L4QAUCT HONOURSSelect
Choose a bar, dot, node, edge, cell or word in either dashboard.
Contribution 03 / Visualisations
Every finding should have a way back to the source.
Linked dashboards connect analytical patterns to source text, supporting evidence inspection and human validation.
Research focus
How can linked visualisations help researchers interpret, trace and validate LLM-generated analysis?
Context & contribution
A chart can reveal a pattern without explaining why it exists. The visualisation contribution links results across views and resolves selections into their supporting material, so researchers can investigate a result in context.
Built with Dash, Plotly and PyVis, the dashboards connect code overviews, coverage and co-occurrence views with sentiment and aspect views. They display available evidence, justifications, audit critiques and human edits, without making additional LLM calls or inventing confidence scores.
The approach
How it works.
Filter
The shared selection updates relevant linked views while the overview remains visible.
Resolve
Grounded Theory selections resolve to supporting quotes; sentiment selections resolve to source documents.
Inspect
Review evidence and justifications, plus approval or critique details when available. Continue exploring with another selection.
One interaction pattern across two dashboards: select, filter, resolve and inspect, with an exploration loop back to selection.
Open full-size diagramEvaluation
What the results show.
Three expert reviewers informed refinements, followed by a task-based study with five research-literate participants.
User study · n = 5Expert reviewers · n = 3
Usability · SUS
80.5 / 100
Mean reported score
Workload · Raw NASA-TLX
19.5 / 100
Mean reported workload
Task completion
96%
48 of 50 tasks completed
Error-spotting confidence
2.8 / 5
Mean reported confidence
Scores are means. Task completion covers 50 tasks.
Reading the resultsUsers valued the network view, consistent colours and access to evidence. The code-coverage grid and critique markers remained difficult to interpret; confidence in spotting errors averaged 2.8/5.
Source: Theo Mashaba’s final paper (PDF), Table 3 and Sections 4–6.
03 / Follow the evidence
A finding is a starting point.
Select a topic to see how a view can connect an interpretation with the words that support it.
Linked evidence explorerIllustrative example · invented excerpts
Select a topic
The selection changes the linked excerpt and interpretation together.
“The first steps were clear, but I still wanted an example before coding on my own.”
Suggested interpretation
Clear instructions can coexist with a need for guided practice. The source contains both views.
This small interaction demonstrates the linking concept. It is not the research application and does not run an LLM.
Scope & limitations
What to keep in mind.
- Five participants used one dataset in one session, with no control condition. These formative findings do not establish improved analytical accuracy.
- Error-spotting confidence is self-reported, not a measure of detection accuracy. The workload score also does not demonstrate a reduction against another tool.
- Evidence access depends on the engine output: Grounded Theory resolves to quote excerpts; sentiment resolves to whole documents because text offsets are unavailable.
More about the evaluation
Expert review consolidated 33 heuristic violations into 13 problems. User tasks covered exploration, targeted comprehension and interpretation, followed by questionnaires and interviews. SUS averaged 80.5 (SD 14.2); RawTLX averaged 19.5 (SD 5.9). Both evaluations highlighted the code-coverage grid and critique visibility.
Supporting material
All project resourcesRead the final paper.
Visualisations final paper
Theo Mashaba · Visualisations engine.