an SL sequence annotated with grammatical & syntactic information.
LexicoLexiconn
LookuLookupp
The SL sequence is enhanced by translation equivalents & PoS info, thus resembling a TL pattern.
Core Core EnginEnginee
The core engine of METIS II system is fed with a sequence of TL-like patterns, handled by the pattern-matching algorithm. It proceeds in 2 stages involving wider and narrower contexts, thus generating a TL sequence.
Web Web InterfaInterfacece
The end user selects the preferred SL and enters the text to be translated.
Token Token GenerationGeneration
The token generation module receives as input a sequence of translated lemmas & their respective tags; it is responsible for the production of tokens out of lemmas.
Weights
Evaluation Setup
For the system evaluation an experimental corpus extracted from real texts, mainly from newspapers, was used. It consisted of 200 200 sentences, 5050 per language pair. The test sentences were of relative complexity, containing one to two clauses each and covered various syntactic phenomena such as word-order variation, NP structure, negation, modification etc.
The reference translations have been restricted to 33 and were produced by humans, while BLEUBLEU & NISTNIST metrics have been used for the evaluation.
Evaluation Results
Greek Results
Dutch Results
Fig. 1: Comparative
analysis of the score
ranges obtained for
METIS II and
SYSTRAN using the
BLEU metric
NIST
50 4945
4135
30 2821
49 4845
3531
22
115
210
4
43
0
10
20
30
40
50
60
Score range
Num
ber
of Sen
tence
s METIS II
Systran
BLEU
3230 29
22
16 1512
10
31 3129
10 95 4
2 2
88
23
0
5
10
15
20
25
30
35
Score range
Num
ber
of Sen
tence
s METIS II
Systran
Fig. 2: Comparative
analysis of the score
ranges obtained for
METIS II and
SYSTRAN using the
NIST metric
Fig. 3: Comparative
analysis of the
scores obtained for
METIS II and
SYSTRAN using the
BLEU metricNIST RESULTS
0,0000
1,0000
2,0000
3,0000
4,0000
5,0000
6,0000
7,0000
8,0000
9,0000
10,0000
best NIST avg NIST
VERBATIM
BASE
TRANSFER
NO S LEVEL
SYSTRAN
Fig. 4: Comparative
analysis of the scores
obtained for METIS II
and SYSTRAN using
the NIST metric
BLEU RESULTS
0,0000
0,0500
0,1000
0,1500
0,2000
0,2500
0,3000
0,3500
0,4000
0,4500
0,5000
best BLEU avg BLEU
VERBATIM
BASE
TRANSFER
NO S LEVEL
SYSTRAN
German Results
Fig. 8: Comparative
analysis of the scores
obtained for different
settings of METIS II and
SYSTRAN using the
NIST metric
Spanish Results
Fig. 7: Comparative
analysis of the scores
obtained for different
settings of METIS II and
SYSTRAN using the
BLEU metric
Future work involves further investigation of METIS II system architecture. More specifically, work towards the system optimisation includes the following:
Further system testing with a big number of test suites that will have more elaborate structures and deal with a wider range of phenomena Algorithm optimisation in terms of accuracy Automatic fine tuning of weights Implementation of a post-editor module
0,43
0,44
0,45
0,46
0,47
0,48
0,49
Metis-No op.
Metis-Insertion
Metis-Deletion
Metis-Permutation
Metis-All ops.
Systran
6,25
6,3
6,35
6,4
6,45
6,5
6,55
6,6
6,65
NIST
Metis-No op.
Metis-Insertion
Metis-Deletion
Metis-Permutation
Metis-All ops.
Systran
METIS-II Systran
0.26
0.265
0.27
0.275
0.28
0.285
0.29
0.295
0.3
0.305
0.31
BLEU
METIS-II Systran
4.7394
4.7396
4.7398
4.74
4.7402
4.7404
4.7406
4.7408
4.741
4.7412
NIST
Fig. 5: Comparative
analysis of the scores
obtained for METIS II
and SYSTRAN using the
BLEU metric
Fig. 6: Comparative
analysis of the scores
obtained for METIS II
and SYSTRAN using the
NIST metric