TournaRank: When retrieval becomes document competition

Abstract

Numerous feature-based models have been recently proposed by the information retrieval community. The capability of features to express different relevance facets (query-or docu ment-dependent) can explain such a success story. Such models are most of the time supervised, thus requiring a leaming phase. To leverage the advantages of feature-based representations of documents, we propose TOURNARANK, an unsupervised approach inspired by real-life game and sport competition principles. Documents compete against each other in toumaments using features as evidences of relevanoe. Toumaments are modeled as a sequence of matches, which involve pairs of documents playing in tum their features. Once a toumament is ended, docu ments are ranked according to their number of won matches during the toumament. This principle is generic since it can be applied to any collection type. lt also provides great fl ex ibility since different alternatives can be considered by changing the toumament type, the match rules, the feature set, or the strategies adopted by documents during matches. TOURNAMNK was experimented on several collections to evaluate our mode! in different contexts and to compare it with related approaches such as Learning To Rank and fusion ones: the TREC Robust2004 collection for homogeneous documents, the TREC Web2014 (ClueWeb12) collec tion for heterogeneous web documents, and the LETOR3.0 collection for comparison with su pervised feature-based models. Information retrieval (IR) and ranking are inherently linked. Document ranking implies evaluating the relevance of documents according to a user need, expressed as a query. Pioneering models are based on manually tunable weighting sc hemes and matching functions such as BM25, Tf.Idf, or Cosine measure (Manning, Raghavan, & Schütze, 2008). More recent works intend to automatically determine these two facets of IR models. They mainly rely on a feature based representation of documents. Due to their variety, features used as relevance predictors enable a rich representatio n of potentially heterogeneous documents. The feature based works can fit into two main categories: supervised and unsupervised techniques. On the first hand, supervised approaches, lmown as Learning To Rank (LTR) methods (Tax, Bockting, & Hiemstra, 2015), have been thoroughly investigated in the literature and often offer very competitive performances provided that enough labelled training data are available (Macdonald, Santos, & Ounis, 2013). On the other hand and to the best of our lmowledge, very few works have investigated the use of features in an unsupervised scenario for adhoc retrieval. They have been proposed to fit very particular scenarios using most of the time linear combination of features.

Unsupervised Graph-based Rank Aggregation for Improved Retrieval

“…Related to the rank aggregation task, re-ranking refers to a prior family of methods that also intend to promote better results, but do not explore the inter-relationships between the ranks from the response objects. Re-ranking approaches are feature-based (Hubert et al, 2018) or rank-based (Bai and Bai, 2016). In this sense, the exploitation of inter-relationships between ranks is a potential advantage for the rank aggregation methods by definition.…”