The document describes using conditional random fields (CRF) for opinion mining tasks including detecting expressions, holders, and targets from text. It involves developing a KAF feature extractor to generate input for CRF in TAB format, training a CRF model, and converting CRF output back to a KAF layer to extract opinions. Key steps are extracting features from KAF, generating the CRF training format, training and using the CRF model, and converting CRF predictions back to the KAF representation.