Introduction
Many common administrative tasks across government - from written communication and completing standardised routine forms, to maintaining documentation - consume large amounts of public sector workers’ time. Generative AI has the potential to help automate many of these tasks where appropriate and where its efficacy can be robustly evaluated. A recurring problem is the need to deal with large quantities of ‘free text’ data provided as input to government (defined as text created by users who are free to type in whatever they want, with little to no specified structure). We are exploring the extent to which the latest developments in text processing and analysis, driven by advances in large language models (LLMs), can aid with this type of government administrative task.
Applications
Government consultations
In our ongoing work with the Department for Transport we are exploring the extent to which advances in LLMs can aid with summarizing and analysing responses to government consultations, which may attract thousands of responses from citizens expressing their views on potential policy changes. Human summaries of consultation responses often necessarily remain at a high-level scope, and may miss out on granular detail particularly within consultations that receive a high level of responses. This is inevitably harder for a small team of individuals to capture in a systematic manner.
Free text
There are many potential use cases beyond consultations such as processing the result of policy evaluations, generating responses to citizen questions, or investigating large volumes of documentary evidence. Summarizing or extracting information from large quantities of free text responses in any of these contexts is a repetitive and difficult task which is hard to achieve with high accuracy. We are developing frameworks to evaluate how well generative AI is at performing these common tasks.
Organisers
Saba Esnaashari
Data Scientist
John Francis
Data Scientist
Youmna Hashem
Researcher
Anton Poletaev
Data ScientistPrevious contributors
