Universal Reordering via Linguistic Typology

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3167–3176, Osaka, Japan, December 11-17 2016. Universal Reordering via Linguistic Typology Joachim Daiber Miloˇs Stanojevi´c Khalil Sima’an Institute for Logic, Language and Computation (ILLC) University of Amsterdam {initial.last}@uva.nl Abstract In this paper we explore the novel idea of building a single universal reordering model from En- glish to a large number of target languages. To build this model we exploit typological features of word order for a large number of target languages together with source (English) syntactic features and we train this model on a single combined parallel corpus representing all (22) in- volved language pairs. We contribute experimental evidence for the usefulness of linguistically defined typological features for building such a model. When the universal reordering model is used for preordering followed by monotone translation (no reordering inside the decoder), our experiments show that this pipeline gives comparable or improved translation performance with a phrase-based baseline for a large number of language pairs (12 out of 22) from diverse language families. 1 Introduction Various linguistic theories and typological studies suggest that languages often share a number of prop- erties and that their differences fall into a small set of parameter settings (Chomsky, 1965; Greenberg, 1966; Comrie, 1981). While this intuition has influenced work on multilingual parsing (Zeman and Resnik, 2008; McDonald et al., 2011), it has found less practical use in other areas of natural language processing, such as the task of machine translation. In machine translation, significant word order differ- ences between languages often constitute a challenge to translation systems. Word order differences are frequently given special treatment, such as in the case of preordering (Xia and McCord, 2004; Neubig et al., 2012; Stanojevi´c and Sima’an, 2015, inter alia), which is a technique heavily used in practice as a means to improve both translation quality and efficiency. In preordering, word order is predicted based on manually created rules or based on statistical models estimated on word-aligned training data exploiting only source language features. This approach works well for some language pairs, however it usually demands a separate, dedicated preordering model for every source-target language pair, trained on a word-aligned corpus specific for the particular language pair. But if the similarities and differences between languages can indeed be captured with a small set of features, as linguistic theory suggests, then it seems more expedient to try to benefit from the similarities between target languages in the training data, which is not possible when training a separate preordering model for every new target language. Ideally, the word-aligned data obtained for various target languages should be combined to train a single, universal reordering model with a single set of features. The questions addressed in this paper are (1) could a linguistically inspired universal reordering model show any promising experimental results and (2) how can such a universal reordering model be built? For building an effective universal reordering model, we need access to a resource that describes the …

Content truncated to 500 words