Supervised Morphology Generation Using Parallel Corpus

Proceedings of Recent Advances in Natural Language Processing, pages 408–414, Hissar, Bulgaria, 7-13 September 2013. Supervised Morphology Generation Using Parallel Corpus Alireza Mahmoudi, Mohsen Arabsorkhi† and Heshaam Faili School of Electrical and Computer Engineering College of Engineering University of Tehran, Tehran, Iran {ali.mahmoudi,hfaili}@ut.ac.ir †marabsorkhi@ece.ut.ac.ir Abstract Translating from English, a morphologi- cally poor language, into morphologically rich languages such as Persian comes with many challenges. In this paper, we present an approach to rich morphology predic- tion using a parallel corpus. We focus on the verb conjugation as the most important and problematic phenomenon in the con- text of morphology in Persian. We define a set of linguistic features using both En- glish and Persian linguistic information, and use an English-Persian parallel cor- pus to train our model. Then, we pre- dict six morphological features of the verb and generate inflected verb form using its lemma. In our experiments, we generate verb form with the most common feature values as a baseline. The results of our experiments show an improvement of al- most 2.1% absolute BLEU score on a test set containing 16K sentences. 1 Introduction One of the main limitations of statistical machine translation (SMT) is the sensitivity to data sparse- ness, due to the word-based or phrased-based ap- proach incorporated in SMT (Koehn et al., 2003). This problem becomes severe in the translation from or into a morphologically rich language, where a word stem appears in many completely different surface forms. Therefore, morphological analysis is an important phase in the translation from or into such languages, because it reduces the sparseness of model. So, modeling rich morphol- ogy in machine translations (MT) has received a lot of research interest in several studies. In this paper, we present a novel approach to rich morphology prediction for Persian as target language. We focus on the verb conjugation as a highly inflecting class of words and an important part of morphological processing in Persian. Our model incorporates decision tree classifier (DTC) (Quinlan, 1986), which is an approach to multi- stage decision making. In order to train DTC, we use both English and Persian linguistic informa- tion such as syntactic parse tree and dependency relations obtained from an English-Persian paral- lel corpus. Morphological features which we pre- dict and use to generate the inflected form of verb are voice (VOC), mood (MOD), number (NUM), tense (TEN), negation (NEG) and person (PER). Our proposed model can be used as a component to generate rich morphology for any kind of lan- guages and MTs. The reminder of the paper is organized as fol- lows: Section 2 briefly reviews some challenges in Persian verb conjugation, Section 3 presents our proposed approach to generate rich morphology, in Section 4 our experiments and results are pre- sented, in Section 5 we cover conclusions and fu- ture work, and finally, in Section 6 we describe related works. 2 Morphology Challenges of the Persian Verbs Verbs in Persian have a complex inflectional sys- tem (Megerdoomian, 2004). This complexity ap- pears …

Content truncated to 500 words