Antonin Descampe — Augmented Images (ULiège)

Antonin Descampe — Augmented Images (ULiège)

by Antonin Descampe · View the channel on PSI

Humanities, Social Sciences & Thought Computing & Cybersecurity UYComputer scienceUYWAugmented reality (AR)
🎙 Antonin Descampe 👥 178 📅 October 15, 2022 ⏱ 38 min 👁 44 📄 original study 🧭 2026-10-09
Available in: English (current) Français

Keywords

subjectivitymultimodalitynatural language processingnews framingtransformer models

Summary

The talk presents an interdisciplinary research project aimed at detecting subjectivity in multimodal news content using machine learning, with a focus on French-language Belgian press. The speaker, Antonin Descampe, outlines three main steps: first, evaluating linguistic subjectivity in written articles using both handcrafted features and language models, achieving high accuracy but differing interpretability. Second, using language models to analyze bias and stereotypes in news articles by examining word-vector spaces, showing how semantic contexts reveal media framing (e.g., vaccination coverage differences between Le Monde and Le Soir). Third, extending to automated news frame analysis using multimodal transformers, which integrate text and image processing. The talk discusses challenges such as the black-box nature of transformers, the need for explainability, and visual-specific properties (analogical quality, indexicality, lack of explicit propositional syntax). It concludes with future directions including human annotation and combining uncrafted features with deep learning for better performance and interpretability. The presentation is well-structured and accessible, highlighting the intersection of AI, linguistics, and media studies.

163 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into practical applications of machine learning for media analysis, bridging computational and social sciences. It demonstrates concrete results (e.g., 90% and 99% accuracy for handcrafted vs. transformer models) and proposes a hybrid approach to combine performance with explainability. The argumentation is coherent, moving from definitional issues to methodological steps, and it effectively illustrates conceptual points with examples (e.g., the ‘bank’ example, gender bias in job associations). The speaker candidly acknowledges limitations, such as the lack of explainability in transformers and the ongoing nature of the work, which strengthens credibility. The visual framing analysis is thoughtfully linked to real media examples, making the case for multimodal analysis compelling.

Scientific Rigor, Source Quality, Title Accuracy

The talk references established scholars and works (e.g., Entman’s framing theory, Rodriguez & Dimitrova’s visual framing levels, the Reuters Digital News Report, and a quote from Wesley Lowry). These sources lend theoretical grounding. However, no explicit URLs or detailed citations are provided, and the claims rely on the speaker’s descriptions. The title of the video is generic and does not reveal the specific content, which may reduce discoverability. The content itself is rigorous within the scope of an ongoing research presentation, but it is not peer-reviewed and should be treated as a preliminary report.

221 words

Title / Content Match

The video title names the speaker and the symposium, but the actual talk content is about subjectivity detection in multimodal news. It thus provides only generic information and may not immediately signal the specific topic.

Quality & Reliability

7/10

The talk presents original research in progress, grounded in established theories (framing, agenda-setting) and uses recognized NLP methods. The speaker is an academic and acknowledges the collaborative, ongoing nature of the work. However, it is a conference presentation without peer review, and some methodological details are left open for Q&A.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • Reuters Digital News Report — Cited to illustrate declining trust in news media.
  • Wesley Lowry on objectivity — Quoted to argue that objectivity is a myth and journalism should aim for fairness.
  • Entman's framing theory — Referenced when discussing how news frames influence audience perception.
  • Rodriguez & Dimitrova, visual framing levels — Used to categorize levels of visual framing in the multimodal analysis.

Contribution & Novelties

The talk contributes a concrete methodological synthesis for detecting subjectivity in news using both text and image, addressing the gap between high-performance black-box models and interpretable handcrafted features. It proposes a pipeline that combines these approaches for explainable and accurate analysis. The discussion on using language models as tools for bias analysis (rather than only debiasing) offers a novel perspective in social science research. The extension to multimodal transformers for visual framing analysis is an emerging area, and the speaker’s engineering background brings a practical angle.

Pour aller plus loin :

  • Transformer (machine learning) — The core architecture behind modern NLP and multimodal models; directly relevant to the talk’s discussion of transformers.
  • Multimodal learning — Overview of integrating multiple data modalities; essential for understanding the multimodal approach.
  • Visual framing — Rodriguez and Dimitrova’s levels; provides analytical depth for the visual analysis part.
  • Reuters Institute Digital News Report — Source of trust and media consumption data; useful for contextualizing the problem.
  • Word embedding — The basis for bias detection via vector space; explains the ‘king - man + woman = queen’ example.

181 words

Radar Profile

The radar profile shows high scores in quantity and technical level, reflecting the dense, expert presentation. Quality of information and reliability are slightly lower, consistent with an interim research talk that lacks peer-reviewed validation. The overall shape suggests a technically robust yet not fully vetted contribution.

Reliability 7/10