Towards the Use of Language Models in Scientific Paper Recommender Systems

Within the educational and research community, Research Paper Recommender Systems debuted in the late 1990s and today, they constitute a specific research area. In this work, it is explored how the use of neural networks together with the incorporation of Natural Language Processing techniques, such as word embeddings and language models, affect the recom-mendation process of scientific papers. Three Deep Learning-based recommenders are explored: a neural collaborative filtering recommender, a recommender that uses word embeddings, and a recommender that incorporates language models. In addition, the results obtained are evaluated on two different datasets to see the effect of each of them on the recommendation process. While the first dataset only includes papers that have interested the user, the second one also includes papers that have not interested the user. The collaborative Deep Learning-based recommender constitutes the baseline against which to compare the rest of the developed recommenders. To evaluate the recommenders, each model is used to recommend 10 research papers for each user. The recommendations are evaluated and considered appropriate if they are related to the research field the user is interested in. The results confirm that the use of NLP techniques improves the performance of pure collaborative recommenders.