Recommending reviewers for theses using artificial intelligence

Peer review in the area of scientific contributions, such as publishing papers to conferences,plays a crucial role to evaluate the integrity and correctness of the respective work. This is also the case for the process of obtaining a Master’s degree at TU Wien, where students must defend their thesis in front of an examination board. The examination board consists of the supervisor and two additional reviewers, which must be selected manually as part of the process. The manual selection of those reviewers can be prone to humanmisjudgment, causing a faulty evaluation during the examination.In this thesis, we aim to develop a recommendation engine driven by state-of-the-artmethodology in the area of artificial intelligence. This process is done in three steps:Firstly, we extract necessary data from TU Wien internal databases. Then, we define aset of different deep learning architectures and train them on our data set. The presented models are inspired by and incorporate the architectures of LSTMs, Autoencoder and Siamese Neural Networks. Each model is trained based on three text embedding modules:BERT, GPT2 and XLNet. Finally, we evaluate the models on the task of ranking potentialr esearch profiles given a thesis as input and compare it with BM25, an established stateof-the-art baseline. Furthermore, a manual evaluation for selected use cases is performed to further validate their respective performances. We conclude that models based ona Siamese Neural Network architecture achieve promising results and, in the setting of neural re-ranking, even out perform BM25. Based on the conducted experiments,we also observe that BERT as embedding module results in the best scores across all architectures. Finally, we come to the conclusion that the matching and selection process of reviewers can be optimised using the above presented state-of-the-art deep learning methods.xi