Text and NLP annotation for language models and search

Entities, intents, sentiment, document fields and moderation labels, marked to written span rules and checked against a gold set before delivery.

Start with a pilot batch See task types

Text annotation types

Choose one task or combine them. Every label type gets written rules with examples before work starts.

Named entity recognition

People, organisations, products, dates, amounts and any custom entity, marked at the exact character span.

Used for: Search, document AI, support routing

Text classification

Single or multi-label categories for messages, documents, reviews or tickets.

Used for: Routing, topic tagging, spam filters

Sentiment and emotion

Positive, negative or neutral, or finer emotions, at document, sentence or aspect level.

Used for: Reviews, social listening, CX analytics

Intent and slots

What the user wants, plus the details needed to act on it, such as date, product or order ID.

Used for: Chatbots, voice assistants, IVR

Relation extraction

Links between entities: who works where, which product has which issue, which clause refers to which party.

Used for: Knowledge graphs, contracts, research

Document and OCR annotation

Key-value fields, tables and layout regions on invoices, receipts, forms and IDs.

Used for: Invoice processing, KYC, form digitisation

Question and answer pairs

Questions and correct answers written from your documents, for search and assistant training.

Used for: RAG, FAQ bots, reading comprehension

Content moderation

Posts and comments labeled against your policy, with the clause and severity.

Used for: Community safety, marketplaces, UGC

Span rules decide your data quality

Most NER disagreements come from boundaries, not entity types. These are the kinds of rules we write down with you before labeling starts.

QuestionExampleExample rule
Include titles?"Dr. Meera Iyer"Title is outside the PERSON span
Legal suffixes?"Acme Foods Pvt Ltd"Suffix is inside the ORG span
Currency symbols?"₹2,499"Symbol is inside the MONEY span
Nested entities?"State Bank of India Delhi branch"Label the longest span only, unless your schema allows nesting
Typos and slang?"amzon", "flipkrt"Label as written, add the normalised value as an attribute

How a project runs

Pilot first, scale only after you have seen real output.

  1. Share samples and schema

    A sample of your text, your label list and what the model needs to do.

  2. Write the guidelines

    Definitions, span rules and edge cases with examples. You sign them off.

  3. Pilot batch and QA report

    A small batch, double-labeled in part and scored against your gold set.

  4. Refine

    You review. We update rules and fix what you flag.

  5. Scale

    Full delivery in batches, each with the same QA report.

What the QA report shows

Measured, not claimed. Acceptance criteria are agreed before the pilot.

  • Span-level accuracyPrecision, recall and F1 against your gold set, per entity type.
  • Agreement between annotatorsFor classification and sentiment, how often two annotators choose the same label.
  • Confusion between labelsWhich labels get mixed up most, so the guidelines can be tightened.
  • Edge-case logNew cases found during work, with the rule applied or a question for you.
  • Batch summaryItems delivered, reviewed and reworked.

Tools and formats

We work in your tool or set one up, and deliver in the format your pipeline reads.

Tools

Label StudioDoccanoProdigyLabelboxYour in-house tool

Export formats

JSONL with character offsetsspaCyCoNLL / BIOCSVLabel Studio JSON

Data security and personal data

Text often contains names, phone numbers and account details. We plan for that upfront.

  • NDA first. Signed before any text or guidelines are shared.
  • Masking rules agreed upfront. If your text has personal data, we agree what to mask or skip before work starts.
  • Need-to-know access. Only annotators and reviewers on your project can open your data.
  • Your tool, your rules. Data can stay inside your platform if you prefer.
  • Used only for your project. Never shared, sold or reused, and deleted on request.

Where teams use it

The task types above, applied to real products.

Support and chatbots

Intents, slots and routing labels on tickets, chats and emails.

E-commerce

Product attributes, review sentiment and search relevance.

Finance and insurance

Entities and fields in statements, claims and KYC documents.

Legal

Clauses, parties, dates and obligations in contracts.

Social and marketplaces

Moderation labels, spam and abuse detection.

Indian-language products

Hindi, Hinglish and regional-language text, with code-switching rules written into the guidelines.

Questions teams ask

If yours isn't here, ask in the form below.

How is text annotation priced?

Per item, per entity or per hour, depending on text length and how many label types each item needs. We quote after seeing samples and confirm the rate after the pilot.

Which languages do you support?

Tell us the languages and scripts you need, including Hinglish or other code-switched text. We confirm annotator availability for each before the pilot.

Can you handle nested or overlapping entities?

Yes, if your schema and tool allow it. We agree the rule upfront so every annotator handles them the same way.

What about personal data in the text?

We agree masking or exclusion rules before work starts, and only annotators on your project can access the data.

Can you work in our tool?

Yes. We can label inside your platform or set up Label Studio or Doccano and deliver in your format.

Do you also do other data types?

Yes: image & video, audio & speech, llm evaluation, data collection.

Start with a pilot batch

Tell us about your text data and what your model needs to understand. We'll reply with questions and a pilot plan.

Prefer email? Write to hello@norbeon.com.

Run an annotation or data team? Register as a vendor.

By sending this form, you agree to our Privacy Policy.