Till startsida
University of Gothenburg
To content Read more about how we use cookies on gu.se

Aarne Ranta "Explainable Machine Translation"

Explainable Artificial Intelligence (XAI) is an attempt to mitigate the black-box character of machine learning based AI techniques. A prime example of such techniques is Neural Machine Translation (NMT), which works via an end-to-end string transformation where the intermediate stages are machine-created vectors of floating point numbers. While NMT has reached average scores higher than older methods, its problem is that the resulting translations are sometimes completely wrong, and it is hard tell good translations from bad ones without knowing both the source and the target language.

Explainable Machine Translation (XMT) applies the XAI idea to the field of translation. It is inspired by Kurt Mehlhorn’s notion of certifying algorithms: algorithms that don’t just produce a result but a certificate - a piece of evidence that can be inspected by the user. The certificate that we propose is a formula in constructive type theory, which encodes the semantic and abstract syntactic structure of the translatable content. The formula is obtained by a parser, which itself may be a black box, but which has a formally defined relation to the source and target text. In our implementation of the idea, the parser is a neural dependency parser returning Universal Dependency (UD) trees. These trees are interpreted as type theoretical formulas, which form a translation interlingua. As the final step, the interlingua linearized into the target language by using the Grammatical Framework (GF). The last step is defined by formal rules and therefore easy to verify, whereas the parsing step - which involves interpretation of unrestricted human language - is a natural playground for machine learning approaches. By producing a formula as the outcome of this step, we can increase the reliability of the entire translation process.

Page Manager: Stergios Chatzikyriakidis|Last update: 5/25/2017

The University of Gothenburg uses cookies to provide you with the best possible user experience. By continuing on this website, you approve of our use of cookies.  What are cookies?