[PhD defence] 09/12/2024 - Arthur AMALVY: "Natural language processing applied to the representation of narrative texts using character networks" (UPR LIA)

Arthur AMALVY will submit his thesis on 9 December 2024 on the subject of "Natural language processing applied to the representation of narrative texts using character networks".

Laboratory

UPR 4128 LIA - Avignon Computing Laboratory

Composition of the jury

Mr Vincent LABATUT	Avignon University	Thesis supervisor
Ms Claire GARDENT	University of Lorraine	Rapporteur
Mr Richard DUFOUR	University of Nantes	Thesis co-director
Ms Farah BENAMARA	Paul Sabatier University	Examiner
Mr David BAMMAN	University of Californa, Berkeley	Examiner
Mr Christophe CERISARA	University of Lorraine	Rapporteur

Summary

A character network represents characters as vertices in a graph, and their relationships as the edges between these vertices. In the case of literary works, they make it possible to model an entire narrative using a single mathematical object. Depending on requirements, their edges can represent different types of interaction: co-occurrence, conversation, direct action, etc. In addition, temporal changes in relationships can be modelled with dynamic networks. Thanks to this flexibility, character networks have been used to tackle a number of tasks, such as literary genre classification, story segmentation, recommendation or automatic summarisation. However, extracting these networks manually is costly, so many researchers are interested in automating this process. This automation requires solving various natural language processing tasks such as Named Entity Recognition (NER), coreference resolution or speaker attribution.

In this thesis, we present contributions to this automatic extraction process in the case of novels, as well as to applications of character networks. We propose Renard, a modular extraction pipeline that we make available under an open licence. We use it to better understand the performance of existing pipelines by studying the impact of REN and coreference resolution errors on the quality of extracted networks. We observe that the performance of both tasks is important, and strongly depends on the novel studied. For coreference resolution, we also note that the impact depends on the type of error: the precision of the coreference links extracted is particularly important in order to detect characters. In addition, we identify and contribute to two challenges for character network extraction systems. The first is the lack of literary data to train these systems. We address this by 1) proposing a new literary dataset covering REN and alias resolution and 2) proposing to use a data augmentation technique, mention replacement, in the case of cross-domain REN.

The second challenge we identify is the limited scope of transformer-based models, which can be detrimental to the performance of certain tasks. We propose to recover relevant context at the document level to alleviate the lack of information induced by this limited scope, and show that this can increase the performance of the REN task. Finally, we present contributions to character network applications in the context of two case studies. First, we use networks modelling different types of interactions in an analysis of Alfred de Musset's Lorenzaccio. Using community detection, we identify the plots of the play, quantify their relative importance and determine the interactions between them. Furthermore, we propose an automatic method to detect conspiracies. Secondly, we propose to use character networks to solve the narrative alignment task on three adaptations of George R. R. Martin's Iron Throne: the original novels, the comic books adapted from them and the TV series.

Our results show that network-based methods can be better than text-based methods, and can be combined with the latter to improve performance. We also highlight the importance of performing the alignment task on commensurable narrative units. In these two case studies, we demonstrate the value of dynamic networks.

Keywords Automatic natural language processing, Deep learning, Complex networks, Narrative documents