Discourse and Document-level Information for Evaluating Language Output Tasks

Evaluating the quality of language output tasks such as Machine Translation (MT) and Automatic Summarisation (AS) is a challenging topic in Natural Language Processing (NLP). Recently, techniques focusing only on the use of outputs of the systems and source information have been investigated. In MT, this is referred to as Quality Estimation (QE), an approach that uses machine learning techniques to predict the quality of unseen data, generalising from a few labelled data points. Traditional QE research addresses sentencelevel QE evaluation and prediction, disregarding document-level information. Documentlevel QE requires a different set up from sentence-level, which makes the study of appropriate quality scores, features and models necessary. Our aim is to explore documentlevel QE of MT, focusing on discourse information. However, the findings of this research can improve other NLP tasks, such as AS.