Umfrage-Analyse-System¶
Umfrage-Analyse-System is a web-based application for the structured evaluation of survey data at universities and research institutions. The application processes surveys with mixed question types — from single choice to free text — and turns raw data into reproducible reports. A defining feature is the multi-stage, LLM-supported pipeline that chains translation, item extraction, thematic clustering and textual summaries. Intermediate results are persisted, so every processing step remains traceable and can be corrected manually.
At a glance¶
- Import and clean survey exports including multilingual responses and LimeSurvey structure files
- Convert free-text responses into thematic clusters without manual coding, with the option to rework them
- Segment evaluations by country, institution type and size — including for free-text clusters
- Have statistical significance and effect size computed automatically
- Generate structured reports as Word documents with charts, detail tables and cluster examples
- Provide data for an accompanying dashboard website in a machine-readable format
- Generate evaluation texts (description, interpretation, segment comparison) per question automatically
Highlights¶
In contrast to a direct LLM prompt or an ad-hoc script, the application captures the entire evaluation process as a persisted, multi-stage pipeline. Every stage can be resumed, reviewed and corrected manually — essential for the reliability of the results.
- Seven question-type handlers in one tool. Single choice, multiple choice, yes/no matrices, Likert scales, rankings, free text and cooperation matrices are processed via a common registry, with question-type-specific aggregation and visualisation.
- Workbench for cluster quality assurance. Cluster names, descriptions and assignments can be edited after the fact; items can be renamed, merged, split or reassigned. The data can subsequently be re-clustered against the corrected categories.
- Multi-stage LLM pipeline instead of single prompts. Translation, item extraction (a hybrid of rule-based and LLM-supported methods), clustering with dynamic min/max thresholds, cluster example selection and per-question summaries are separate, parametrised stages with their own persistence.
- Segmented evaluation also for free text. The breakdown by country, institution type and size is performed not only for closed questions but also for free-text clusters — including comparison tables per cluster.
- Statistical grounding. Chi-square tests, Cramér's V as effect size and a correlation analysis between questions are integrated into the evaluation and reporting flow; sample warnings flag critical segment sizes.
- Automatically generated evaluation texts. For each question, three text blocks — distribution description, interpretation of notable findings and a note on significant segment differences — are produced in three languages and cached in the database.
- Multilingual output with a domain glossary. Reports and charts are produced in German, French and English; paragraph-wise translations use a configurable glossary for consistent terminology.
- Accompanying website via dashboard export. All analysis results including segmented data, correlations and theme assignments are exported as JSON and can be read by a separate dashboard application.
- Connection to three sources and services. Tabular survey exports (CSV, Excel), LimeSurvey structure files (.lss) and an OpenAI-compatible LLM API.
- Resumable batch processing. Long pipeline runs can be continued after an interruption; previously computed analyses, translations and clusters are loaded from the database rather than recreated.