In the context of ERC Demoserie european project, the ERC Demoserie team and the Japanese French Laboratory of Informatics organize an interdisciplinary workshop on TV content analysis bringing together researchers in social science and researchers in computer science.

Detailed program

February 29th

9:45 : Opening  -  Pr. Philippe Codognet, Japanese French Laboratory for Informatics, The University of Tokyo

Session 1. TV series and AI (chair: Sandra Laugier)

• 10:00 - Maxime Valette, CTO and founder of BetaSeries
  History and Methods of Recommendation through Data" - Why and how data have transformed cultural recommendation, businesses in the cultural and creative industries, and platforms.
  This presentation delves into the evolution and methodologies of data-driven recommendations within the cultural and creative sectors. It aims to explore the pivotal role that data analytics and machine learning have played in revolutionizing how cultural content is recommended to users across various platforms. Participants will gain a comprehensive understanding of the role of data in cultural recommendations, the technologies behind these systems, and their implications for businesses, creators, and consumers in the cultural and creative sectors.

• 10:45 - Alexandre Gefen, Thalim, CNRS
  Making series talk
  We propose to think about and experiment with how LLM-based generative AIs can make viewers and critics of a series talk to enable dialogue-based recommendations. Even riskier, we will attempt to train a language model on texts extracted from subtitles: what do series tell us if they are mediated by a language model? Will the moral orientation of the AI's responses be modified and enhanced by training on the discourse of the series, in the analogous way that our language, i.e. our ethics, is semantically educated by the series?

Lunch break

Session 2. Multimodal analysis of multimedia documents (chair: Camille Guinaudeau)

• 14:00 - Benjamin Renoust,  Osaka University, Institute for Datability Science & Median Technologies
  (Multilayer) networks visual analytics for humanities
  In this presentation we introduce how networks and multilayer networks can help model humanities subjects as a complex system. The models make a good support to use the panel of graph analysis, and we propose their integration in different types of visualization and interactions to explore the complex systems in depth, and to localize and identify interesting trends and outliers. With the recent advances in "artificial intelligence" models at large, the combination of multiple modalities analysis has never been easier, and we present additional strategies to leverage on this network analysis to improve classification and explainability of models and predictions. The present will cover multiple topics of human science applications: news analysis, multimedia archives, movies recommendations, social networks, art analysis, and a glimpse of medical analysis.

• 14:45 - Yusuke Mori, Research Center for Advanced Science and Technology, The University of Tokyo
  Emotion, Humor, and Morality in Stories and Narratives
  Attention to emotion and humor is essential in human communication. Humor that puts people in a pleasant mood captures the recipient's attention and makes the message more memorable. However, sometimes morally problematic jokes can hurt people, which must be considered when exercising creativity. Knowing how humans exercise creativity in text and multimodal media and how they understand it will lead to better communication among humans and between humans and computers. It will also lead to the further creation of exciting works. We have been working on stories, narratives, jokes, and (image-based) memes, considering their relationship to emotion and humor. This presentation will introduce such efforts.

Coffee break

Session 3. Multimodal analysis of multimedia documents (chair: Philippe Codognet)

• 15:45 - Akiko Ogasawara and Kimiko Aoki, Broadcasting Culture Research Institute, NHK, Japan Broadcasting Corporation
  On-Screen Gender Representation in Japanese Television Programs – from 2021/2022 Annual Snapshot Surveys by the Diversity Research Project, NHK Broadcasting Culture Research Institute
  NHK Broadcasting Culture Research Institute’s annual surveys of gender representation in Japanese television programs aims to address the gap in data for media content diversity in Japan. Since 2021, the institute’s Diversity Research Project has conducted annual snapshot surveys of gender representations in television programs broadcast on NHK and five major commercial channels. The surveys look at programs of all genres using metadata, and weekday-evening national news & current affairs programs using coding analysis. For the latter, the coding team recorded the attributes of all those who spoke or were quoted, including their gender, age, occupation, whether they were named, what news item they appeared in, and in what role. This presentation will explain the methods and outcomes of the 2021 and 2022 surveys.

• 16:30 - Camille Guinaudeau, Japanese French Laboratory for Informatics, National Institute of Informatics
  Role indentification in TV news
  In the context of monitoring gender equality, estimating the speaking time and appearance time of men and women is a first step in assessing the balance between genders. However, these initial measures are not sufficient to precisely analyze the data and highlight potential inequalities in treatment between men and women (supporting role vs. leading role, expert vs. presenter role, etc.). In this context, identifying the role of men and women in television news allows for a more detailed analysis of their media presence, particularly in terms of expertise and dominance. In this presentation we will present our first experiments on Japanese and French data.

• 17:15 - Prof. Antoine Laurent, Laboratoire d'Informatique de l'Université du Mans, Le Mans University et Martin Lebourdais, Institut de Recherche en Informatique de Toulouse
  Speakers Interactions: From overlapped speech to interruption detection
  The ANR GEM project, initiated by the National Audiovisual Institute, aims to study the differences in treatment and representation between women and men in the media. This project encourages collaboration between research in media and language sciences and research in computer science. One of the project's objectives is to promote the creation of automated tools to generalize and facilitate social sciences and humanities studies on large corpora. In this presentation, we focused on signal processing tools that facilitate the characterization of speaker representations. Specifically, we proposed methods to automatically detect and characterize interruptions during conversations from television debate programs.

March 1st 

9:30 : Opening  -  Pr. Sandra Laugier, University Paris 1 Panthéon-Sorbonne

• 9:45 - Quention Gervasoni,  Université Paris 1 Panthéon-Sorbonne
  Routinizing hype : exploring how series shape the temporality of public online emotions
  This presentation is an exploratory analysis of the temporality of hype produced by series. I draw on the results of my PhD dissertation which conceptualizes hype as a central emotional regime (a concept borrowed and adapted from William Reddy's work) of today's online cultural reception. Since my dissertation is based on the case study of Pokémon online fans, this presentation aims to translate this framework to the study of series in order to further assess how the cultural industries and the broader sociotechnical context of online reception shapes public emotions. I rely on an exploratory analysis of a corpus built by scraping the #LBDL which relates to the French securitary series Le Bureau des Légendes. 

• 10:30 - Jérémy Poiroux,  Japanese French Laboratory for Informatics, The University of Tokyo
  Competition and collaboration: recommending music as an organisational activity.
  Our work is based on the premise that the music recommendation activity should no longer be analysed in exclusively cultural terms, at the risk of overlooking its organisational issues. This change of perspective makes it possible to reconsider the way in which certain values are instantiated in recommendations, especially when they are algorithmic. In a study based on a dozen interviews with Deezer employees, we show, for example, that exploration and diversification are not aims in themselves – even if they are high ideals – but are entirely appropriate means for recommending music and, by extension, getting people to consume it. This commercial objective is the driver of competition between recommendation modules, as well as collaboration between the actors involved in their design. Our main result highlights the importance of algorithms in organising both music recommendation (competition) and organisation (collaboration). 

• 11:15 - Agathe Chabrol, Université Paris 1 Panthéon-Sorbonne
  TV series recommandations: exploring AI solutions
  This presentation offers an update on our ongoing project to develop a TV series recommandation system using AI analysis. We'll discuss our progress, including the exploration of Large Language Models (LLMs) and the utilization of the OpenAI API to understand TV show characteristics. We're leveraging Betaseries' database and analyzing subtitles, audio descriptions, and synopsis to enhance recommandations. Additionally, we'll share preliminary user statistics. Our main focus remains on refining our system through diverse data and advanced algorithms to deliver personalized viewing suggestions.

Lunch break

• 14:00 - 18:00 - Working session  : RECO+ projet

Organizing committee 

• Dr. Camille Guinaudeau, Japanese French Laboratory for Informatics, National Institute of Informatics

• Pr. Philippe Codognet, Japanese French Laboratory for Informatics, Tokyo University

• Pr. Sandra Laugier, University Paris 1 Panthéon-Sorbonne

• Anastasia Krutikova, University Paris 1 Panthéon-Sorbonne