English-to-Hindi system description for WMT 2014: Deep Source-Context Features for Moses

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 79–83, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics English-to-Hindi system description for WMT 2014: Deep Source-Context Features for Moses Marta R. Costa-juss`a1, Parth Gupta2, Rafael E. Banchs3 and Paolo Rosso2 1Centro de Investigaci´on en Computaci´on, Instituto Polit´ecnico Nacional, Mexico 2NLE Lab, PRHLT Research Center, Universitat Polit`ecnica de Val`encia 3Human Language Technology, Institute for Infocomm Research, Singapore 1marta@nlp.cic.ipn.mx,2{pgupta,prosso}@dsic.upv.es, 3rembanchs@i2r.a-star.edu.sg Abstract This paper describes the IPN-UPV partici- pation on the English-to-Hindi translation task from WMT 2014 International Evalu- ation Campaign. The system presented is based on Moses and enhanced with deep learning by means of a source-context fea- ture function. This feature depends on the input sentence to translate, which makes it more challenging to adapt it into the Moses framework. This work reports the experimental details of the system putting special emphasis on: how the feature func- tion is integrated in Moses and how the deep learning representations are trained and used. 1 Introduction This paper describes the joint participation of the Instituto Polit´ecnico Nacional (IPN) and the Uni- versitat Polit`ecnica de Valencia (UPV) in cooper- ation with Institute for Infocomm Research (I2R) on the 9th Workshop on Statistical Machine Trans- lation (WMT 2014). In particular, our participa- tion was in the English-to-Hindi translation task. Our baseline system is an standard phrase- based SMT system built with Moses (Koehn et al., 2007). Starting from this system we propose to in- troduce a source-context feature function inspired by previous works (R. Costa-juss`a and Banchs, 2011; Banchs and Costa-juss`a, 2011). The main novelty of this work is that the source-context fea- ture is computed in a new deep representation. The rest of the paper is organized as follows. Section 2 presents the motivation of this seman- tic feature and the description of how the source context feature function is added to Moses. Sec- tion 3 explains how both the latent semantic in- dexing and deep representation of sentences are used to better compute similarities among source contexts. Section 4 details the WMT experimental framework and results, which proves the relevance of the technique proposed. Finally, section 5 re- ports the main conclusions of this system descrip- tion paper. 2 Integration of a deep source-context feature function in Moses This section reports the motivation and descrip- tion of the source-context feature function, to- gether with the explanation of how it is integrated in Moses. 2.1 Motivation and description Source context information in the phrase-based system is limited to the length of the translation units (phrases). Also, all training sentences con- tribute equally to the final translation. We propose a source-context feature func- tion which measures the similarity between the input sentence and all training sen- tences. In this way, the translation unit should be extended from source|||target to source|||target|||trainingsentence, with the trainingsentence the sentence from which the source and target phrases were extracted. The measured similarity is used to favour those translation units that have been extracted from training sentences …

Content truncated to 500 words