← All articles
Data labeling

Text labeling for intents and entities

Build a text annotation schema for user goals and meaningful spans.

Decide what to do with multiple goals

A message can contain several actions. In “Change my delivery address and tell me when the parcel arrives,” both requests matter. Decide whether the application needs multiple intent labels or a single routing decision. For single-label tasks, define precedence so similar messages reach the same queue.

Specify exact span boundaries

For entity labeling, state whether words such as “on,” punctuation, currency symbols, or unit names belong inside the span. Include repeated mentions and nested phrases in the pilot. Check intent errors separately from boundary errors because each calls for a different change to the guidelines.

Annotate a sentence from start to finish

Consider “Move my order to 12 Pine Street and check its delivery date.” Mark address change and delivery tracking if multiple intents are allowed. For entities, specify whether “12 Pine Street” is a single address span and whether “its” receives a reference label. Run this example through your expected storage format so reviewers see how the two intents and the extracted span become structured fields.

Write rules for overlaps and repeated mentions

Decide whether a location entity inside a full address should also be labeled. If the training format only permits one span at a position, choose a precedence rule that preserves the downstream requirement. For repeated mentions, determine whether every mention counts or only the first. Also define whether spans are measured by character or token and whether offsets refer to the original text.

Include difficult language in the pilot

Use examples with typos, shorthand, negation, and indirect requests. “I’m not asking for a refund” contains a refund keyword but does not express refund intent. “Where is it?” requires context that a standalone sentence may lack. When intent cannot be determined under your context rules, use a documented unresolved label and check how often it is applied.

Separate classification errors from extraction errors

Review intent confusion by class and score extraction both for exact boundaries and partial overlap when that distinction matters. A correct entity type with a trailing comma points to a span rule. A missing delivery intent points to taxonomy or precedence. Report these error types separately and rerun the specific slice after revising the relevant guideline.