Short answer
Issue clustering converts each submission into a representation of its meaning, groups similar ones into themes, and arranges the themes into a tree you can drill into down to the original text. It identifies form-letter campaigns as campaigns and surfaces the unusual submission nobody else raised. Used well, it makes a consultation report more defensible: every count traces to listable submissions, and people make and record the decisions about boundaries, names and conclusions.
Every engagement professional who has run a major consultation knows the arithmetic of the submission mountain. An environmental impact statement exhibition can attract thousands of responses. Each must be read, categorised and reflected in a report that the community, the assessing agency and sometimes a court will scrutinise. The manual version is slow and, less discussed, inconsistent: the same submission categorised by two analysts, or by one analyst on Monday and Friday, can land in different buckets.
How does issue clustering work, in plain language?
Text is converted into a mathematical representation, called an embedding, in which meaning determines proximity. Two submissions worried about construction noise near a school sit close together even if one says “unacceptable acoustic impacts on educational facilities” and the other says “the kids won’t be able to hear their teacher”. The embedding captures what is meant, not just which words were used, which is where keyword approaches fall down.
Once every submission is a point in that space, clustering algorithms find the natural groupings, which become candidate themes. Good systems then build a theme tree, so “noise” branches into construction noise, operational noise and night works, each drillable to the submissions in it. That traceability matters most: when the report says 412 submissions raised construction traffic, the 412 can be listed. That is a stronger evidentiary position than a manual tally.
How should clustering handle form-letter campaigns?
By counting the campaign honestly as one theme with many participants, and keeping the personal additions. Organised campaigns are a legitimate part of participation, and they distort naive analysis in two directions: five hundred near-identical submissions can read as five hundred independent concerns, or be dismissed wholesale, erasing real sentiment and the personal paragraphs people often add. Near-duplicate detection identifies the campaign, reports it as exactly that, and separates the personalised text so it is analysed on its own merits.
What good clustering output looks like
- A theme tree, not a flat list: high-level issues branching into specific concerns.
- Every theme drillable to the verbatim submissions in it.
- Campaigns identified and counted as campaigns, with personal content preserved.
- Outliers surfaced, not discarded: the one submission raising a groundwater issue nobody else saw.
- Reproducible counts and a record of the human decisions made on top of them.
Where does a person need to stay in the loop?
At four points, and each decision should be recorded:
- Boundaries. A submission about “trucks waking the street at 5am” belongs to noise, traffic or both. An analyst should decide how the report treats overlaps.
- Names. The label for a cluster becomes the language of the report and of the project’s response. Machines can propose names; people should own them.
- Outliers and small clusters. The point of surfacing them is that someone qualified reads them in full.
- Synthesis. What the pattern means and what should change in response was never the machine’s job.
Framed this way, the division of labour is clean. The machine does the parts where consistency and scale matter: reading everything, grouping by meaning, counting reproducibly. People do the parts where judgement and accountability matter. The result is much faster than a fully manual process and easier to defend, because the method is explicit and the working can be shown, a property government AI policies increasingly expect, as we discuss in our NSW AI Assessment Framework piece.
What makes the resulting consultation report defensible?
Traceability from every statement in the report back to the record. For each theme, keep the list of submissions, the human decisions about boundaries and naming, and how the project responded. Where a submission raised something the project acted on, link the submission to the change. That chain, from what people said to what was done about it, is what a regulator, a community member reading the report, or a court will look for. Our guide to writing a defensible consultation report from thousands of submissions goes further, and the wider set of AI capabilities is mapped in our guide to AI in stakeholder engagement.
Where VivoAI fits
VivoAI clusters submissions into a theme tree you can drill into, from the high-level issue down to individual submissions, with campaigns identified and unique submissions kept. Themes, sentiment and classifications carry a confidence level, AI-assisted decisions keep an audit trail, and inference runs in the AWS Sydney region. Your team reviews and names the themes before anything is relied on.
See how the AI keeps the record.
Frequently asked questions
What is automated issue clustering?
It is a machine learning technique that groups free-text feedback, such as consultation submissions or complaints, by meaning rather than by keyword. Similar concerns land in the same theme even when they are worded differently, and each theme can be traced back to the submissions in it.
Is AI clustering accurate enough for a consultation report?
It is a strong first pass, not a final answer. It reads everything consistently, but boundaries between themes, theme names, outliers and the overall conclusions need a person. A report is defensible when every count traces to listable submissions and the human decisions about boundaries and naming are recorded.
How should form-letter campaigns be counted?
Report the campaign honestly as one theme carried by many participants, and separate any personal text people added to the template so it is analysed on its own merits. Neither inflating a campaign into hundreds of independent concerns nor dismissing it is fair analysis.
Does using AI to analyse submissions need an AI assessment?
For government agencies it may. Commonwealth agencies follow the Policy for the responsible use of AI in government and NSW agencies apply the NSW AI Assessment Framework, both of which ask agencies to assess and take accountability for AI use cases. Check which framework applies to you and what it requires.
Sources
- Policy for the responsible use of AI in government (Digital Transformation Agency)
- NSW AI Assessment Framework (Digital NSW)
Next step
Bring us your submission mountain
See a consultation dataset become a theme tree that traces every count to the submissions behind it.