Sequence Tagging for Verb Conjugation in Romanian

Proceedings of Recent Advances in Natural Language Processing, pages 215–220, Hissar, Bulgaria, 7-13 September 2013. Sequence Tagging for Verb Conjugation in Romanian Liviu P. Dinu and Octavia-Maria S, ulea Faculty of Mathematics and Computer Science Center for Computational Linguistics University of Bucharest ldinu@fmi.unibuc.ro mary.octavia@gmail.com Vlad Niculae University of Wolverhampton vlad@vene.ro Abstract Verbs in Romanian sometimes manifest local irregularities in the form of alternat- ing letters. We present a sequence tag- ging based method for learning stem al- ternations and ending sequences. Super- vised training is based on a morphologi- cal dictionary, with a few regular expres- sion paradigms encoded by hand. Our best model improves upon previous ma- chine learning approaches to Romanian verb conjugation, and can generalize to unseen paradigms that can be constructed as variations of the ones in the training set. 1 Introduction Romanian has a rich inflectional morphology which, in the verbal domain, manifests through complex conjugational patterns. In Table 1, we give an example comparing from left to right: a regular verb, which exhibits an invariable stem, another regular verb, which also exhibits an in- variable stem but receives an additional infix -ez, a partially irregular verb, which exhibits stem al- ternation, and a completely irregular verb, which exhibits stem suppletion. The example also shows different syncretism patterns between different conjugated forms. Namely, the 1st and 4th verbs (a merge and a fi) exhibit 1sg and 3pl syncretism, the 2nd and 3rd verbs (a dansa and a purta) ex- hibit 3sg and 3pl syncretism. Given the richness in ending sequences, stem alternations, and syncretisms, many attempts have been made throughout Romanian linguistics to give conjugational classifications with stronger predictive power than the traditional, Latin- inspired one introduced by Tiktin (1905) which divided verbs into four conjugation classes based on the theme vowel surfacing as the ending in the infinitive form (Costanzo, 2011) and attributed to each of these classes only one general conjuga- tional ending sequence. The traditional analysis was followed by struc- turalist ones: Lombard (1955) arrived at 6 classes investigating 667 verbs, Felix (1964) proposed 12 classes, Gut¸u-Romalo (1968) investigated over 400 verbs and proposed 38 ending sequences, which she reduced to 10 verb classes by em- ploying specifically designed homonymy argued against, however, by Avram (1969). When at- tempting to combine the information gathered about stress shift, ending sequences, and stem alternations, Gut,u-Romalo unfortunately ended up with a very extensive classification mirroring a near-exhaustive enumeration of the verbs em- ployed. More recently, Barbu (2007) distinguished 41 conjugational classes for all tenses and 30 for the indicative present, covering 7, 295 contempo- rary Romanian verbs. Her classes did not take into account stem alternations but only ending sequences, making her classification similar to Gut,u-Romalo’s 38 ending sequences. On the op- posite end, new studies like (Feldstein, 2004) and (S¸ulea, 2012) take a unifying approach to Roma- nian conjugation that is elegant in theory but, like a merge a dansa a purta a fi to walk to dance to wear to be merg-λ dans-ez-λ …

Content truncated to 500 words