Introducing Mouse Actions into Interactive-Predictive Neural Machine Translation

Introducing Mouse Actions into Interactive-Predictive Neural Machine Translation Angel Navarro annamar8@prhlt.upv.es Francisco Casacuberta fcn@prhlt.upv.es Patter Recognition and Human Language Technology Research Center, Universitat Polit`ecnica de Val`encia - Camino de Vera s/n, 46022 Valencia, Spain Abstract The quality of the translations generated by Machine Translation (MT) systems has highly improved through the years, but we are still far away to obtain fully automatic high-quality translations. To generate them, translators use Computer-Assisted Translation (CAT) tools, among which we find the Interactive-Predictive Machine Translation (IPMT) systems. This paper uses bandit feedback as the principal and only information needed to generate new pre- dictions that correct the previous translations. Furthermore, the application of bandit feedback reduces the number of words that the translator needs to type in an IPMT session. In conclu- sion, this technique saves valuable time, and effort for translators. Moreover, its performance improves improves with the future advances in MT, so we recommend its application in the actuals IPMT systems. 1 Introduction In recent years there had been a large number of advances in the Machine Translation (MT) field that has led to a significant improvement in the quality of the translations. Currently, even with all the new advances, the MT systems are still not able to generate perfect ready to use translations (Toral, 2020). Indeed, MT systems usually require human post-editing in order to achieve perfect translations. The Computer-Assisted Translation (CAT) tools aim to generate high-quality translations using the knowledge and experience of professional translators while reducing the effort that they need to do. There is a large variety of CAT tools approaches, among which we focus on the Interactive-Predictive Machine Translation (IPMT) systems. Some of the recent projects in this field are TransType (Langlais et al., 2000; Esteban et al., 2004; Cubel et al., 2003), Matecat (Federico et al., 2014), CasMacat (Alabau et al., 2014, 2013; Sanchis-Trilles et al., 2014) and MMPE (Herbig et al., 2020). They aim to create a workbench with an array of innovative features that were not available in other tools when they started. IPMT is one of the main paradigms that include these projects, where an expert translator provides feedback to the system, typically using the keyboard and mouse, to generate new predictions that correct previous errors. There are two main IPMT approaches, both use usually the keyboard and mouse as the main feedback interface, but the validation process changes between prefix (Foster et al., 1997) and segments (Peris et al., 2017; Domingo et al., 2017). In this project, we use the validation by prefix approach. Figure 1 illustrates a conventional IPMT session. Initially, the user is provided with a source sentence x to be translated. At iteration 0, the IPMT system generates the first Proceedings of the 18th Biennial Machine Translation Summit Virtual USA, August 16 - 20, 2021, Volume 1: MT Research Track Page 270 SOURCE (x): Una versi´on traducida de un texto. REFERENCE (y): A translated version of a text. ITER-0 (p) (ˆsh) ( ) A written version of a story. …

Content truncated to 500 words