A deep learning based method for extracting semantic information from patent documents

The text-based patent analysis is grounded in information extraction technique. However, such technique suffers from obvious defects such as low degree of automation and unsatisfactory extraction accuracy. To deal with these problems, after an information schema is pre-defined, which contains 17 types of entities and 15 types of semantic relations, a dataset of 1010 patent abstracts is annotated and opened freely to the research community. Then, a novel patent information extraction framework is proposed, in which two deep-learning models, BiLSTM-CRF and BiGRU-HAN, are respectively used for entity identification and semantic relation extraction. Finally, to demonstrate the advantages of the new framework, extensive experiments are conducted, and the SAO method and PCNNs model are taken as respective baselines on the framework and module levels. Experimental results show that our framework out-performs the traditional one in terms of automation and accuracy, and is capable of extracting fine-grained structured information from patent texts.