Skip to main content

Introduction

Planning Applications hold a wealth of information that can tell us about the current state of development in the UK. From the types of developments proposed (or previously constructed/ already underway), to the locations which are attracting interest (or not). Access to this information could help better inform decision making (think future applications), however, it is locked up in lengthy documents, each of which with different formatting and language. This project employs an AI-enabled solution to extract information from Planning Applications in a uniform manner – summarising length documents into a manageable dataset.

Challenges

The biggest difficultly faced when trying to ascertain the types and quantity of development taking place in the UK is the disaggregation of Planning Applications. Because each of the 326 Local Planning Authorities in the UK is individually responsible for planning/ development in their area, this creates 326 separate stores of information, each of which is accessible at a different location and with varying degrees of sophistication. Furthermore, the wide-ranging formatting and language-use within the documents (which are written by individual Planning Officers), leads to inconsistent use of terminology etc. Fortunately, a well-trained Large Language Model (LLM), is able to ‘read’ through thousands of documents (of which there are more than 1,000 added per day) much faster than a human and create a standardised output.

Solutions

To overcome the issue(s) of formatting and language, we employed a Large Language Model (LLM) to ‘read’ through thousands of Planning Applications and create a standard view of each document. This process generates an easy-to-analyse database of all Planning Applications in the UK, allowing users to identify micro and macro trends in development. A LLM is the technology which underpins popular ‘chatbots’ such as Open AI’s ChatGPT or Microsoft Copilot. Through our partnership with Microsoft, we are able to access the model on which ChatGPT runs. Furthermore, we are able to add to the model’s ‘prebuilt’ knowledge with specialist learning relating to the nuance of the Planning System – effectively creating a virtual expert in the matter.

Results

By carefully specifying the desired information, we are able to generate a standardised database of Planning Applications from previously unstructured and inconsistent documents. The resultant database features information pertaining to an Application, such as reference numbers, dates, parties involved etc., but also a summary of the longer prose from the document. For example, if a Planning Officer gives a detailed breakdown of the reasons behind their decision, we can categorise the reason into a word or two for easier analysis. The well-curated database would take thousands of hours for a human to collate manually (if it was at all possible), and with more than 1,000 new Applications submitted each day on average, any effort would quickly become obsolete. Instead, our AI-enabled solution is able to ingest documents, quickly retrieve the desired information and reflect that back in the database almost instantly (1 document ≈ 10 seconds to ingest).

An innovation-led social enterprise

The directory is brought to you by the Digital Task Force for Planning, a not-for-profit organisation. Our ambition is to promote digital integration and advancement in Spatial Planning to tackle the grand challenges in the 21st Century.
Brought to you by:
Seed funded by: