76 Integration of Machine Translation in CAT Tools: State of the Art, Evaluation and User Attitudes Anna Zaretskaya, Gloria Corpas Pastor and Miriam Seghiri Abstract There have been proposed various techniques for combining machine translation (MT) and translation memory (TM) technologies in order to enhance retrieved TM matches and increase translators’ productivity. We provide an overview of these techniques and propose a way of classifying them. According to the results of our user survey, many translators are not aware of MT feature in their computer-assisted translation (CAT) tool. However, more than a half of the population perceive such combination as useful. We argue that it is necessary to take into account user perspective when evaluating MT and CAT integration and suggest characteristics of such evaluation. 1. Introduction Complementing translation memory (TM) software 1 with automatic translation appears to boost translators’ productivity. SMT (statistical machine translation) toolkits such as Moses (Koehn et al., 2007), Microsoft Translator Hub 2 and others made it possible for companies to train their own domain-, company- or project-specific engines that provide better results compared to generic engines available for free use. For instance, the Sybase IT company (Bier, 2012) reports productivity increase from combining MT and TM, with the condition that the engine is trained on large-scale company-specific data. In the individual translator scenario, a study was carried out by (Kanavos and Kartsaklis, 2010), which showed significant productivity increase in workflows that involved MT integration. Indeed, it would be ideal if translators could not only make use of already translated texts (i.e. translation memories), but also have some technology that can help them with new parts, which are not in the TM. For instance, these parts can be automatically translated and presented to the translator. The problem is, however, to decide how exactly we should present these MT suggestions in a CAT (computer-assisted translation) tool environment. Another problem is that free publicly available engines do not always satisfy the quality requirements, which is even more true for specialised texts, where general SMT systems cannot account for specific vocabulary. While agencies can train domain-specific engines, independent freelance do not have the possibility to do that and often just refuse using any MT at all. In addition, not all agencies have resources to train good-quality engines for all language pairs they need. Finally, some customers restrict translators from using online MT services because of confidentiality issues. Despite all these issues, most state-of-the-art CAT tools do allow automatic translation integration in one way or the other. Translators receive MT suggestions along with TM matches, termbase matches, online resources, glossaries. Most CAT tools allow to install a plugin from one of the main MT service providers or to connect a proprietary engine. Moreover, it seems to be a trend in the field, as most recent CAT software releases claim to employ advanced MT technologies. The question is how MT can be integrated in the workflow in the most convenient way for users.