SIMFIC: An Explainable Book Search Companion

Consider a digital library of fiction books. A user has a certain book in mind and is searching for books which are similar in writing style, sentiment and general content. Classic retrieval techniques applied in such scenarios lack the support to explain why a set of top-K items are relevant to the query. Explainable AI (XAI) is attempting to explain the working mechanism of complex models like retrieval systems, helping humans to trust systems as companions. XAI research suggests two prominent directions: either develop add-on methods to peek inside complex models or design simple and explainable models. We adopt the latter and present a simple and explainable model for fiction books called SIMFIC (similarity in fiction). We partition books into smaller portions called chunks and extract features aligned with human cognition like writing style, sentence complexity and sentiment per chunk. In a query by example setting, a relevance ranked list of books is created based on similarities between chunks. A novel reward-penalty scheme is used while accumulating the similarities to ensure a fair comparison between short and long books. We perform feature selection using global feature vectors and pose them as plausible explanations, arguing them as global key factors differentiating between relevant and non-relevant books. SIMFIC is compared with a benchmark retrieval model and evaluated by domain experts in a study. Majority of the users found SIMFIC to provide more helpful results with respect to writing style, sentiment and general content, compared to the baseline. The results are statistically significant.