Free open-ended survey response analyzer
Paste open-text answers, or choose one column of a CSV file. An AI model that runs in your browser groups similar answers into themes, also across languages. See the count, the percentage and example answers of each theme, and download the result as a CSV file. No signup and no email address.
Your answers stay on this device. When you start the first analysis, your browser downloads the model files (about 135 MB) from Hugging Face and the program files from jsDelivr. These downloads contain no answer text. Your browser keeps a copy of the files for the next analysis.
Put one answer on each line. Empty lines are skipped. To keep line breaks in an answer, use a CSV file. Up to 2,000 answers.
Your browser reads the file on this device. Save the file as CSV UTF-8. Values can be separated by commas, semicolons or tabs. A value in quotes can contain commas, line breaks and quotes written as two quotes (""). Up to 2,000 answers.
You can change this setting and analyze again. For the same answers, the model does not read the answers again.
Correct this, then analyze again:
Draft themes
These themes are a draft. The model groups answers by similar meaning, and it can make mistakes. Read the answers in each theme before you use the numbers.
One row for each answer, with its theme. If a cell starts with =, +, - or @, the file adds an apostrophe at its start, so that a spreadsheet does not run it as a formula.
What this tool does
An open-ended question gives answers in the words of each person, so you cannot count the answers directly. The usual method is coding: you read each answer, tag it with a theme, and count the tags. Our guide to qualitative vs quantitative survey questions explains the manual method step by step. This tool makes a first draft of the themes, so that you start with groups of answers and not with a long list.
How the tool groups answers
The model converts each answer into a list of numbers, called an embedding. Answers with a similar meaning get similar embeddings, also when they are in different languages. The model is paraphrase-multilingual-MiniLM-L12-v2. Its model card says that it can be used for tasks like clustering, and its language list has 49 languages. The model runs in your browser with Transformers.js.
Then the tool does these steps:
- It calculates the cosine similarity of each pair of answers. Two answers are neighbours if their similarity is at least the value of your setting: 0.7 for “very similar”, 0.6 for “similar” and 0.55 for “loosely similar”.
- The answer with the most neighbours starts the first theme, together with its neighbours. The tool does this again with the answers that are not in a theme. A group that is smaller than the minimum size does not become a theme.
- If the centres of two themes have a similarity of at least the same value, the two themes become one theme.
- An answer that is not in a theme joins the theme with the most similar centre, if the similarity is at least the same value. All other answers go to “Other”.
- The most typical answer of a theme is the answer that is nearest to its centre. The example answers are the next nearest answers.
Each answer is in exactly one theme or in “Other”. So the percentages of all groups add up to 100%, with rounding. The minimum size of a theme is 2 answers, or 1% of all answers if that is more. The same answers with the same setting always give the same result. The setting values and the minimum size are choices of this tool, and we tested them with invented answers only.
The tool does not write names for the themes. It shows the most typical answer of each theme. Read the answers, and write a name for each theme yourself.
Example: answers in English, German and Japanese
We tested the tool with 36 invented answers to the question “What almost stopped you from signing up?”: 12 in English, 12 in German and 12 in Japanese, about price, setup, data privacy and “nothing”. With the default setting, one theme had all 9 answers about price, 3 in each language. Two of them were “The price is too high for us.” and “私たちには価格が高すぎます。”. The answers about the other three meanings were in two themes each, and no theme had answers with different meanings. So the model can put answers with the same meaning in one theme across languages, but it can also split one meaning into two themes.
Limits
- The themes are a draft. A person must read the answers in each theme and correct the themes before the result goes in a report or supports a decision.
- The model can put answers with different meanings in one theme, and it can split answers with the same meaning into two themes.
- Each answer gets one theme only. An answer with two ideas is in the theme of one of them.
- The model reads only the first 128 tokens of each answer. A token is a word or a part of a word. The rest of a long answer does not change its theme.
- Very short answers, such as “ok” or “N/A”, can make a theme of their own.
- The percentages describe the answers that you analyze. They are not estimates for all your customers.
Your answers stay in your browser
The tool has no signup and asks for no email address. It sends no answer text to a server: the page has no form to submit, and the model runs on your device. The page downloads files only: the model files from Hugging Face and the program files from jsDelivr. We checked this with a network log of the browser during the analysis of a test file. The page also loads Plausible Analytics, which counts page views. It does not get your answers.
How to use the result
- Read the most typical answer and the example answers of each theme. Then write a name for the theme.
- Open the full list of each theme, and find the answers that do not fit. In the CSV file, change their theme.
- Read the answers in “Other”. They can contain a new theme or a problem that few people have.
- Use the corrected themes as a codebook. Our guide to qualitative vs quantitative survey questions explains how a second person can check the codes.
Frequently asked questions
Are the themes correct?
They are a draft, and not a final result. The model groups answers by similar meaning, and it can make mistakes. A person must read the answers in each theme and correct the themes before the result is used.
Do my answers go to a server?
No. Your browser reads the answers and runs the model on your device. The network requests of the page are downloads of the model and program files, and page view counts by Plausible Analytics. None of them contains answer text.
Do I need an account or an email address?
No. The tool is free, and it has no signup.
Which languages does it support?
The language list of the model card has 49 languages, including English, Spanish, French, German, Italian, Portuguese, Dutch, Danish, Swedish, Japanese and Korean. Answers with the same meaning in different languages can be in one theme. The model can be less accurate in some languages, so check the themes of each language.
How many answers can it analyze?
Up to 2,000 answers at one time. The time depends on your device and the length of the answers. For a larger survey, analyze a random sample of the answers.
Why is the first analysis slow?
Before the first analysis, your browser downloads the model files and the program files. After that, it uses its saved copy, so the next analysis starts faster.
Why are some answers in “Other”?
An answer goes to “Other” if it is not similar enough to a theme, or if its group is smaller than the minimum size. Read these answers. A theme can be important even if few people mention it.
Can one answer have two themes?
No. The tool puts each answer in one group, so that the counts add up to the number of answers. In manual coding, one answer can have more than one code. If an answer has two ideas, add the second theme yourself in the CSV file.
What does the similarity setting change?
It sets how similar two answers must be to share a theme. “Very similar” gives more themes with fewer answers in each. “Loosely similar” gives fewer, larger themes, and more answers with different meanings can be in one theme. Try two settings and compare the results.
Which file format can I use?
A CSV file in UTF-8, with one answer in each row of a column. In Excel, save the file as “CSV UTF-8”. In Google Sheets, download the sheet as comma-separated values. A text file with one answer on each line also works: paste its text.