Abstract
Group messaging has become one of the main places where friendships are maintained and, incidentally, recorded. This article examines a private, French-language group chat spanning nine years and some 208,000 messages, and asks whether everyday conversation (the apparently trivial exchanges usually dismissed as small talk) carries the structure of a collective memory. Three complementary computational approaches are combined. Neural topic modelling (BERTopic) identifies recurrent themes and traces their evolution over time; a multilingual transformer-based sentiment classifier, suited to informal chat language, follows shifts in conversational tone; and directed reaction networks model who responds to whom, describing the interactional structure of the group. Read together, these signals offer a longitudinal account of how a small community's shared references emerge, persist and fade.
Research questions
- Which themes structure the conversation, and how do they appear, recur and disappear over nine years?
- How does the emotional tone of the exchanges evolve, and does it vary with the themes discussed?
- What interactional structure emerges from reactions between participants, and how stable is it over time?
Data
The corpus is the complete export of a single Messenger group chat between friends, covering nine years and approximately 208,000 messages, written mostly in informal French. Messages were ordered chronologically and segmented into periods so that themes, tone and interactions can be compared over time.
Privacy statement. The corpus is private and is not distributed. This page presents the research questions and methods only; no message, participant name, excerpt or individual-level result is reproduced here.
Methods
Topic modelling
Messages are embedded with a transformer model and clustered with BERTopic (Grootendorst, 2022), which describes each cluster with class-based TF-IDF terms. Topic frequencies are then computed per period to follow how themes rise, recur and decline.
Sentiment analysis
Each message is labelled negative, neutral or positive by a multilingual transformer classifier trained on social-media text (Barbieri et al., 2022), chosen for its robustness to the abbreviations, slang and emoji typical of chat language. Scores are aggregated by period and by topic.
Reaction networks
Reactions and replies are modelled as a weighted, directed graph in which an edge from A to B counts how often A responds to B. Standard network measures describe the centrality of participants and how the structure of the group changes across periods.
Scope
The study is a longitudinal case study of one small community and does not claim to generalise to all group conversations. Its contribution is methodological as much as descriptive: it shows how established NLP and network tools can be combined to read informal, long-running conversation as a record of shared memory.
References
Barbieri, F., Espinosa Anke, L., & Camacho-Collados, J. (2022). XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond. Proceedings of the Thirteenth Language Resources and Evaluation Conference (LREC 2022), 258–266.
Grootendorst, M. (2022). BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint arXiv:2203.05794.
How to cite
[SURNAME], T. ([YEAR]). Du bavardage à la mémoire collective. [WHERE PUBLISHED].
The full text is not publicly available. For questions about the methods, get in touch.