A distributed inflection model for translating into morphologically rich languages

A Distributed Inflection Model for Translating into Morphologically Rich Languages Ke Tran m.k.tran@uva.nl Arianna Bisazza a.bisazza@uva.nl Christof Monz c.monz@uva.nl Informatics Institute, University of Amsterdam, Science Park 904, 1098 XH Amsterdam, The Netherlands Abstract Lexical sparsity is a major challenge for machine translation into morphologically rich lan- guages. We address this problem by modeling sequences of fine-grained morphological tags in a bilingual context. To overcome the issue of ambiguous word analyses, we introduce soft tags, which are under-specified representations retaining all possible morphological attributes of a word. In order to learn distributed representations for the soft tags and their interactions we adopt a neural network approach. This approach allows for the combination of source and target side information to model a wide range of inflection phenomena. Our re-inflection ex- periments show a substantial increase in accuracy compared to a model trained on morpholog- ically disambiguated data. Integrated into an SMT decoder and evaluated for English-Italian and English-Russian translation, our model yields improvements of up to 1.0 BLEU over a competitive baseline. 1 Introduction In morphologically rich languages (MRLs), words can have many different surface forms de- pending on the grammatical context. When translating into MRLs, standard statistical ma- chine translation (SMT) models such as phrase translation models and n-gram language mod- els (LMs) often fail to select the right surface form due to the sparsity of observed word se- quences (Minkov et al., 2007; Green and DeNero, 2012). While neural LMs (Bengio et al., 2003; Schwenk, 2007) address lexical sparsity to a certain degree by projecting word sequences to distributed vector representations, they still suffer from the problem of rare words which is particularly exacerbated in MRLs (Botha and Blunsom, 2014; Jean et al., 2015; Luong et al., 2015). A potential solution to overcome data sparsity in MRLs, is to use word representations that separate the grammatical aspects of a word, i. e. inflection, from the lexical ones. Such word representations already exist for many languages in the form of morphological analyzers or lexicons. However, using these resources for statistical language modeling is far from trivial due to the issue of ambiguous word analyses. Table 1 illustrates this problem in Italian, for which a fine-grained morphological lexicon but no sizable disambiguated corpus exists. These morphological analyses1 clearly contain information that is useful to encourage grammatical agreement and, in this case, detect the highlighted error. Unfortunately, though, the needed 1In this work we use the terms analysis and tag interchangeably to denote fine-grained word annotations provided by a morphological analyzer or lexicon. Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 145 SMT idee ribelli che circola Gloss ideas rebellious that circulate Analyses noun-f:p noun-f:p con ver:impr+pres+2+s noun-m:p wh-che ver:ind+pres+3+s adj:pos+f+p det-wh:f+p adj:pos+m+p det-wh:f+s ... ... Table 1: Example of morphological error in Italian SMT output: the verb form should be plural (circolano) and not singular (circola) to agree in number with the subject. Most of the words have multiple analyses according to …

Content truncated to 500 words