SMT at the International Maritime Organization experiences with combining in-house corpus with more general corpus

SMT at the International Maritime Organization: Experiences with Combining In-house Corpora with Out-of-domain Corpora Bruno Pouliquen World Intellectual Property Organization 34, chemin des Colombettes CH-1211 Geneva, Switzerland bruno.pouliquen@wipo.int Marcin Junczys-Dowmunt Adam Mickiewicz University ul. Umultowska 87 61-614 Pozna´n, Poland junczys@amu.edu.pl Blanca Pinero International Maritime Organization 4 Albert Embankment London SE17SR, United Kingdom bpinero@imo.org Michał Ziemski United Nations 8-14, Avenue de la Paix CH-1211 Geneva, Switzerland mziemski@unog.ch Abstract This paper presents a machine transla- tion tool – based on Moses – developed for the International Maritime Organiza- tion (IMO) for the automatic translation of documents from Spanish, French, Rus- sian and Arabic to/from English. The main challenge lies in the insufficient size of in- house corpora (especially for Russian and Arabic). The United Nations (UN) granted IMO the right to use UN resources and we describe experiments and results we obtained with different translation model combination techniques. While BLEU results remain inconclusive for combina- tions, we also analyze user preferences for certain models (when choosing betweeen IMO only or combined with UN). The combined models are perceived by transla- tors as being much better for general texts while IMO only models seem better for technical texts. 1 Introduction This paper describes the installation and training of TAPTA, an MT tool, for the automatic trans- lation of IMO documents. TAPTA has been pre- viously installed at other international organiza- tions (Pouliquen et. al 2013). IMO is a special- ized agency of the United Nations system deal- ing with safe and secure seas and the protection c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. of the marine environment. It has three working languages (English, French and Spanish) with par- allel corpora of ca. 60 million words each. A much smaller number of documents (conventions and re- ports totaling ca. 6 million words per language) are translated into the other official languages of the United Nations (Arabic, Chinese1, Russian). IMO felt that the introduction of MT would help translators in their daily work, given similar ex- periences in other UN agencies and repetitive na- ture of their documentation due to periodic report- ing. While building the SMT corpora, the spe- cific terminology used in the maritime domain and the “house style” were also important considera- tions. The large imbalance in the number of paral- lel documents between language poses a problem which we try to solve by integrating larger par- allel corpora which “complete” the IMO models (especially for Russian and Arabic)2. The corpora provided by the United Nations Secretariat were thought to be ideal for this purpose as the language pairs are the same and working practices in both translation services are very similar. Authorization was granted to merge the corpora. 2 Data and preprocessing The International Maritime Organization has 6 of- ficial languages (Arabic, English, Spanish, French, 1Work on Chinese data has been postponed and is not de- scribed in this paper. 2Documents were provided by the Documentation Division (New …

Content truncated to 500 words