Coupling Knowledge-Based and Data-Driven Systems for Named Entity Recognition

Within Information Extraction tasks, Named Entity Recognition has received much attention over latest decades. From symbolic / knowledge-based to data-driven / machine-learning systems, many approaches have been experimented. Our work may be viewed as an attempt to bridge the gap from the data-driven perspective back to the knowledge-based one. We use a knowledge-based system, based on manually implemented transducers, that reaches satisfactory performances. It has the undisputable advantage of being modular. However, such a hand-crafted system requires substantial efforts to cope with dedicated tasks. In this context, we implemented a pattern extractor that extracts symbolic knowledge, using hierarchical sequential pattern mining over annotated corpora. To assess the accuracy of mined patterns, we designed a module that recognizes Named Entities in texts by determining their most probable boundaries. Instead of considering Named Entity Recognition as a labeling task, it relies on complex context-aware features provided by lower-level systems and considers the tagging task as a markovian process. Using thos systems, coupling knowledge-based system with extracted patterns is straightforward and leads to a competitive hybrid NE-tagger. We report experiments using this system and compare it to other hybridization strategies along with a baseline CRF model.