Post-editing (PE) combines the advantages of artificial intelligence and human intelligence, but also shifts the focus of translation work: Instead of generating text, translators correct errors in otherwise helpful suggestions in the target language. Improving the frequently recurring machine translation (MT) errors is tedious; fixing hard-to-find or complex errors makes the job cognitively demanding. "While AI is good at quickly suggesting translation drafts, only a human with in-depth knowledge of the source and target languages can analyze lexical and semantic nuances and ensure that the meaning of the translation is identical," says project leader Prof. Dr. Josef van Genabith, outlining the benefits.
The scientists from the research areas Cognitive Assistants, led by Prof. Dr. Antonio Krüger, and Multilinguality and Language Technology, led by Prof. Josef van Genabith, investigated how translation environments can support multimodal input and take into account cognitive aspects of post-editing. They also addressed the question of how automatic post-editing helps to avoid recurring errors. The team created a translation environment through a user-centered design process. The environment allows text to be crossed out or added by hand, words to be reordered by dragging and dropping, or voice commands to be used for editing.
An evaluation with professional translators shows that these interaction modalities are good extensions to mouse & keyboard, with pen and touch input proving suitable for deletion and reordering tasks, while voice commands and multimodal combinations of select & speak work well for substitutions and insertions.
However, post-editing also changes the cognitive dimension of translation. It requires a sense of the sentence in the original language and the error-prone output of the machine translation, of the surrounding context, including the readership and its cultural background. Robust approaches to automatically estimate this altered cognitive load (CL) during post-editing will enable a better understanding of whether and when machine translation tends to help or hinder the work process.
Therefore, the project team developed a sensing framework that uses a wide range of physiological and behavioral data to estimate perceived cognitive load and tested it in several studies. They demonstrated that multimodal measures of eye-, heart-, and skin-based data can be used to adapt translation environments to cognitive load.
Not only do actual errors occur during machine translation, the MT sometimes makes the same lexical or stylistic choices over and over again, with which the translator may disagree. Similar modifications are then required throughout the text. The researchers have thus investigated various deep-learning architectures for automatic post-editing (APE) that can adapt the output of each black-box MT system to a particular domain or style. Rather than learning to translate, APE systems learn from recurring human corrections and apply them to machine translation proposals for new text.
The international visibility of the scientific results achieved in the project is documented by outstanding publications, e.g., at the Conference on Human Factors in Computing Systems (CHI), the Annual Meeting of the Association for Computational Linguistics (ACL), the Conference on Computational Linguistics (COLING) or in the Machine Translation Journal.
The MMPE project is now available as open source on Github.