Extracting meaning from text.
WHAT THE TASKS INCLUDE
Classifying text
Identifying entities: names, places, organisations, amounts
Determining intent Extracting relationships Answering questions Summarising
WHAT INTENT CLASSIFICATION DOES
Determines what the user wants from what they said.
WHAT ENTITY EXTRACTION DOES
Identifies the specific values the intent needs.
WHY THOSE TWO TOGETHER
They convert an utterance into an action with parameters.
WHAT CHANGED WITH LARGE LANGUAGE MODELS
Many of these tasks became achievable without task-specific training.
WHAT THAT MEANS PRACTICALLY
Building understanding components is far cheaper than it was.
WHAT REMAINS DIFFICULT
Ambiguity Context spanning a long conversation Domain-specific vocabulary Languages with limited data
WHY THAT LAST POINT MATTERS HERE
Performance in Nigerian languages and in local English varieties is substantially worse than in the languages models were mainly trained on.
WHAT TO TEST WITH
Actual user language, including code-switching between languages.
WHAT TO NEVER ASSUME
That published benchmark performance applies to your users.
WHAT TO PROVIDE
A route to a human, always.
