Sentiment Analyzer
Detect positive or negative sentiment in any text using AI — runs entirely in your browser, nothing uploaded.
All processing runs entirely in your browser using WebAssembly. No image or text is ever uploaded to a server. First run downloads the AI model (~44–70 MB) and caches it locally — subsequent runs are instant.
Characters: 0
Sentiment Analyzer — AI-Powered Text Sentiment Detection
This tool uses DistilBERT fine-tuned on the Stanford Sentiment Treebank 2 (SST-2) dataset to classify text as positive or negative. DistilBERT is a lightweight transformer model that retains 97% of BERT's accuracy at 40% of the size, making it practical for in-browser inference via WebAssembly. The model (~70 MB) downloads once and is cached locally. Use it to analyse customer reviews, social media posts, survey responses, or any free-text feedback. Results include a confidence percentage for each class. Nothing is uploaded; all inference runs on your device.
Built and maintained by Meet Shah · Last updated
What this tool is used for
- Getting a rough read on the tone of a batch of feedback.
- Checking whether a draft reads more negatively than intended.
- Sorting responses into rough buckets before reading them.
- Comparing the tone of two pieces of copy.
- Screening a long comment thread for the most negative entries.
Frequently Asked Questions
- What model is doing the classification?
- DistilBERT fine-tuned on SST-2, running in your browser through Transformers.js. It is a binary classifier — positive or negative — trained on movie-review sentences, which is the corpus that shapes what it is good and bad at.
- Why is there no neutral option?
- Because SST-2 is a two-class dataset, so the model has no neutral to output. A genuinely neutral sentence is forced into one side with a confidence near 50%, which is why a low-confidence result should be read as "no opinion detected" rather than a weak verdict.
- What does it get wrong?
- Sarcasm, negation at a distance, and domain shift. "Not bad at all" and a glowing review of a terrible product are both classic failures, and text from a domain far from movie reviews — clinical notes, legal writing — degrades noticeably.
- Is the confidence score a probability?
- It is a softmax output, which behaves like one and is not calibrated like one. Models of this size are routinely overconfident, so treat the number as a ranking signal — useful for sorting a list of comments, unreliable as "87% likely to be positive".
- Is my text uploaded?
- No. The model downloads once and then runs entirely in the tab via WebAssembly, so the text never leaves the machine. That is the practical reason to run a small model locally rather than call an API — it works on content you would not send anywhere.
Common errors and gotchas
- Trusting the label on sarcasm or irony, which sentiment models handle badly.
- Applying it to domain language where 'negative' words are neutral, such as a bug report.
- Treating a score as a measurement rather than a rough classification.
- Running it on very short text, where there is too little signal.
- Making a decision about a person based on a model's read of one message.
Related Text Tools tools
Case Converter
Convert between Sentence, Title, camelCase, snake_case, and more.
Word Counter
Count words, characters, and analyze keyword density.
Text to Binary
Convert text into its 8-bit binary representation.
Binary to Text
Decode 8-bit binary back into readable text.
Text Reverser
Reverse text by characters, words, or lines.
Remove Duplicate Lines
Delete repeated lines and keep your list unique.
Remove Empty Lines
Strip blank and whitespace-only lines from text.
Sort Lines Alphabetically
Sort lines A–Z or Z–A, case-sensitive or not.