Key points are not available for this paper at this time.
Named Entity Recognition (NER) is a key building block of any Natural Language Processing (NLP) system, making possible the detection and classification of entities (e.g., Person, Location) in any given text. While a large number of NER software exist today, it remains difficult for NLP and NER practitioners to clearly and objectively identify what software perform(s) the best. One of the reasons is the difference in results across the literature and the lack of information needed to be able to fully reproduce the experiment. To overcome this problem, this paper presents a comprehensive and replicable study to assess the performance of NER software, thus laying the groundwork for future benchmarking and meaningful comparison studies. As part of our experiments, the latest version of five well-known NER software were selected, along with two distinct corpora. We observe a discrepancy between the result we get and the result found in the literature being around 50% in certain cases. We also found that StanfordNLP usually performs the best.
Schmitt et al. (Tue,) studied this question.