Ranking suggestions for black-box interactive translation prediction systems with multilayer perceptrons

Ranking suggestions for black-box interactive translation prediction systems with multilayer perceptrons Daniel Torregrosa dtorregrosa@dlsi.ua.es Juan Antonio P´erez-Ortiz japerez@dlsi.ua.es Mikel L. Forcada mlf@dlsi.ua.es Departament de Llenguatges i Sistemes Inform`atics Universitat d’Alacant, E-03690 Sant Vicent del Raspeig, Spain Abstract The objective of interactive translation prediction (ITP), a paradigm of computer-aided trans- lation, is to assist professional translators by offering context-based computer-generated sug- gestions as they type. While most state-of-the-art ITP systems are tightly coupled to a machine translation (MT) system (often created ad-hoc for this purpose), our proposal follows a resource- agnostic approach, one that does not need access to the inner workings of the bilingual resources (MT systems or any other bilingual resources) used to generate the suggestions, thus allowing to include new resources almost seamlessly. As we do not expect the user to tolerate more than a few proposals each time, the set of potential suggestions need to be filtered and ranked; the resource-agnostic approach has been evaluated before using a set of intuitive length-based and position-based heuristics designed to determine which suggestions to show, achieving promising results. In this paper, we propose a more principled suggestion ranking approach using a regressor (a multilayer perceptron) that achieves significantly better results. 1 Introduction Translation technologies are frequently used to assist professional translators. Common ap- proaches use machine translation (MT) (Hutchins and Somers, 1992) or translation memo- ries (Somers, 2003, Chapter 3) to produce a first (and usually inadequate) prototype of the trans- lation which is then edited by the professional translator in order to produce a target-language text that is adequate for publishing. In both scenarios, the suggestion may be considered by the professional translators as a source of inspiration: they will assemble the final translation on some occasions by accepting and rearranging parts of the proposal, or on other occasions by introducing their own words when an appropriate equivalent is not found in the suggestion. Our approach, described in a previous work (P´erez-Ortiz et al., 2014), follows a different paradigm known as interactive translation prediction (ITP), which, instead of presenting a translation proposal that gets reshaped by the target-language sentence formed in the translator’s mind, focuses on offering translation suggestions as the translation is carried out. Most state-of- the-art ITP approaches obtain the suggestions by means of a modified (or tailor-made) statistical machine translation system (SMT) (Koehn, 2004) that is able to provide additional information (such as word alignments, alternative translations, and scores or probabilities for the translation). These systems are able to leverage more information from the bilingual resource than if it were used unmodified as a black-box, but doing so they inherit the common requirements of SMT, ������������������������������������������������������ ���������������������������������������� namely, the dependency on the availability of extensive parallel corpora. It is worth noting that integrating other resources of bilingual information would be almost impossible in this kind of systems, as the ITP tool needs the additional information obtained from the underlying SMT engine. Unlike those previously described, the approach described here is able to use any bilingual resource …

Content truncated to 500 words