Passage retrieval in context: Experiments on Patents - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Passage retrieval in context: Experiments on Patents

Résumé

Focused retrieval retrieves and ranks sub-parts of documents according to their estimated relevance to a query. Many approaches akin to Structured Document retrieval exploit documents structure to effectively retrieve logic elements (titles, sections, etc...). Other approaches like Passage Retrieval aim at retrieving arbitrary length text unit (passages), considering the document as a unstructured flat text. In this work, we use the best of the two worlds. We want to (1) retrieve passages to find the best text units to retrieve ; (2) exploit the document's structure to more effectively estimate the passages' relevance. Previous work has shown that leveraging on the passage context was efficient for passage ranking. We believe that the information given by a document's structure can be used to estimate its passages' context. Firstly, we propose several ways to represent and integrate a document's structure and its sub-structures elements (sections) in the estimation of its passages' context. Secondly, we integrate these passage contexts in a state-of-the-art passage retrieval model. We evaluate our approach on two passage retrieval tasks on structured documents: CLEF IP2012 and CLEF IP2013. Our results show that using a document's structure to estimate its passages' contexts improves retrieval performances.
Fichier principal
Vignette du fichier
Coria-lucas.pdf (667.78 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03230421 , version 1 (19-05-2021)

Identifiants

  • HAL Id : hal-03230421 , version 1

Citer

Lucas Albarede, Philippe Mulhem, Lorraine Goeuriot, Claude Le Pape-Gardeux, Sylvain Marie, et al.. Passage retrieval in context: Experiments on Patents. Proceedings of CORIA'21, Apr 2021, Grenoble, France. ⟨hal-03230421⟩
156 Consultations
114 Téléchargements

Partager

Gmail Facebook X LinkedIn More