Semantic recommendations of books using recurrent neural networks

Abstract

Digital transformations led to the development of supportive technologies, new tools for smart education, and emergent branches of research in the domain of digital library services. This paper introduces a content-based recommender system for Romanian books. The reference documents are old and were digitized via Optical Character Recognition (OCR), a process that generated noise in the conversion. The current prototype version of our system is trained on a corpus of 50 OCRed books which are split into corresponding paragraphs; thus, recommendations of related books to the user’s input query are provided only with regards to these reference documents. The trained neural models consider a bidirectional RNN layer with LSTM or GRU cells over pre-trained Romanian FastText embeddings, followed by a global max-pooling layer. The study shows competitive results on predicting books given an input text, as the proposed model achieves an overall accuracy of around 90%.