Improved Named Entity Recognition Through SVM-Based Combination

Named Entity Extraction (NER) consists in identifying specific textual expressions, which represent various types of concepts: persons, locations, organizations, etc. It is an important part of natural language processing, because it is often used when building more advanced text-based tools, especially in the context of information extraction. Consequently, many NER tools are now available, designed to handle various sorts of texts, languages and entity types. A recent study on biographical texts showed the overall indices used to assess the performance of these tools hide the fact they can behave rather differently depending on the textual context, and could actually be complementary. In this work, we check this assumption by proposing two methods allowing to combine several NER tools: one relies on a voting process and the other is SVM-based. Both take advantage of a global text feature to guide the combination process. We extend an existing corpus to provide enough data for training and testing. We implement an open source flexible platform aiming at benchmarking NER tools. We apply our combination methods on a selection of NER tools, including state-of-the-art ones, as well as our custom tool specifically designed to process hyperlinked biographical texts. Our results show both proposed combination approaches outmatch the individual performance of all the considered standalone NER tools. Of the two, the SVM-based approach reaches the highest performance