Developing unsupervised knowledge-enhanced models to reduce the semantic gap in information retrieval

In this thesis we tackle the semantic gap, a long-standing problem in Information Retrieval (IR). The gap can be described as mismatch between users' queries and way retrieval models answer to such queries. Two main lines of work have emerged over years bridge gap: (i) use external knowledge resources enhance bag-of-words representations used by lexical models, (ii) perform matching latent documents. To deal with issue, first an in-depth evaluation through different analyses [Marchesin et al., 2019]. objective is understand what features share, if their signals are complementary, how they combined effectively address gap. particular, focuses on (semantic) neural critical aspects. Each analysis brings perspective understanding relation models. outcomes highlight differences signals, need combine them at early stages IR pipeline Then, build insights develop addressing Specifically, unsupervised that integrate from resources, evaluate for medical domain - high social value, where prominent, large presence authoritative allows us explore effective ways it. For investigate extent concepts relations stored within integrated query improve effectiveness Thus, propose several knowledge-based expansion reduction techniques [Agosti 2018, 2019; Di Nunzio These reformulations increase probability retrieving relevant documents adding or removing original highly specific terms. experimental test collections Precision Medicine particular case Clinical Decision Support (CDS) show proposed reformulations. subset allow achieve top performing results all considered collections. Regarding analyze limitations knowledge-enhanced presented literature. overcome these limitations, SAFIR 2020], framework IR. integrates learning process it does not require labeled data training. learned optimized encode linguistic CDS demonstrate when entire document collection retrieve Pseudo Relevance Feedback (PRF) methods is, pipeline. quantitative qualitative ability affected well combining complementary provide obtain better answers semantically hard