Production Case Study · Data Science & Social Research
Q-Mvutu: Conversational Survey Architecture & Lusophone Sentiment Analytics
Impact Summary: Named after the Kikongo word mvutu (“answer / response”), Q-Mvutu replaces static 40-question web forms with adaptive conversational interviews and real-time aspect-based sentiment pipelines — lifting survey completion rates to 92.5% and increasing respondent engagement by +150%.
Contents
1. The problem: Survey fatigue, shallow checkbox data and delayed reporting
Organisations conducting market research, civic feedback, or service-quality audits face a persistent dilemma. Multiple-choice radio-button grids are easy to tabulate in a spreadsheet, but they suffer from high abandonment rates on mobile screens and force respondents into pre-written boxes that miss the real root cause of dissatisfaction. Conversely, open-ended text boxes yield rich qualitative detail, but traditional research teams take weeks to read, code, and categorize thousands of free-text answers manually.
When we set out to engineer Q-Mvutu, our goal was to combine the statistical rigour of structured questionnaires with the depth of a human interviewer — while delivering classified sentiment and theme distributions in under `1.2 seconds` per completed response.
2. Adaptive inquiry routing and clarifying follow-up probes
In a static survey form, if a respondent selects “Dissatisfied” on question 4 and writes three vague words (“demora muito tempo”), the form blindly moves on to question 5. The researcher never learns which step took too long or at which branch it occurred.
Q-Mvutu models each study as a directed inquiry graph with bounded conversational follow-up probes:
- Deterministic core spine: Every respondent answers the mandatory demographic and baseline Likert-scale anchors identically, preserving statistical comparability across waves.
- Bounded clarifying probes: When an open-ended response is ambiguous or high-salience (for example, reporting a service failure without specifying the channel), a lightweight classification node triggers a single targeted follow-up question in natural Portuguese before advancing to the next core item.
- Mobile-first state persistence: Built on our low-bandwidth engineering standards, every response step is saved locally in `IndexedDB` and synced idempotently so respondents on unstable mobile connections never lose progress.
3. Real-time aspect sentiment and theme distribution pipeline
As responses stream into Q-Mvutu, they pass through our Lusophone natural-language processing pipeline (documented in detail in Beyond Keyword Counting: Sentiment & Theme Classification):
| Telemetry Stage | Technical Mechanism | Output Delivered to Researcher |
|---|---|---|
| 1. Normalisation & Language ID | Diacritic-tolerant subword tokenizer (`pt-AO`, `pt-PT`, `en`) | Preserves informal idioms and mobile shorthand without data loss |
| 2. Aspect-Based Sentiment (ABSA) | Fine-tuned transformer extracting `(aspect, polarity, urgency)` | Separates praise for staff from complaints about pricing or uptime |
| 3. Emergent Theme Clustering | Dense embeddings + HDBSCAN density clustering | Automatically surfaces unanticipated issues outside fixed categories |
| 4. Live Telemetry Dashboard | Incremental aggregation with confidence intervals | End-to-end latency `< 1.2s` from submission to chart update |
4. Response quality guardrails and duplicate/bot filtering
Public and incentivised online surveys are frequent targets for automated form-stuffing, copy-pasted gibberish, and straight-lining (clicking the first option on every row without reading). Q-Mvutu protects dataset integrity without invasive tracking:
- Interaction cadence & entropy checks: Responses completed faster than human reading speed or containing repetitive low-entropy character sequences are quarantined into a separate review bucket rather than polluting the primary distribution.
- Semantic coherence verification: If a respondent selects the highest possible satisfaction rating (`5/5`) but writes an open-ended comment containing severe negative sentiment, the record is flagged for analyst inspection.
5. Deployment results and decision-maker telemetry
Across live research campaigns deployed on Q-Mvutu, organisations recorded:
- 92.5% completed response rate among participants who initiated a conversational flow, compared to `54%–62%` on legacy multi-page matrix forms.
- < 1.2s real-time sentiment classification, eliminating the multi-week delay between closing field collection and presenting executive findings.
- +150% increase in actionable qualitative detail (measured by distinct aspect-opinion pairs extracted per respondent).
Visit the platform at q-mvutu.ticnes.ai, or contact our data science team to design a custom research or customer-intelligence deployment.