Conditional Random Fields model have been a popular method for Named Entity Recognition as it accounts for statistical dependencies between entities and can include observed features that can aid with the classification. It can predict the most likely sequence of labels that correspond to a sequence of inputs.

Implementing this model can give an output of transition weights for the top likely and top unlikely transitions. Furthermore, weights are given to certain word features. When classifying a language for example, the model recognizes that when the last three letters of a word are –ish, the word is likely to concern a language like English or Swedish.

From what I have read, the CRF has a single exponential model for the joint probability of the entire sequence of labels given the observation sequence. However, I do not understand what the weights imply. My question therefore is how I can interpret the transition weights and the feature weights? What does a transition weight of 7.029925 or a feature weight of +2.397 exactly mean? Some more examples are given down below:

Top likely transitions:

B-ORG -> I-ORG 7.029925

I-ORG -> I-ORG 6.672091

B-PER -> I-PER 5.898448

y=B-LOC top features

Weight Feature

+2.397 word.istitle()

+0.099 word.isupper()

-0.152 word.isdigit()


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.