English-French Verb Phrase Alignment in Europarl for Tense ...

1 English-French verb Phrase Alignment in Europarlfor Tense Translation ModelingSharid Lo aiciga , Thomas Meyer , Andrei Popescu-Belis LATL-CUI, University of Geneva Idiap Research InstituteRoute de Drize 7 Rue Marconi 191227 Carouge, Switzerland1920 Martigny, paper presents a method for verb Phrase (VP) Alignment in an english / french parallel corpus and its use for improving statisticalmachine translation (SMT) of verb tenses. The method starts from automatic word Alignment performed with GIZA++, and relies on aPOS tagger and a parser, in combination with several heuristics, in order to identify non-contiguous components of VPs, and to labelthe aligned VPs with their Tense and voice on each side.

2 This procedure is applied to the Europarl corpus, leading to the creation of asmaller, high-precision parallel corpus with about 320 000 pairs of finite VPs, which is made publicly available. This resource is usedto train a Tense predictor for translation from english into french , based on a large number of surface features. Three MT systems arecompared: (1) a baseline Phrase -based SMT; (2) a Tense -aware SMT system using the above predictions within a factored translationmodel; and (3) a system using oracle predictions from the aligned VPs.

3 For several tenses, such as the Frenchimparfait, the Tense -awareSMT system improves significantly over the baseline and is closer to the oracle :machine translation, verb tenses, verb Phrase alignment1. IntroductionThe precise Alignment of verb phrases (VPs) in parallel cor-pora is an important prerequisite for studying translation di-vergences in terms of Tense -aspect-mode (TAM) as well asfor modeling them computationally, in particular for Ma-chine Translation (MT). In this paper, we present a methodfor aligning english and french verb phrases in the Eu-roparl corpus, along with a quantitative study of Tense map-ping between these languages.

4 The resulting resource com-prises more than 300 000 pairs of aligned VPs with theirtenses, and is made publicly available. Using the resource,we train a Tense predictor for EN/FR translation and com-bine its output with the Moses Phrase -based statistical MTsystem within a factored model. This improves the trans-lation of VPs with respect to a baseline system. Moreover,for some tenses, our Tense -aware MT system is closer toan oracle MT system (which has information of the correcttarget Tense from our corpus) than to the baseline paper is organized as follows.

5 We present related workon verb tenses in MT in Section 2. We introduce our high-precision VP Alignment technique in Section 3 and ana-lyzed the obtained resource quantitatively in Section 4, interms of EN/FR Tense mappings. We put our resource to usein Section 5 to train an automatic Tense predictor, which wecombine with a statistical MT system in Section 6, measur-ing the improvement of verb translation and of the Related Work on verb Tense TranslationVerb phrases (VPs) situate the event to which they refer in aparticular time, and express its level of factuality along withthe speaker s perception of it (Aarts, 2011).

6 These Tense -aspect-modality (TAM) characteristics are encoded quitedifferently across languages. For instance, when translat-ing VPs into a morphologically rich language from a lessrich one, mismatches of the TAM categories arise. The dif-ficulties of generating highly inflected Romance VPs fromEnglish ones have been noted for languages such as Spanish(Vilar et al., 2006) and Brazilian Portuguese (Silva, 2010).Research in statistical MT (SMT) only recently started toconsider such verb Tense divergences as a translation prob-lem. For EN/ZH translation, given that Tense is not morpho-logically marked in Chinese, Gong et al.

7 (2012) built ann-gram-like sequence model that passes information frompreviously translated main verbs onto the next verb , withoverall quality improvements o f up to et al. (2007) used a classifier to insert appropriate Chi-nese aspect markers which could also be used for and Fraser (2012) trained a Phrase -based SMT sys-tem using POS-tags as disambiguation labels concatenatedto english words which corresponded to the same Germanverb. This system gained up to BLEU points over asystem without the EN/FR translation, Grisot and Cartoni (2012) haveshown that the english present perfect and simple pasttenses may correspond to eitherimparfait,pass e compos eorpass e simplein french and have identified a narrativity feature that helps to make the correct translation an automatic classifier for narrativity, Meyer et al.

8 (2013) showed that EN/FR translation of VPs in simple pasttense was improved by 10% in terms of Tense choice and In this paper, we build on this idea and labelEnglish VPs directly with their predicted french Tense ENTense ENVP FRTense FRI regret this since we are having to take ac-tion because othershavenotdonetheir le d eplore car nous devons agir du faitque d autres n ontpasfaitleur travailhavedonepresentperfect,activeontf aitpass ecompos e,activeTo this end, I would like to remind you ofthe resolution of 15 September, whichrec-ommendedthat the proposal be presentedas soon as ce sens, je vous rappelle la r esolution du15 septembre, laquellerecommandaitquela proposition soit pr esent ee dans les plusbrefs d ,activerecommandaitimparfait,activeFigur e 1.

9 Two sentences with one VP each (in bold) annotated with Tense and voice on both english and french Method for VP Phrase AlignmentOur goal is to align verb phrases from the english andFrench sides of the Europarl corpus of European Parliamentdebates (Koehn, 2005), and to annotate each with VP labelsindicating their Tense , mode, and voice (active or passive)in both languages. The targeted annotation is exemplifiedin Figure 1 on two sentences with one VP each. The auto-matic procedure proposed here discards the pairs for whichincoherent labels are found (as defined below), with theaim of selecting an unbiased, high-precision parallel cor-pus, which can be used for studies in corpus linguistics orfor training automatic following software is used to align and analyze VPs onboth the english and french sides of Europarl : GIZA++ (Och and Ney, 2003) is used to retrieve wordalignments between the two languages; a dependency parser (Henderson et al.)

10 , 2008) is usedfor parsing the english side; Morfette (Chrupa a et al., 2008) is used for Frenchlemmatization and morphological , the parallel corpus is word-aligned using GIZA++and each language is analyzed independently. From theparsing of the english sentences we retain the position,POS tags, heads and the dependency relation the french side, we use both the morphological tagsand the lemmas produced by Morfette. The three outputsare thereupon combined into a single file which containsthe english parsing aligned to the french analysis accord-ing to the Alignment produced by GIZA++.