Navigating uncertain waters in information retrieval Adapting to domain shifts and complex information needs

Information retrieval (IR) in scenarios where information needs are complex, or information is otherwise scarce, poses several challenges. These complexities can arise due to several factors of variation, including noise, ambiguity, or changes in domains, particularly in low-resource settings. This thesis attempts to address these challenges, focusing on two tasks within IR: text retrieval and item recommendation.
In the realm of text retrieval, relatively little attention has been given to uncertain information needs, where users seek items based on partial or possibly incorrect information. In particular, Part I of this thesis studies the tip-of-the-tongue phenomenon. This section outlines methods to gather datasets in multiple domains with minimal annotation effort and subsequently details how such requests are formulated. In addition, it studies the effectiveness of several methods for tackling this task. Part II shifts the focus from retrieval to recommender systems, exploring how changing preferences, domains, and contexts can be handled by making effective use of available data.
In summary, this thesis advances our understanding of how users express complex tip-of-the-tongue information needs, the effectiveness of retrieval strategies for such needs, and how to cope with data scarcity under domain shifts in recommender systems by efficiently utilizing available data.