PhD defense of Noé Cécillon – 18 January 2024

Date: Thursday, January 18, 2023 at 14:00.

Place: University of Avignon, Campus JH Fabre, Ada Lovelace amphitheater

Jury :

  • Irina Illina, Maîtresse de Conférence, Université de Lorraine (Reviewer)
  • Julien Velcin, Professeur, Université Lyon 2 (Reviewer)
  • Serena Villata, Directrice de Recherche, Institut 3IA Côte d’Azur (Examinator)
  • Harold Mouchère, Professeur, Nantes Université (Examinator)
  • Vincent Labatut, Maître de Conférence, Avignon Université (Advisor)
  • Richard Dufour, Professeur, Nantes Université (Co-advisor)

Title: Combining Graph and Text to Model Conversations: An Application to Online Abuse Detection.

Abstract: Online abusive behaviors can have devastating consequences on individuals and communities. With the global expansion of internet and the social networks, anyone can be confronted with these behaviors. Over the past few years, laws and regulations have been established to regulate this kind of abuse but the responsibility ultimately lies with the platforms that host online communications. They are asked to monitor their users in order to prevent the proliferation of abusive content. Timely detection and moderation is a key factor to reduce the quantity and impact of abusive behaviors. However, due to the sheer quantity of online messages posted every day, platforms struggle to provide adequate resources. Since this implies high human and financial costs, companies have a keen interest in automating this process. Although it may seem a relatively simple task, it turns out to be quite complex. Indeed, malicious users have developed numerous techniques to bypass the standard automated methods. Allusions or implied meaning are other examples of strategies that automatic methods struggle to detect. While usually performed on individual messages taken out of their context, it has been shown that automatic abuse detection can benefit from considering the context in which the message was posted.

In this thesis, we want to focus on the combination of content and structure of conversations to tackle the abuse detection task. Using the textual content of messages is the standard approach which was first developed in the literature. It has the advantage of being easy to set up, but on the other hand, it is vulnerable to text-based attacks such as obfuscation. The structure of the conversation which represent the context is less frequently used as it is more complicated to manipulate. Yet it allows to introduce a contextual aspect which helps detecting abuse occurrences when the text on its own is not sufficient. This context can be modeled as a contextual graph representing the conversation which includes the message. By comparing two methods based on feature engineering on a dataset of conversations extracted from a video games, we could show that a method relying exclusively on conversational graphs and ignoring the content was able to obtain better detection performance. The literature suggest that combining multiple modalities often result in a better detection of abusive messages. We propose multiple strategies to combine the content and structure of conversations and prove that their combination is indeed beneficial to the detection.
A limitation of feature-based methods is that they are costly in time and computational resources. Our study also highlights that only a fraction of the computed features are truly relevant for the task. Representation learning methods can be used to mitigate these issues by automatically learning the representations of text and conversational graphs. For graphs, we demonstrated that using edge weights, signs and directions improved the performance. As no method exists for signed whole-graph embedding, we fill this gap in the literature by developing two such methods. We assess them on a newly constituted benchmark of three datasets of signed graphs and show that they perform better than their unsigned counterparts.
Lastly, we perform a comparative study of several lexical and graph-embedding method for abuse detection by applying them to our dataset of conversations. Our results show that they perform better than feature-based approaches on text and are slightly less effective on graphs. Still, they obtain promising results given that they are completely task independent, much more scalable and time-efficient than feature-based approaches.