Deep Neural Networks at the Service of Multilingual Parallel Sentence Extraction

Proceedings of the 27th International Conference on Computational Linguistics, pages 1372–1383 Santa Fe, New Mexico, USA, August 20-26, 2018. 1372 Deep Neural Networks at the Service of Multilingual Parallel Sentence Extraction Ahmad Aghaebrahimian University of Innsbruck Department of Translation Studies Herzog-Siegmund-Ufer 15, A-6020 Innsbruck, Austria Ahmad.Aghaebrahimian@uibk.ac.at Abstract Wikipedia provides an invaluable source of parallel multilingual data, which are in high demand for various sorts of linguistic inquiry, including both theoretical and practical studies. We intro- duce a novel end-to-end neural model for large-scale parallel data harvesting from Wikipedia. Our model is language-independent, robust, and highly scalable. We use our system for collect- ing parallel German-English, French-English and Persian-English sentences. Human evaluations at the end show the strong performance of this model in collecting high-quality parallel data. We also propose a statistical framework which extends the results of our human evaluation to other language pairs. Our model also obtained a state-of-the-art result on the German-English dataset of BUCC 2017 shared task on parallel sentence extraction from comparable corpora. Title and Abstract in German Tiefe Neuronale Netze im Dienste der Extraktion mehrsprachiger paralleler Satzpaare Wikipedia ist eine ¨uberaus wertvolle Quelle von mehrsprachigen Paralleldaten, die f¨ur eine Vielzahl von theoretischen und praktischen sprachbezogenen Fragestellungen ben¨otigt werden. Wir stellen ein neuartiges neuronales End-to-End-System f¨ur das massenhafte Sammeln von Par- alleldaten aus der Wikipedia vor. Das System ist sprachenpaarunabh¨angig, robust und weist eine hohe Skalierbarkeit auf. Wir nutzen es zur Extraktion von parallelen Satzpaaren in den Sprachen- paaren Deutsch-Englisch, Franz¨osisch-Englisch und Persisch-Englisch. Die hohe Genauigkeit unseres Systems wird durch manuelle Evaluation best¨atigt. Dar¨uber hinaus stellen wir einen statistischen Ansatz vor, mit dessen Hilfe menschliche Qualit¨atsurteile auf weitere Sprachen- paare ¨ubertragen werden k¨onnen. Unser System erzielt State-of-the-Art-Ergebnisse gemessen am deutsch-englischen Datensatz der BUCC 2017 Shared Task zur Extraktion von parallelen Satzpaaren aus Vergleichskorpora. 1 Introduction Parallel texts are an important resource in Natural Language Processing (NLP) applications and tasks. From Statistical and Neural Machine Translation (SMT, NMT) (Brown et al., 1990; Och and Ney, 2002; Kalchbrenner and Blunsom, 2013; Cho et al., 2014), to automatic lexical acquisition (Gale and Church, 1993; Melamed, 1997), cross-lingual Information Retrieval (Davis and Dunning, 1995; Oard, 1997) and annotation projection (Yarowsky et al., 2001; Diab and Resnik, 2002) all are dependent on parallel data. Generating parallel corpora from scratch is a highly time consuming and expensive task. Therefore, many studies focus on extracting parallel texts from comparable corpora (Munteanu and Marcu, 2005) such as Wikipedia. Wikipedia is a useful source of parallel sentences since humans already annotated its comparable documents. In Wikipedia, one can find both parallel, and comparable articles1. The reason is that some This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/ 1In a parallel bilingual corpus like Europarl (Koehn, 2005), all sentences in source and target languages are parallel (i.e., translations of each other) while in two comparable corpora the same topic may be described with entirely different sentences. 1373 authors prefer to explain an issue in their own words, while …

Content truncated to 500 words