Post-editing Experiments with MT for a Controlled Language 1 Irina Temnikova and Constantin Orasan Research Institute in Information and Language Processing University of Wolverhampton, UK E-mail: I.Temnikova2@wlv.ac.uk, C.Orasan@wlv.ac.uk Abstract This paper aims to establish whether a new controlled language (CL) for emergency-related texts could facilitate both human and machine translation. To achieve this, an experiment involving an MT engine and human translators and post-editors was conduced. In order to estimate whether the CL pre-editing has an impact on human and machine translation, the time to manually translate, the time to post-edit and the edit distance between the original and the simplified texts were measured. The results of the experiment confirm the hypothesis. 1. Introduction In emergency situations it is vital to be able to quickly and correctly comprehend messages, alerts and protocols especially because it was noticed that human understanding under stress is different than in normal conditions. For this reason the use of a controlled language (CL) is strongly recommended for writing this kind of texts. In the current global society it is rarely the case that a message is written only in one language especially when the message is addressed to various groups of people. This paper investigates to what extent controlled language re-writing can improve human and machine translation (MT) of texts from the emergency- related domain. To achieve this, the time necessary to manually translate emergency-related texts and their simplified versions was measured and compared. At the same time, we also performed another experiment where translators had to post-edit automatically translated texts. The time and the amount of editing for original and simplified texts were calculated and compared. The hypothesis investigated in this paper is that the simplified texts can be translated faster and with less effort than the original texts. If this is true, it will be reflected by the fact that translators require less time and fewer edits to post-edit a simplified text than when they deal with original texts. At the same time, translators will be able to manually translate a simplified text more easily than the original one, which will be reflected by a shorter overall translation time. The remainder of this paper is structured as follows: Section 2 presents some related work, Section 3 gives an overview of the controlled language used, Section 4 describes the experiment, Section 5 discusses the evaluation results and Section 6 provides some conclusions. 2. Related work in CL and MT The issue of whether pre-editing texts according to controlled language guidelines can improve the quality of machine translation has been investigated using different approaches ([8], [11], [12], [1]) and several CLs have been developed with the main purpose of improving the output of translation engines ([9], [6]). Krings [8] measures the post-editing effort from temporal (time necessary to post-edit a text), technical (number of additions, deletions, and cuts-and-pastes) and cognitive point of view (think-aloud protocols). O'Brien ([11], [12]) also measures the time and number of additions, deletions, and cuts-and-pastes, but employs Choice Network Analysis for measuring the cognitive effort. In contrast, Aikawa et al., [1] relies on human evaluation scores and BLEU scores ([13]) for estimating the impact of CL on MT quality, and character-based edit distance for measuring the post- editing effort. On the basis of the obtained results, they 1 This paper was been produced in the context of MESSAGE Project (full title: Alert Messages and Protocols), project financed by the European Union (JLS/2007/CIPS/022). With the support of the Prevention, Preparedness and Consequence Management of Terrorism and other Security-related Risks Programme European Commission - Directorate-General Justice, Freedom and Security. This project has been funded with support from the European Commission. This publication reflects the views only of the author, and the Commission cannot be held responsible for any use which may be made of the information contained therein