CanDoYa
All tools
Saved toolsYour favorites

Add to favorites

Saved in this browser only.
Private browser utility / Social & Creator

Free Named Entity Recognition Tool

English only

Runs entirely in your browser - no upload, no sign-up.

Live workspaceLocal processing

Paste English text to find names, organizations, locations, and miscellaneous entities.

Up to 5000 characters. Longer text is split into smaller model-safe passages without being uploaded.

Your text is processed in a browser worker and is not uploaded. The first run fetches the public AI runtime and model files.

Four focused entity labels

Run the English DistilBERT model to highlight full entity spans and group repeated mentions into a structured table.

The model recognizes people, organizations, locations, and miscellaneous named entities. It does not have date, email, phone, money, or product labels.

The model learned from English Reuters news. Informal writing, specialist terms, unusual names, and text in other languages can be missed or mislabeled.

Share this tool
named entity recognition / browser utility
01 / Overview

What is named entity recognition?

Named entity recognition, or NER, finds spans of text that refer to named people, organizations, locations, and other proper-name concepts. This tool runs an English DistilBERT model in your browser, highlights each detected span, groups repeated mentions, and exports a CSV without sending your text for server-side analysis.

02

How to use

  1. 01
    Paste an English passage

    Enter up to 5,000 characters from a news story, report, transcript, or meeting note. Load the sample if you want to see the expected result shape.

  2. 02
    Run the local NER model

    Click Find named entities. Allow the first model download to finish; later visits can usually reuse the browser cache.

  3. 03
    Review and filter

    Inspect the highlighted spans and grouped table. Raise or lower the minimum score to hide or reveal weaker model predictions.

  4. 04
    Export the useful rows

    Copy the grouped entity table or download it as CSV, then verify names and categories against the source text before reuse.

03

Who it's for

  • Researchers and analysts turning English news, reports, or interview notes into a quick list of people, organizations, and places.
  • Meeting and transcript reviewers finding recurring named entities before creating tags, summaries, or follow-up notes.
  • Editors and fact-checkers scanning a draft for names that deserve spelling, context, or identity checks.
  • Developers and data teams testing an NER workflow and exporting grouped examples before building a larger processing pipeline.

This named entity recognition tool uses dslim's 66-million-parameter DistilBERT NER model, converted to ONNX for Transformers.js. It was fine-tuned on the English CoNLL-2003 dataset of Reuters news and has four labels: person, organization, location, and miscellaneous. It does not include dedicated labels for dates, money, emails, phone numbers, products, or street addresses.

The first analysis downloads quantized model files of about 70 MB plus the browser AI runtime. A worker keeps loading and inference away from the main interface. WebGPU is preferred when a compatible graphics adapter is available; WebAssembly is the CPU fallback. Browsers normally cache the files, but private browsing, storage cleanup, or cache eviction can trigger another download.

FAQ

Is my text uploaded to a server?

No. Your text is passed to a worker inside your browser and is not included in the requests that fetch Transformers.js or the public model files. The AI runtime and weights come from external hosts on first use, so network activity is required until those files are cached.

Is this named entity recognition tool free?

Yes. There is no account, payment, API key, or analysis quota. Your device performs the inference. The practical costs are the initial model download, browser storage, memory, and processing time. Your network provider may still charge for the data used to fetch the model.

How much text can I analyse?

Each run accepts up to 5,000 Unicode characters. The tool splits the text into smaller passages before inference so the model stays within a practical token window and the interface remains responsive. For longer documents, analyse coherent sections separately and combine the reviewed CSV files afterward.

Which entity types can the model detect?

The model has four trained categories: person, organization, location, and miscellaneous. Miscellaneous covers named concepts that do not fit the first three classes. There are no dedicated date, time, money, percentage, email, phone, product, event, or address labels, so the tool does not claim to extract them.

Does the NER model work in languages other than English?

The model is English-specific because it was fine-tuned on the English CoNLL-2003 dataset. It may output something for other languages, but those results do not have the same training basis and can be misleading. Use this page for English text and choose a language-specific or multilingual NER model elsewhere when needed.

What does the confidence threshold mean?

The threshold hides entity spans whose model score falls below your chosen value. It is useful for exploring the precision and recall tradeoff, but the score is not a calibrated probability that the entity is correct. Domain shift, ambiguous names, tokenization, and context can still produce confident mistakes.

Will named entity recognition find all personal information?

No. This is not a complete PII detector or redaction system. The model may find person names and some locations or organizations, but it has no dedicated labels for email addresses, phone numbers, account numbers, dates, or street addresses. Use a purpose-built process and human verification for privacy or compliance work.

Does the tool work without WebGPU?

Yes. It tries WebGPU first when the browser exposes a compatible graphics adapter. If WebGPU is missing or model startup fails, the worker retries with the WebAssembly CPU backend. The same four labels are used, but CPU loading and inference can take longer on low-power devices.