An interdisciplinary exploration of the Constitutional Council’s decisions

How did the idea for the Isovote project come about?

Samuel Ferey: It was initially motivated by some rare materials that became available. Thanks to a 2008 law on the dissemination of the Constitutional Council’s deliberations, we were able to make use of all the minutes from 1959 to 1995. Then, we exploited the power of NLP (natural language processing) tools, combined with AI. Everything came together for an unprecedented project of high scientific quality.

Maxime Amblard: Up until that point, researchers studying the Constitutional Council could only focus on a few deliberations at a time, due to a lack of resources. Our major innovation was to undertake exhaustive and mostly automated analysis, to establish trends across multiple years or decades. This approach drew colleagues from a range of backgrounds: computer science, of course, but also economics, public law, history, political sciences and political philosophy. This interdisciplinarity enabled us to better understand how the Constitutional Council has accompanied the French government’s transformations since the beginning of the 5^th Republic, by combining our insights.  

Concretely, how did you organise the work within the team?

S. F.: At the start, we had over 300 minutes in PDF format. The challenge was to convert them into a corpus that could be critically examined for the needs of all the disciplines involved. So we reflected together, as specialists in computer science and humanities, about its characteristics: what, precisely, did we want to explore? What proportion of errors could we accept in the final texts? How could we structure the content of the minutes to differentiate between formal language, reading the preliminary reports, dialogue between councillors, votes, etc.?

M. A.: As early as 2021, there were powerful software programs to convert PDFs to text, but the error rate was too high to produce a usable corpus. Going from an error rate of 10% to 0.5% took time, especially since there was a lot of structuring work involved. We got there by combining NLP tools, AI tools that were directly derived from NLP, and manual corrections for certain tasks: for example, reassigning certain fragments of dialogue to the right councillor, which can change the entire meaning of a debate.

What does having this standardised and structured corpus available bring to the table?

M. A.: It's the fundamental result, the base that will allow our colleagues to carry out research in their respective disciplines. In my area, NLP, it is rare to achieve such a technical corpus – full of legal references – made up of dialogue, i.e. non-formatted interactions, non-verbal annotations, etc. I had to use specific methods to get there. As for AI, it made it possible to automate a lot of tasks and is now used to evaluate the corpus, for example, to quantify the frequency of words or arguments, analyse similarities or detect long-term shifts.  

At this point, has there already been research using the corpus? 

S. F.: The team members first evaluated whether councillors’ votes were influenced by the political affiliation of the person that nominated them. They concluded that it was not the case: the deliberation process of the Constitutional Council inherently limits the possibility of partisan votes. This work will be published by Springer. In parallel, we have published a more complex analysis on voting paradoxes, to see whether Constitutional Council votes are affected by the doctrinal paradox. Lastly, our tools and corpus are starting to be used by other researchers. For example, there’s the work by Dominique Rémy-Granger, Les Mots de Robert Badinter, which will come out in 2025.

M. A.: Beyond these initial results, there are many articles in progress in order to better understand the deliberation process. In particular, how are the decisions of the Constitutional Council established, step by step? What kind of legal arguments are used? Which councillors take the floor most? Do they form coalitions? What is the influence of the constitutional principle of equality between citizens in issuing decisions? Eventually, other teams of researchers will be able to use the corpus when it is made available.

What do you already see as the impact of Isovote?

S. F.: The project sparked a great deal of interest at the Constitutional Council, which has been highly engaged in sharing its internal documents with the public since the presidency of Robert Badinter (1986-1995) The Council is particularly interested in the role of AI in constitutional law, in order to understand the opportunities and limitations; we will discuss this subject with its departments in autumn. 

M. A.: Isovote was an opportunity for researchers from a wide range of disciplines to work together and learn from each other’s research. The project also allowed us to improve Tacteo, our collaborative platform for transcription and corpus annotation using digital sources. It’s a step forward for humanities and social sciences researchers, who often have access to documents hundreds of thousands of words long, or even more, but who are not always equipped to use such documents in their raw state. Furthermore, Tacteo guarantees strict compliance with research data regulations, which is not necessarily the case with proprietary software.  

* ANR-21-CE41-0008, Isovote 2021-2025.