Since the start of the 16th legislature of the Fifth French Republic in June 2022, no less than 85,345 amendments have been tabled. If only 5% of them are adopted, the related work is considerable. Each of the amendments submitted for examination by one of the two chambers is analyzed and processed by agents of the central administration working for the government, with the aim of facilitating the smooth running of discussions, debates and votes in Parliament .
In 2020, the State launched a platform called Signale (Interministerial digital management system for legislative amendments) which allows administrative agents to coordinate and work more efficiently on their tasks related to bills (in particular amendments). For example, it facilitates the writing of summary notes and the creation of tables for the preparation of interministerial meetings concerning legislative texts.
Optimizing the work of administrative agents on bills, this platform represented a considerable advance given the number of amendments which increases over time, often reaching thousands for a given bill.
Tedious work transformed by generative AI
Continuing its digital transformation work, the public service has therefore adopted generative AI to manage the work related to the processing of amendments. Thus, last fall, the digital transformation department (DTNum) implemented a solution which automatically assigns the parliamentary amendments received during the finance laws to the right teams, and is responsible for summarizing them. automatically by a generative AI solution.
The great baptized language model LTheMandement is based on Meta's open source LlaMa-2 70B model, and has been “subjected to fine-tuning by Government agents” via the low-rank adaptation (LORA) technique to refine the selected LLaMA model. This technique introduces adaptability into model learning, particularly beneficial for models like LLaMA, by inserting additional parameters.
Parameters that are specifically designed to adapt the model's response to the intricacies of legislative language without intensive training or significant alteration of its structure. For its training, the teams indicate having used 15,397 pairs of amendments and summaries, data from the Signale platform.
Neutral summaries and considerable time savings
With LlaMandement, the objective is to improve the efficiency of processing French parliamentary work (including the drafting of bench notes and preparatory work for interministerial meetings) through the production of neutral summaries of bills and proposed laws. According to Esther MacNamaradelegate for digital transformation (DTNum) of the general directorate of public finances (DGFIP) and attached to the Ministry of the Economy and Finance, “this work allows agents to concentrate on preparing the political part of the response by freeing up valuable time, in an emergency context.”
Thus, as part of the implementation of AI tools, a self-attribution rate of 94% was achieved for nearly 5,400 amendments in less than 10 minutes, an encouraging result demonstrating their potential for effectiveness and of accuracy.
The fruit of joint work by different administrations
This solution is the result of joint work by the Tax Legislation Directorate (DLF) – attached to the General Directorate of Public Finances (DGFiP) – as well as the data scientist teams of the Digital Transformation Directorate (DTNum) attached to the General Directorate of Public Finances and Dinum.
Selected for you