Inproceedings - Full-Paper Contribution

English

Meindl, Benjamin 
Ott, Ingrid 
Zierahn, Ulrich 

Proceedings of The 1st Workshop on Patent Text Mining and Semantic Technologies (PatentSemTech 2019) 

2019

1st Workshop on Patent Text Mining and Semantic Technologies (PatentSemTech 2019) 

12-Sep-2019

Karlsruhe, Germany

5

patent classification; small sample size; convolutional neural network; language model pre-training; fast pre-training

In this paper, we develop binary patent classification algorithms for ambiguous concepts and small sample sizes. These are particularly useful for economic questions, which often require binary classification for implementing ambiguous and subjective concepts, where human classification is time-consuming, so that sample sizes are small. This covers examples such as whether workers are susceptible to automation or not, or whether a device is an automat or not. We compare the performance of naive Bayes, support vector machine, random forest and k-nearest neighbor classifiers with a the spaCy convolutional neural network (CNN) model, as well as spaCy CNN model pre-trained with patent data. The results show overall highest accuracy for the CNN models, with a significantly improved performance through pre-training. Our analysis suggests that the spaCy pre-trained CNN model provides a highly accurate NLP model, feasible for implementation without extensive computation capacity required. Pre-training was particularly beneficial for small sample sizes. Already 100 labeled patents lead to an accuracy of 77.2%. The low sample size required, may encourage researchers in various fields to use manually labeled patent data, for evaluating their specific question.