Contextual Scripture Recommendation for Writers

Abstract

Recommendation of book passages, quotes, or citations based on a given text can aid writing, research, literary analysis, and the incorporation of legal references (e.g. laws, previous cases). Each of these applications requires domain-specific data or knowledge. This thesis focuses on recommending verses of scripture, including the highly cited King James Bible, given a paragraph as input. To represent verses and queries, I compare the TF-IDF Bag-of-Words representation with the Universal Sentence Encoder. I experiment with a novel approach to weight the context of citations in the training data. I propose the use of a transformation matrix to recommend uncited verses using the contexts of cited verses. I discuss the need to evaluate both utility and diversity of recommendations, and compare automatic metrics with a small human evaluation.