Optimizing the Features of CRF-based Named Entity Recognition for Patent Documents

摘要

This paper proposes a feature optimization of named-entity recognition (NER) system for patent documents that automatically recognizes named entities such as product names, service names, and technology names that are topic words in patent description sentences. Named entity recognition problem can be approached as a kind of tagging problem in which NER tags are attached to the named entities in a given sentence. We tried to apply a machine learning model CRF that is widely accepted as an excellent method to a tagging problem. We constructed an NER-tagged patent corpus in which 668,260 tokens (approximately 15,000 sentences) are included for training data and 74,248 tokens for test data. We evaluated each CRF features and tried to find an optimal CRF feature set. The accuracy of our NER system was 93.69% with an F1 score 65.40.