Automatic Construction of Morphologically Motivated Translation Models for Highly Inflected, Low-Resource Languages

Automatic Construction of Morphologically Motivated Translation Models for Highly Inflected, Low-Resource Languages John Hewitt johnhew@seas.upenn.edu Department of Computer and Information Science, University of Pennsylvania Philadelphia, PA 19104 Matt Post post@cs.jhu.edu David Yarowsky yarowsky@jhu.edu Center for Language and Speech Processing, Johns Hopkins University Baltimore, MD 21211 Abstract Statistical Machine Translation (SMT) of highly inflected, low-resource languages suffers from the problem of low bitext availability, which is exacerbated by large inflectional paradigms. When translating into English, rich source inflections have a high chance of being poorly es- timated or out-of-vocabulary (OOV). We present a source language-agnostic system for auto- matically constructing phrase pairs from foreign-language inflections and their morphological analyses using manually constructed datasets, including Wiktionary. We then demonstrate the utility of these phrase tables in improving translation into English from Finnish, Czech, and Turkish in simulated low-resource settings, finding substantial gains in translation quality. We report up to +2.58 BLEU in a simulated low-resource setting and +1.65 BLEU in a moderate- resource setting. We release our morphologically-motivated translation models, with tens of thousands of inflections in each of 8 languages. 1 Introduction Statistical machine translation systems are typically trained on large bilingual parallel corpora (bitext). Low-resource machine translation focuses on translation of languages for which there exists little bitext, and where translation quality is subsequently often poor. Highly inflected languages—those that exhibit large inflectional paradigms of words with a common dictio- nary entry—excacerbate the problems of a low-resource setting. Many inflections of words in an inflectional paradigm are complex and rare, and their translations are unlikely to be well- estimated even in a moderately large parallel corpus. For example, Koehn (2005) point to the highly inflected nature of Finnish as a reason for poor translation performance into English even in high-resource settings. However, even where bitext may be lacking or scarce, there are often many other resources available. One source of rich morphological information is Wiktionary.1 This paper describes a method for using resources extracted from Wiktionary to automatically map inflections in paradigms of morphologically rich languages to ranked sets of English phrasal translations. This is done by the following procedure: 1https://www.wiktionary.org/ ������������������������������������������������������ ����������������������������������������� Figure 1: The translation model (TM) construction pipeline. This depicts the process by which we map each morphologically annotated inflection to a ranked set of English phrasal transla- tions. 1. We begin with resources from the UniMorph project (Sylak-Glassman et al., 2015), which produced millions of tuples pairing inflected words forms with their lemmas and a rich morphological tag (which we refer to as a UniMorph tag or vector) that was designed to be a universal representation of morphological features (§3). 2. We next take a small set of pairs of UniMorph vectors and short English sentences that were produced in an Elicitation Corpus, designed to collect inflections that in English are expressed phrasally instead of morphologically (§4). 3. We then produce phrasal translation pairs by extracting English phrases from these sen- tences and pairing them with the foreign language through the UniMorph tag (§5). We investigate …

Content truncated to 500 words