Conditional random fields with semantic enhancement for named-entity recognition

We propose a novel Named Entity Recognition (NER) system based on a machine learning technique and a semantic network. The NER system is able to exploit the advantages of semantic information, coming from Expert System proprietary technology Cogito. NER is a task of Natural Language Processing (NLP) which consists in detecting, from an unformatted text source and classify Named Entities (NE), i.e. real-world entities that can be denoted with a rigid designator. To address this problem, the chosen approach is a combination of machine learning and deep semantic processing. The machine learning method used is Conditional Random Fields (CRF). CRF is particularly suitable for the task because it analyzes an input sequence of tokens considering it as a whole, instead of one item at a time. CRF has been trained not only with classical information, available after a simple computation or anyway with little effort, but with the addition of semantic information. Semantic information is obtained with Sensigrafo and Semantic Disambiguator , which are the proprietary semantic network and semantic engine of Expert System, respectively. The results are encouraging, as we can experimentally prove the improvements in the NER task obtained by exploiting semantics, in particular when the training data size decreases.