Named Entity Enrichment Based on Subject-Object Anaphora Resolution

Abstract

Named Entity Recognition (NER) is an early stage processing of an Information Extraction, which identifies and classify entities in text. The outcomes of the task have become the foundation of building complex Information Extraction applications. With the enormous amount of information available everywhere, the area has gained lots of attention from the research community. Currently, there are two main approaches to perform the Named Entity Recognition; rule-based approach and machine learning approach. In order to improve the accuracy of the classification and performance of the recognizer, some researchers have implemented a hybrid approach, which is the combination of both approaches. Even though many research and works have been done in the Named Entity Recognition, there is still room available for improvement. This paper proposed to increase the accuracy of entities detected by implementing anaphora resolution during the preprocessing phase and a hybrid approach to classify the detected tokens during the classification phase. The hybrid approach is combined the Conditional Random Field (CRF) classifier with a gazetteer and pattern rules to perform classification. The result has shown that the application of anaphora and gazetteer has increased 46% accuracy of the detected entities for the person class.