Automatic Annotation of Direct Speech in Written French Narratives

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7129–7147 July 9-14, 2023 ©2023 Association for Computational Linguistics Automatic Annotation of Direct Speech in Written French Narratives Noé Durandard Deezer Research & EPFL noe.durandard@epfl.ch Viet-Anh Tran Deezer Research vatran@deezer.com Gaspard Michel Deezer Research gmichel@deezer.com Elena V. Epure Deezer Research eepure@deezer.com Abstract The automatic annotation of direct speech (AADS) in written text has been often used in computational narrative understanding. Meth- ods based on either rules or deep neural net- works have been explored, in particular for En- glish or German languages. Yet, for French, our target language, not many works exist. Our goal is to create a unified framework to design and evaluate AADS models in French. For this, we consolidated the largest-to-date French narrative dataset annotated with DS per word; we adapted various baselines for sequence la- belling or from AADS in other languages; and we designed and conducted an extensive evalu- ation focused on generalisation. Results show that the task still requires substantial efforts and emphasise characteristics of each baseline. Al- though this framework could be improved, it is a step further to encourage more research on the topic. 1 Introduction Prose fiction makes whole worlds emerge. Authors make use of different strategies to create narratives and convey the storyworld. Novels intertwine nar- rators’ words to build the atmosphere and tell the story, with words stemming from characters inhab- iting the fictive world that disclose their personality and depict them directly via dialogues or direct speech (DS) (James, 2011; Hühn et al., 2014). The development of algorithms to perform the automatic annotation of direct speech (AADS) in written text has been of high interest for literary studies. This task consists in retrieving lines uttered by the characters of a narrative in contrast to words delivered by the narrator of the story. One goal of AADS has been to compare fiction works by differ- ent authors or stemming from different genres or time periods. DS was then studied as a literary de- vice carrying specific purposes and disclosing com- pelling cultural information (Muzny et al., 2017; Egbert and Mahlberg, 2020). AADS is also central Figure 1: Excerpts of Madame Bovary by Gustave Flaubert (1856). Translation by Eleanor Marx-Aveling. in narrative understanding endeavors. DS has been then considered as the main realisation of charac- ters, their means to gain volume and depth, and come alive to the readers. In this context, AADS is often regarded as a pre-processing step that enables downstream analysis such as DS speaker attribu- tion (Cuesta-Lazaro et al., 2022), that can in turn serve to assemble characters networks (Labatut and Bost, 2019), or model personas (Sang et al., 2022). AADS has been widely performed for English literature, leveraging strict formatting conventions (e.g. quotes or long dashes) to extract DS through simple regular expression—regex (Bamman et al., 2014; O’Keefe et al., 2012; Elson and McKeown, 2010). Yet, in other languages, dialogues may be less strictly segregated from narration and …

Content truncated to 500 words