CRE-Orange : Lot1 - Etat de l'art sur les techniques d'extraction de relations n-aires autour de frame - A survey of extraction techniques for n-ary relations and their links with frames
Résumé
Relation Extraction is an important Information Extraction task and the literature is vast. A number of excellent surveys of the field exist. The first one was proposed by Bach and Badaskar (2007), followed by Sharma et al. (2016) who focused on binary and complex relation extraction techniques in the biomedical domain. Recently, Pawar et al. (2017) surveyed advances in supervised and semi-supervised methods while Smirnova and Cudré-Mauroux (2018), Niklaus et al. (2018) and Kumar (2017)delt with distant supervision, Open IE and deep learning methods for RE, respectively. Finally, a very good survey of RE (and IE in general) in a semantic web settings has been proposed by Martinez-Rodriguez et al. (2018). Based on these surveys and relevant literature in the field, this document attempts to review main existing techniques on first-order relation extraction, discussing in particular: . Binary RE, although this was not initially the focus of the project (cf. Chapter 1). As most techniques are also used for n-ary RE, this allows for introducing basis for the next chapter. We exclude from our survey EEL subtasks and relation linking. The reader can refer to Appendix 3 for a non exhaustive list of existing EEL tools. . N-ary relation extraction, seen as a semantic roles labelling task à la FrameNet (cf. Chapter 2). RE techniques that rely on other resources (e.g., PropBank, VerbNet (Kipper et al., 2000), SemLink (Bonial et al., 2013), Framester (Gangemi et al., 2016), ConceptNet (Speer and Havasi, 2012), etc.) will be the focus of the next deliverable. Also, relation linking will not be addressed here but in the final deliverable about frame-based knowledge patterns and frame representation in the Linguistic Linked Open Data (LLOD). . Open questions and challenges regarding future RE highlighting some aspects that still need to be improved or developed further including cross-sentence RE, domain/language adaptation, and lack of annotated data. (cf. Chapter 3).