Multi-aspect and diffErenTiable Evaluation of Rankings

ended

Information Retrieval (IR) deals with the automatic retrieval and ranking of information conveying items, which are relevant to a specific information need, from a large collection of items. Search engines are the most popular and well known examples of IR systems. State-of-the-art IR systems use sophisticated Machine Learning (ML) and Deep Learning (DL) models. Those models usually minimize a loss function which is built upon an IR evaluation measure, i.e. a measure that evaluates the quality of a ranked list of items. This project, Multi-aspect and diffErenTiable Evaluation of Rankings (METER), will tackle two open challenges for state-of-the-art IR systems. First, traditionally IR systems ranks items only by relevance, estimated as the semantic similarity between the user query and the information conveying items. However, beside relevance, understandability and trustworthines are fundamental for health search, or credibility and correctness should be considered for news search. Therefore, the first goal of METER will be to extend IR evaluation measures to deal with mutiple aspects. Then, these new evaluation measures will be integrated in ML algorithms, to develop multi-aspect IR systems. Second, IR measures are non-continuous and non-differentiable. This represents an issue for ML algorithms, which usually exploit gradient based approaches to minize the loss function. Therefore, the second goal of METER will be to thoroughly analyze IR evaluation measures and propose differentiability like properties which will help for the search of minima of the loss function. Therefore, METER has the potential for making both a scientific and a societal impact: 1) multi-aspect measures will be used to account for several aspect and improve the effectiveness of IR systems in different domains; 2) differentiability like properties will be exploited to improve the search of local minima for IR loss functions and better understand how this search is performed.