Optimizing Entity Recognition

SpaCy (SpacyEntityExtractor) - ner_spacy

• Pretrained entity extractors

• Statistical BILOU transition model

Duckling

• Number related information (dates, distance, duration)

• Run server on docker image ner_duckling

NER_CRF (CRFEntityExtractor)

• trained from scratch

• Need to annotate training data yourself

• Annotate training examples EVERYWHERE in training data (even if entity is not relevant for intent)

• Use of lookup tables makes ner_crf prone to overfitting

  • If training data matches Regex or Lookup, it will ignore other features, so if you have message with entity that is not matched by Regex, ner_crf will not detect

Common Problems

Entities are not recognizing unseen values

• Could be:

  • Lack of training data

  • Overfitting of ner_crf (try training model without regex or look up)

Map extracted entity to different value

• ner_synonyms