Hypothesis Refinement Using Agreement Constraints in Machine Translation

International Joint Conference on Natural Language Processing, pages 429–437, Nagoya, Japan, 14-18 October 2013. Hypothesis Refinement Using Agreement Constraints in Machine Translation Ankur Gandhe Carnegie Mellon University, USA ankurgan@andrew.cmu.edu Rashmi Gangadharaiah IBM Research, India rashgang@in.ibm.com Abstract Phrase-based machine translation like other data driven approaches, are often plagued by irregularities in the transla- tions of words in morphologically rich lan- guages. The phrase-pairs and the lan- guage models are unable to capture the long range dependencies which decide the inflection. This paper makes the first at- tempt at learning constraints between the language-pair where, the target language lacks rich linguistic resources, by automat- ically learning classifiers that prevent im- plausible phrases from being part of de- coding and at the same time adds consis- tent phrases. The paper also shows that this approach improves translation quality on the English-Hindi language pair. 1 Introduction Data driven Machine Translation approaches have gained significant attention as they do not require rich linguistic resources such as, parsers or manu- ally built dictionaries. However, their performance largely depends on the amount of training data available (Koehn, 2005). When the source language is morphologically rich and when the amount of data available is limited, the number out-of-vocabulary (OOV) in- creases thereby reducing the translation quality. Popovic and Ney (2004) applied transformations to OOV verbs. Yang and Kirchoff (2006) used a back-off model to transform unknown words, where, the phrase-table entries were modified such that words sharing the same root were replaced by their stems. Others (Freeman et al., 2006; Habash, 2008) found in-vocabulary words that could be treated as morphological variants. Translating into a language that is rich in mor- phology from a source language that is not mor- phologically rich also has limitations. The main reason for this is that the source language does not usually contain all the information for inflecting the words in the target half. For language-pairs that have limited amounts of training data, it is un- likely that the Translation model comes across all forms of inflections on the target phrases. Hence, some mechanism is required in order to generate these target phrases with all possible inflections and at the same time be able to filter out the im- plausible hypotheses. Certain approaches (Toutanova et al., 2008; Minkov et al., 2007; Green et al., 2012) predict inflections using syntactic and rich morphologi- cal sources for the target language. This approach cannot be applied on resource poor languages such as, Hindi or other Indian languages, which lack such rich knowledge sources. Ramanathan et al. (2009) use factored models to incorporate se- mantic relations and suffixes to generate inflec- tions and case markers while translating from En- glish to Hindi but do not consider the problem of agreement between phrases in the target sentence. William and Koehn (2011) suggested an approach to eliminate inconsistent hypotheses in a string-to- tree model by adding unification-based constraints to only the target-side of the synchronous gram- mar. Although tranfer-based MT (Lavie, 2008) uses rich feature structures, grammar rules …

Content truncated to 500 words