Conditional Random Fields (CRF): Short Survey

CRF is not very good for keywords extraction as soon as it cannot handle unknown words. Moreover, adding new data to the training dataset forcers us to re-train the whole CRF model – and it may be quite time-consuming due to the high complexity of the training phase of the algorithm. CRF shows good performance when dealing with entity recognition (any types of entities, including named entities, time expressions, etc.). It can use both linguistic (characters, words) and non-linguistic information (upper/lower case, punctuation marks, spaces etc.). The achievable quality of entity recognition is about 0.7-0.85