Consultation Response Clustering
Consultation Response Clustering groups written responses by the substance of the point they make, to help analysts summarise a large consultation fairly.
Keywords
- Tags: Clustering
- Tags: Consultation
- Tags: Large language model
- Tags: Transparency
- Stage: Pilot
- Type: Embedding clustering with generated cluster labels
- Sector: Public administration
- Language: English
How does our product work?
Responses are grouped by similarity of argument, and each group is given a proposed label describing the point it makes. An analyst reviews and rewrites the labels and checks a sample of each group.
Cluster size is reported but is explicitly not treated as a vote. A point made once may still be decisive.
Overview
A consultation attracting several thousand written responses previously had to be summarised by reading a sample, which systematically favoured whichever arguments happened to be most common.
Clustering makes it possible to see the full range of distinct arguments, including the rare ones, and to state honestly how many people made each.
Owner and responsibility
More detailed information on the system
Here you can get acquainted with the information used by the system, the operating logic, and its governance in the areas that interest you.
System description
Generated cluster labels are treated as a starting point and are always rewritten by an analyst before publication, because a label written by a model is a summary no one is accountable for.
The published analysis names the method and includes the number of responses reviewed individually.
Known limitations
Clustering performs poorly on responses that make several unrelated points in one submission; these are separated manually during review. Non-English responses are excluded from clustering and are analysed by a translator.
References
Consultation analysis method note, published with each consultation report.