Presented at: 9th Extended Semantic Web Conference (ESWC2012)
This paper describes an approach for the task of named entity recognition in structured data containing free text as the values of its elements. We studied the recognition of the entity types of person, location and organization in bibliographic data sets from a concrete wide digital library initiative. Our ap-proach is based on conditional random fields models, using features designed to perform named entity recognition in the absence of strong lexical evidence, and exploiting the semantic context given by the data structure. The evaluation results support that, with the specialized features, named entity recognition can be done in free text in structured data with an acceptable accuracy. Our approach was able to achieve a maximum precision of 0.91 at 0.55 recall and a maximum recall of 0.82 at 0.77 precision. The achieved results were always higher than those obtained with Stanford Named Entity Recognizer, which was developed for well-structured text. We believe this level of quality in named entity recognition allows the use of this approach to support a wide range of information extraction applications in structured data.
Keywords: conditional random fields, metadata, named entity recognition, structured data
Resource URI on the dog food server: http://data.semanticweb.org/conference/eswc/2012/paper/research/6
Explore this resource elsewhere: