Multilingual Language Models for the Evaluation and Selection of auto-generated Abstract Wikipedia Articles

Abstract

To enrich Wikipedia to more topics with less cost, Abstract Wikipedia project, an initiative from the Wikimedia foundation, is considered to be created . The general architecture of Natural Language Generation part of the project to automatically generate articles from wiki-data has been basically built. However, the same input wiki-data may be transformed to several sentences with different sentence structures. This thesis built multilingual data sets and utilized Natural Language Processing techniques (e.g. n-gram model and RoBERTa model) to evaluate the quality of these sentences. The report concludes, that a suitable language model is capable of evaluating and selecting auto-generated Abstract Wikipedia articles and has the potential to improve Abstract Wikipedia project. The model performance slightly varies according to the model architecture and the data set.