Machine translation with source-predicted target morphology

Machine Translation with Source-Predicted Target Morphology Joachim Daiber J.Daiber@uva.nl Khalil Sima’an K.Simaan@uva.nl Institute for Logic, Language and Computation, University of Amsterdam Science Park 107, 1098 XG Amsterdam, The Netherlands Abstract We propose a novel pipeline for translation into morphologically rich languages which consists of two steps: initially, the source string is enriched with target morphological features and then fed into a translation model which takes care of reordering and lexical choice that matches the provided morphological features. As a proof of concept we first show improved translation performance for a phrase-based model translating source strings enriched with morphological features projected through the word alignments from target words to source words. Given this potential, we present a model for predicting target morphological features on the source string and its predicate-argument structure, and tackle two major technical challenges: (1) How to fit the morphological feature set to training data? and (2) How to integrate the morphology into the back-end phrase-based model such that it can also be trained on projected (rather than predicted) features for a more efficient pipeline? For the first challenge we present a latent variable model, and show that it learns a feature set with quality comparable to a manually selected set for German. And for the second challenge we present results showing that it is possible to bridge the gap between a model trained on a predicted and another model trained on a projected morphologically enriched parallel corpus. Finally we exhibit final translation results showing promising improvement over the baseline phrase-based system. 1 Introduction Translation into a morphologically rich language poses a challenge for statistical machine trans- lation systems. Rich morphology usually goes together with relatively freer word order of the target language, which makes it difficult to predict morphology and word order in tandem. Technically speaking this difficulty could be due to data sparsity, but possibly also due to mor- phological agreement between words over long distances. In this paper we explore the idea of combating sparsity by conducting translation in a probabilistic pipeline (chain rule), whereby morphological choice may precede lexical choice and reordering. Whenever the predicate-argument structures of the source and target strings are similar, we expect that the linguistic information required for determining the morphological inflection of a plausible translation resides in the source sentence and its syntactic dependency structure. Consequently, we explore target morphology as a source-side prediction task which aims at enriching the source sentence with useful target morphological information. Practically (see Figure 1), after word aligning the sentence pairs, we project a subset of the target morphological attributes to the source side via the word alignments, and then train a model to predict these attributes on predicate-argument aspects of source dependency trees (i.e., without the source word order). Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 283 1. align 2. tag pred. model 3. project s t s′ m t′ (a) Morphology projection and pred. model training. pred. model MT MT …

Content truncated to 500 words