Two-Step MT: Predicting Target Morphology

Two-Step MT: Predicting Target Morphology Franck Burlot, Elena Knyazeva, Thomas Lavergne, Franc¸ois Yvon LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris Saclay, 91 403 Orsay, France firstname.lastname@limsi.fr Abstract This paper describes a two-step machine translation system that addresses the issue of translating into a morphologically rich language (English to Czech), by performing separately the translation and the generation of target morphology. The first step consists in translating from English into a normal- ized version of Czech, where some morphological informa- tion has been removed. The second step retrieves this infor- mation and re-inflects the normalized output, turning it into fully inflected Czech. We introduce different setups for the second step and evaluate the quality of their predictions over different MT systems trained on different amounts of parallel and monolingual data and report ways to adapt to different data sizes, which improves the translation in low-resource conditions, as well as when large training data is available. 1. Introduction When translating into a morphologically rich language, sta- tistical machine translation (SMT) systems generally per- form poorly, generating several incorrect word forms that show errors in agreement within a noun phrase, that encode the wrong grammatical function of the word in the sentence, or that simply convey the wrong meaning from the source. Such errors come from an important assumption upon which current SMT systems rely: translation is based on a source- target mapping of one or several word forms that are memo- rized in the model disregarding any broader context. This assumption can be problematic when both lan- guages involved lack symmetry in their linguistic systems. A more analytical language such as English encodes grammat- ical information using distinct words (prepositions, negative particles, auxiliary verbs), while a language showing synthet- ical tendencies such as Czech encodes the same information in word inflection. Moreover, inflection encodes more ex- plicitly the grammatical function of a word in the sentence, while such information is in English encoded in its position in the sentence (e.g. the object is on the right side of the verb). At the word level, extracting one-to-one word mappings from the parallel data based on alignment links has two main consequences: • One source word can translate into several target words, leaving difficult choices to be made by the sys- tem. This is typically the case of the English adjective that is invariable and translates into potentially forty- two Czech word forms (varying according to seven cases, three genders and two numbers), not consider- ing lexical ambiguities. When the information is ex- tracted, the context accountable for an inflection is lost. • Such a variety of word forms that may need to be gen- erated on the target side face important sparsity issues. At test time, one may need to produce on output a word form that has not been seen in the training data, which is challenging and may require extra process- ing or enriched representations (e.g. factored models). It may also occur that the word form is contained in the model, but …

Content truncated to 500 words