An Extensive Evaluation of Anaphora Resolution Based Abstract Generation System

This paper presents an extensive evaluation of anaphora resolution (AR) algorithm proposed for Abstract Generation System (AGS) and discusses the metrics used to measure its performance. The evaluation exercise is conducted with the test sets chosen from different domains and the results are compared with other existing open-source Anaphora Resolution Systems (ARSs) like Mitkov’s Anaphora Resolution System (MARS) and Java Resolution of Anaphora Procedure (Java RAP). We have selected news documents belong to five different domains and it is trained with AGS among which some set of documents and their corresponding extract is taken as the test set and executed in various ARSs for comparison. The proposed system is evaluated and its performances are studied with standard metrics of Information Extraction (IE) systems such as Success rate, Precision, Recall and F-measure. KeywordsAbstract Summary; Dangling Anaphora; Natural Language Processing (NLP); Text-mining Framework; IE Metrics