Named Entity Recognition Online — Free, Server-Side, No Signup

The recognizer is a three-stage pipeline. The model itself is a single network — BERT-base fine-tuned on CoNLL-2003 — but most of what makes the output useful happens on either side of that inference call.

1. Tokenisation and inference

Your text is sent once to api-inference.huggingface.co and passed to dslim/bert-base-NER — the BERT-base checkpoint fine-tuned on the CoNLL-2003 English NER dataset (Tjong Kim Sang & De Meulder, 2003). The Inference API tokenises the input with BERT's WordPiece tokeniser, runs the model, and returns one prediction row per sub-word token. We ask for aggregation_strategy="none", so each row carries a BIO label (B-PER, I-LOC, O, etc.) plus a softmax confidence and the source-text offsets. No model weights leave the server — only the small JSON payload.

2. Span aggregation

The browser merges the per-token predictions into entity spans. Two strategies are exposed: simple, which extends a span as long as adjacent tokens share the same base type, and first, which starts a fresh span on every B-tag. Sub-word continuations marked with ## always inherit the parent span's label so a word like “Wickremesinghe” survives intact. The per-token confidences are averaged across the merged span. You can change the strategy without re-running the model — only the merge step replays.

3. Threshold and dedupe

Each merged span's averaged confidence is compared against the slider value (default 75%). Spans below the threshold are dropped; the count is reported in the “dropped” tally. What remains is grouped by the tuple (entity type, lowercased trimmed surface form), so the same name mentioned three times collapses into one row with count: 3 and the average confidence across all three. The entity table is sorted by count descending, then by first occurrence — most-discussed entities surface first.

The pipeline is deterministic given the same text. The HF Inference API is stateless — every call is independent and we send no identifiers along with your input. The four-class taxonomy (Person, Organisation, Location, Misc.) comes straight from the CoNLL-2003 shared task and is what every off-the-shelf English NER tutorial in the ecosystem uses today.

• Hugging Face — dslim/bert-base-NER (model card)
• Tjong Kim Sang & De Meulder, 2003 — Introduction to the CoNLL-2003 Shared Task (the four-class entity taxonomy)
• Hugging Face Inference API — official documentation
• Hugging Face — Token Classification task guide (BIO scheme and aggregation strategies)
• Devlin et al., 2018 — BERT: Pre-training of Deep Bidirectional Transformers (arXiv:1810.04805)

The model card, API endpoint, and taxonomy were last cross-checked on 2026-05-12. The page is reviewed whenever the Hugging Face Inference API contract changes or the upstream model repo is re-uploaded.