On September 28, 2023, Judicaël Poumay will publicly defend his thesis entitled:

"NLP Methods for Weak Signals Detection from Unstructured Text"

at 15:00 at HEC Liège, N1D building, Classroom 0/86

Jury members 

• Prof. Ashwin ITTOO, Supervisor (HEC Liège, Management School of the University of Liège)
• Prof. Pierre GEURTS (Faculty of Applied Sciences, University of Liège)
• Prof. Michael SCHYNS (HEC Liège, Management School of the University of Liège)
• Prof. Antal van den BOSCH (Utrecht University)
• Prof. Shankar VENKATAGIRI (Indian Institute of Management Bangalore)
• Prof. Le Minh NGUYEN (Japan Advanced Institute of Science and Technology)

Summary

In today's competitive and ever-evolving world, organisations must adapt to emerging challenges and opportunities. A crucial aspect of achieving this goal is environmental scanning, which involves continuously collecting, analysing, and interpreting external information to identify impactful elements. This thesis focuses on a subset of environmental monitoring: the detection of weak signals.

The explosion of unstructured data (articles, blogs, social networks, etc.) produced by humanity provides a wealth of exploitable information, but manual processing has become impossible given the existing volume. Natural Language Processing (NLP) offers a solution to this by using machine learning to understand and process human language. Advances in NLP have enabled applications such as automatic translation and sentiment analysis. More recently, technologies like ChatGPT, emerging from the field of NLP, are revolutionising the world.

Hence, NLP has the potential to offer solutions for detecting weak signals by analysing large volumes of textual data. With this in mind, this thesis explores and develops the possibility of using innovative NLP methods through topic extraction, combining hierarchical and temporal dimensions to extract subjects and events in a non-parametric manner.

Consequently, the detection of weak signals relies on identifying infrequent and growing subtopics. Additionally, by studying the correlation between the discovered topics, our model provides contextual information that facilitates strategic planning and scenario building.

The quality of the extracted topics is crucial. However, evaluating this type of model is challenging. Therefore, this thesis also focuses on studying past evaluation methods and proposes a new method that uses labelled data to assess the ability to extract known topics. This offers a quantitative evaluation of the quality of hierarchical topic models.