This paper describes di erent pivot approaches to built SMT systems for language pairs with scarce parallel resources. The strategy is particularly interesting for Spain, a country with three o cial languages (Catalan, Basque, and Galician) besides Spanish, where it is di cult to nd parallel corpora between two of the rst three mentioned languages but it is relatively easy to collect it between Spanish and any of them. This characteristic, however, allow us to develop machine translation systems from major languages like English, to Catalan for instance, using Spanish as pivot. Such systems help these minority languages giving them global presence and promoting their use in content collaboration. We describe a English-Catalan baseline system built following the synthetic approach, we compare it with the transfer approach and comment about future enhancement that could be implemented for this language pair.
1.
Spain is a multilingual country with four o cial languages: Catalan, Euskera, Galician and Spanish. Catalan is spoken by 11.5 million people, Euskera by 1.2 million people, Galician by 3.2 million people and Spanish by 400 million people. Given the high number of Spanish speakers compared to the other languages, Spanish has much more linguistic and data resources.
The quantity of resources is relevant in statistical machine translation. The more parallel text we have, the better the translation quality. In order to face the lack of resources in translation, there are many research works on pivot approaches which consist on using a pivot language to perform a source to target translation (Bertoldi et al., 2008a) (Costa-jussa, Henr quez, y Banchs, 2011). For example, in order to translate from Galician to Catalan, we could use Spanish as pivot language. There are much more resources in Galician-Spanish and SpanishCatalan than between Galician and Catalan directly. The same could happen when interested in translating Catalan, Euskera or Galician into English. In this work, we introduce a state-of-the-art English-Catalan translation system recently built for the free web translator N-II1.
The main di erences with the
CatalanEnglish SMT system presented in
The remainder of this paper is organized The research leading to these results has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement 247762 (FAUST) and from the Spanish Ministry of Science and Innovation through the Juan de la Cierva research program and the Buceador project (TEC2009-14094-C04-01).
1available at http://www.n-ii.org as follows. Section 2 reports a brief description of the phrase-based and Ngram-based translation approaches. Section 3 presents the pivot approaches used in this paper. Section 4 describes the English-Catalan SMT system. Section 5 compares the pivot strategies in terms of translation quality and Section 6 presents the most relevant conclusions. 2.
As mentioned in the previous section, we
are working with two SMT systems: the
phrase-based
This approach to SMT performs the
translation splitting the source sentence in
segments and assigning to each segment a
bilingual phrase from a phrase-table.
Bilingual phrases are translation units that
contain source words and target words, e.g. <
unidad de traduccion j translation unit >,
and have di erent scores associated to them.
These bilingual phrases are then selected to
maximize a linear combination of feature
functions. Such strategy is known as the
loglinear model
X m=1
# mhm (e; f ) (1) where hm are di erent feature functions with weights m. The two main feature functions are the translation model (TM) and the target language model (LM). Additional models include POS target language models, lexical weights, word penalty and reordering models among others.
Moses
The base of the Ngram approach is the concept of tuple. Tuples are bilingual units with consecutive words both on the source and target side that are consistent with the word alignment. They must provide a unique monotonic segmentation of the sentence pair and they cannot be inside another tuple in the same sentence. This unique segmentation allows us to see the translation model as a language model, where the language is composed of tuples instead of words. That way, the context used in the translation model is bilingual and implicitly works as a language model with bilingual context as well. In fact, while a language model is required in phrasebased and hierarchical phrase-based systems, in Ngram-based systems it is considered just an additional feature.
This alternative approach to a translation model de nes the probability as: P (f; e) =
N Y P (f; e)n j (f; e)n 1 ; : : : ; (f; e)1 n=1 (2) where (f; e)n is the n-th tuple of hypothesis e for the source sentence f .
As additional features, we used a Part-OfSpeech (POS) language model for the target side and a target word bonus model.
We used the open source decoder MARIE (Crego, de Gispert, y Marin~o, 2005) to build the Ngram-based system. 3.
The best approaches to build a SMT
system through a pivot language are: the
cascade system, also known as the transfer
approach and the pseudo-corpus or synthetic
approach. Other pivot approaches do not
outperform these two
This approach considers the language
pairs source-pivot and pivot-target
independently. It consists in training and tuning two
di erent SMT systems and combine them in
a two-step process: rst, we translate a source
sentence using the source-pivot system; then,
we use the resulting sentence as input for the
pivot-target translation. A common variation
for this strategy presented in
Instead of considering the two language pairs independently, this approach produces a single source-target SMT system. Assuming we have a source-pivot and a pivot-target parallel corpus, we build and tuned a pivottarget SMT system and we use it to translate the pivot part from the source-pivot corpus. This results in a source-target synthetic corpus (hence the name) which is nally used to build the source-target SMT system. For the tuning process, we could also use a synthetic development corpus but an actual sourcetarget corpus is prefered, if possible. A simple variation for this approach is to build a pivot-source SMT system in order to translate the pivot part of the pivot-target corpus, and use the resulting source-target synthetic corpus to build the nal system. 4.
We present an English-Catalan SMT
baseline system, using Spanish as the pivot
language. In this case, the parallel corpus
available for the Catalan-Spanish language pair
was provided by the bilingual newspaper \El
Periodico"2 and the English-Spanish
corresponds to the train corpora provided during
the 2010 WMT's translation task3, i.e.
Europarl and News Commentary. We followed
the synthetic approach described before to
build the nal system. Therefore, the Spanish
part from the WMT Corpus was translated
into Catalan and a English-Catalan
phrasebased SMT system was built using the
resulting synthetic corpus. Table 1 shows a
summary of the statistics of both corpora. We
also used the Catalan-Spanish baseline
together with the Spanish-English baseline system
presented in the 2010's WMT
Training sents.
Running words
Vocabulary Development sents.
Running words
Vocabulary
Training sents.
Running words
Vocabulary Development sents.
Running words
Vocabulary
Test sents.
Running words
Vocabulary
As mentioned before, the Spanish-Catalan SMT system (named N-II) is based on the corpus provided by the bilingual newspaper \El Periodico". It is a Ngram-based SMT system that includes several improvements speci c to the language pair: a homonym disambiguation for the Catalan verb `soler' and Catalan possessives, special consideration for pronominal clitics, upper-case words and the Catalan apostrophe, gender concordance, numbers and time categorization and text processing for common mistakes found when writing in Catalan. The full description can be found in (Farrus et al., 2011). 4.2.
Once obtained the Catalan translation
from the Spanish section of the WMT corpus,
a phrase-based SMT system was built using
Moses as the decoder. Apart from the
baseline pipeline, the system also includes a POS
target language model computed with TnT
(Brants, 2000), numbers and time
categorization similar to N-II and the parallel corpus
was aligned considering the Catalan lemmas
computed with Freeling
Direction cat-eng eng-cat cat-eng eng-cat
We have presented an English-Catalan SMT system built using Spanish as pivot language, given the scarce resources for EnglishCatalan.
Similarly to previous research work, we have seen here that, in the particular translation task under consideration, the pseudocorpus approach constitutes the best strategy for pivot translation. Although the cascade approach clearly performs worse than the pseudo-corpus approach, it could be also bene cial to consider a system combination between these two strategies to further boost the quality of the translations.
Further work should focus on building Spanish-pivot systems between all the o cial languages and English, as well as among them. The similarities between the languages (except Basque) and the availability of parallel corpora between Spanish and the others encourage the approach.
Bertoldi, N., R. Cattoni, M. Federico, y M. Barbaiani. 2008a. FBK @ IWSLT2008. En Proc. of the International Workshop on Spoken Language Translation, paginas 34{38, Hawaii, USA.
Bertoldi, Nicola, Madalina Barbaiani, Marcello Federico, y Roldano Cattoni. 2008b. Phrase-Based Statistical Machine Translation with Pivot Languages. En Proceedings of IWSLT.
Brants, T. 2000. TnT { a statistical part-of-speech tagger. En Proc. of the Sixth Applied Natural Language Processing (ANLP-2000), Seattle, WA.
Casacuberta, F. y E. Vidal. 2004. Machine translation with inferred stochastic nitestate transducers. Computational Linguistics, 30(2):205{225.
Cohn, T. y M. Lapata. 2007. Machine Translation by Triangulation: Making E ective Use of Multi-Parallel Corpora. En Proc. of the ACL.
Costa-Jussa, M. R. y J. A. R. Fonollosa. 2009. Phrase and ngram-based statistical machine translation system combination. Applied Arti cial Intelligence: An International Journal, 23(7):694{711, August. Costa-jussa, M.R., C. Henr quez, y R. Banchs. 2011. Evaluacion de estrategias para la traduccion automatica estad stica de chino a castellano con el ingles como lengua pivote. En Proc. of the SEPLN, Huelva.
Crego, J.M., A. de Gispert, y J.B. Marin~o. 2005. An Ngram-based Statistical Machine Translation Decoder. En Proceedings of 9th European Conference on Speech Communication and Technology (Interspeech). de Gispert, A. y J.B. Marin~o. 2006. CatalanEnglish Statistical Machine Translation without Parallel Corpus: Bridging through Spanish. En Proc. of LREC 5th Workshop on Strategies for developing Machine Translation for Minority Languages (SALTMIL'06), paginas 65{68, Genova.
Farrus, M., M. R. Costa-jussa, J. B. Marin~o, M. Poch, A. Hernandez, C. Henr quez, y J. A. R. Fonollosa. 2011. Overcoming statistical machine translation limitations: error analysis and proposed solutions for the catalan-spanish language pair. Language Resoures and Evaluation, 45(2):181{208. Association for Computational Linguistics Conference (HLT/NAACL'04), paginas 169{176, Boston, USA, May.