Neural Multi-Source Morphological Reinflection

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 514–524, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Neural Multi-Source Morphological Reinflection Katharina Kann CIS LMU Munich, Germany kann@cis.lmu.de Ryan Cotterell Department of Computer Science Johns Hopkins University, USA ryan.cotterell@jhu.edu Hinrich Sch¨utze CIS LMU Munich, Germany Abstract We explore the task of multi-source mor- phological reinflection, which generalizes the standard, single-source version. The input consists of (i) a target tag and (ii) multiple pairs of source form and source tag for a lemma. The motivation is that it is beneficial to have access to more than one source form since different source forms can provide complementary information, e.g., different stems. We further present a novel extension to the encoder-decoder recurrent neural architecture, consisting of multiple encoders, to better solve the task. We show that our new architecture out- performs single-source reinflection models and publish our dataset for multi-source morphological reinflection to facilitate fu- ture research. 1 Introduction Morphologically rich languages still constitute a challenge for natural language processing (NLP). The increased data sparsity caused by highly in- flected word forms in certain languages causes oth- erwise state-of-the-art systems to perform worse in standard tasks, e.g., parsing (Ballesteros et al., 2015) and machine translation (Bojar et al., 2016). To create systems whose performance is not de- terred by complex morphology, the development of NLP tools for the generation and analysis of morphological forms is crucial. Indeed, these con- siderations have motivated a great deal of recent work on the topic (Ahlberg et al., 2015; Dreyer, 2011; Nicolai et al., 2015). In the area of generation, the most natural task is morphological inflection—finding an inflected form for a given target tag and lemma. An example for English is as follows: (trg:3rdSgPres, bring) Present Ind Past Ind Past Sbj Sg Pl Sg Pl Sg Pl 1 treffe treffen traf trafen tr¨afe tr¨afen 2 triffst trefft trafst traft tr¨afest tr¨afet 3 trifft treffen traf trafen tr¨afe tr¨afen Table 1: The paradigm of the strong German verb TREFFEN, which exhibits an irregular ablaut pattern. Different parts of the paradigm make use of one of four bolded theme vowels: e, i, a or ¨a. In a sense, the verbal paradigm is partitioned into subparadigms. To see why multi-source models could help in this case, starting only from the infinitive treffen makes it difficult to predict subjunctive form tr¨afest, but the additional information of the fellow subjunctive form tr¨afe makes the task easier. �→ brings. In this case, the 3rd person singular present tense of bring is generated. One general- ization of inflection is morphological reinflection (MRI) (Cotterell et al., 2016a), where we must pro- duce an inflected form from a triple of target tag, source form and source tag. The inflection task is the special case where the source form is the lemma. As an example, we may again consider generating the English past tense form from the 3rd person singular present: (trg:3rdSgPres, …

Content truncated to 500 words