Solutions · Social research Analyze open-ended research data systematically and review it with subject-matter expertise.
Work with your own categories, examine the original responses and document how you classified them.
What concerns respondents? The reasons are in the open-ended responses.
By the thousands, in several languages, wave after wave. And every category has to stand up in the methods report.
The challenge: Coding, one response at a time
Open-ended responses are coded by hand. In large surveys, there is often only time for a subsample.
With deepsight: Your codebook for every response
Import your codebook with descriptions and parent categories as an Excel (.xlsx) or CSV file. Topic analysis applies it to all responses rather than a subsample – in minutes rather than weeks.
Codebooks in topic analysisThe challenge: Waves that don’t line up
Coders change, categories get reinterpreted. Changes between waves then become hard to demonstrate.
With deepsight: One category system across all waves
Save your codebook – or the topics from the first wave as a codebook – and apply it in every following wave. You carry over the study context and glossary as a versioned analysis profile, so every wave starts from the same settings.
Go to analysis profilesThe challenge: Personal details in free text
Open-ended responses contain names, addresses or clues to individual people – even when the questionnaire doesn’t ask for them.
With deepsight: Masked before the analysis
Our own anonymization masks details such as names, email addresses or phone numbers without a language model before the analyses run – with translation switched on, only after translating. It doesn’t catch every detail or indirect clues; check results before sharing them and mask anything missed by hand.
Go to anonymizationThe challenge: Separate codes for each language
In multilingual studies, each language team often codes on its own. The categories drift apart, and comparisons become unreliable.
With deepsight: One category system for all languages
Analyze all language versions with the same codebook. A glossary in the study context helps translate technical terms consistently, and the original stays available.
Go to translation
Classifying a submission
A submission from a public consultation on redesigning a town-center street.
“I welcome the new bike lane on the main street. But if the parking spaces in front of the shops disappear as a result, many older customers will stop coming.”
- Suggestion
- Documentation
The added category is saved as a new topic version. Earlier versions can be restored.
Select a category or the review decision to see the passage.
Illustrative example.
What you define in your project
What open-ended responses can tell you
- What do respondents name as the most important problem – and how does that shift from wave to wave?
- Which concerns and objections do residents raise in a public consultation?
- How do respondents explain their stance on a policy measure?
- What lies behind “Other, please specify …” in semi-open questions?
- Do younger and older people, urban and rural respondents raise different topics?
- What experiences do participants describe in the evaluation of a program?
What you start with
The Excel or CSV export from your survey software or consultation process: one row per case, with open-ended responses and attributes in their own columns.
From raw file to documented results
- Prepare the data
Upload the file and add the research question and target group to the study context. Filter out empty or content-free responses such as “n/a” up front.
Data quality check - Assign categories
Apply your codebook strictly or have additional topics suggested – deductive, inductive or both. No training required.
Topic analysis - Review and refine
Correct assignments, merge or split categories. On request, deepsight searches the residual “other” category again for additional topics.
Topic refinement - Compare groups
Compare categories by attribute, such as wave or level of education, with significance indicators if you wish.
Group comparison - Document
Record the codebook version and your review decisions for the methods report.
Methodology
A new AI run can differ – your reviewed results stay saved.
Report from the reviewed results, and review every new run like the first one.
What goes into your report and dataset
Good to know: Do you produce your tabulations in another program? The slide factory turns Excel tables into slides based on your PowerPoint template.
Behind every percentage are the responses it was built from.
So you can back up every category – in the methods report, in peer review and with clients.
What matters with open-ended responses
Define categories, don’t just name them
A short description for each category sets out what belongs in it and what doesn’t. The descriptions in your codebook feed into the assignment; fuzzy boundaries show up as misassignments when you review.
Compare with hand coding
Code a sample independently by hand and compare it with the exported assignments. deepsight doesn’t output an intercoder reliability coefficient; you can calculate one yourself if needed. The discrepancies show which categories need sharpening.
Base and multiple mentions
Percentages and key figures refer to the analyzed responses in each group; anything the data quality check filters out doesn’t count toward the base. One response can be assigned to several categories, so totals can exceed 100%. The base limits are fixed: below 10 responses, values are flagged or hidden in group comparison and dashboards, and significance indicators appear from 30 responses per group.
Put differences in context
Significance indicators compare one group at a time with the rest of the sample. They show where to look more closely, not why a difference arises. Whether results can be generalized to the population depends on your sample and design.
Questions about social research
If the wave is an attribute in your data, compare the categories by wave in the group comparison. Read significance indicators for panel data with care: the test treats the rows as independent cases, even though the same people appear more than once. For recurring studies there is also wave comparison; it is included in every paid plan – see the plan comparison.
Yes. Topic shares and group comparisons can be weighted – with a weight column from your data or to target distributions (rim weighting). Weights do not automatically make a sample representative; whether results can be generalized still depends on your sample and design. Weighting is included in every paid plan – see the pricing page.
Describe the codebook and how you applied it (strictly or with additional suggested topics), along with exclusions, language approach, review and the categories you changed. The methodology shows which records deepsight keeps for this.
Only to a limited extent: deepsight cloud analyzes text row by row, one response or submission per row. For coding qualitative interviews, we develop custom solutions – such as Qualytis, built for Produkt + Markt and not part of deepsight cloud. We work out together what fits your project: Custom projects.
deepsight cloud runs on Microsoft Azure in the EU. For AI features we use Azure OpenAI; Microsoft does not use your data to train AI models. Processing is governed by a data processing agreement under Art. 28 GDPR. DeepL or Google are only used if you choose them for translation. If the data must not leave your organization, we look into a solution with local language models as a custom project. More under Hosting & locations.
Related to social research
Discuss your research project
We will show you the analysis with your own category system using an example.