关键词: 潜在标准必要专利识别, Bert, CNN, 语义特征

Abstract:    Under the background of innovation leading development and the strategy of rejuvenating the country through science and education, the standard essential patent (SEP), as the main carrier of technological innovation, is an important way for the country to master the core industrial technology. Holding many SEP of key technologies in advance is the centralized embodiment of national innovation ability, and an important link for enterprises to control technology dominance. SEP has extremely high strategic and economic value. Enterprises can dominate the entire market by licensing SEP to each other, charge high license fees, and establish market barriers to competitors. Although the Standardization Committee has disclosed many SEPs, more patents will be included in the standards in the future with the development of product technology. At the same time, enterprises apply for patents from multiple angles around the technical points of the standard, continue to layout patents around the existing standards. Therefore, exploring potential SEP has important practical significance for building an innovative country, optimizing the patent distribution of enterprises, enhancing the competitiveness of core enterprises in China, increasing license income and avoiding business risks.This paper proposes a model to identify potential standards from prospect of semantic feature, using global semantic features and higher-dimensional semantic features of patents. Firstly, construct the patent sample set. Collect the declared standard essential patents data, including the patent number, claims, title, abstract and other information, and identify these as the SEP. The same number of non-SEP patent data were randomly sampled as negative samples and mixed with SEP patents to complete the construction of patent sample set. Secondly, Bert model is used to extract structured implicit global semantic features from the context of patent claims, title and abstract, and output high-dimensional semantic vectors. Thirdly, CNN neural network is used to extract the high-dimensional semantic features of the high-dimensional vector output by Bert, and the potential SEP is identified according to the feature extraction results. Lastly, according to the semantic similarity measure of vector, the predicted standard code is outputted for the potential SEPs. The findings are as follows. Firstly, when the amount of data is large, the accuracy and consistency of identifying potential SEPs based on the semantic features of claims are better than that based on the title and abstract, so the claims can provide richer semantic information of patent semantic features. Secondly, compared with the Doc2Vec-RF model, in terms of potential SEP prediction, the greater the number of predicted patents data, the better the prediction effect. For corresponding standard prediction, the performance of this method is nearly 10% higher than that of previous studies, and the test results are stable. Lastly, changing the amount of data tested in this paper, it is found that the experimental results float less and have good robustness.The conclusion of this paper not only provides reference value for the innovation strategy, but also provides practical significance for management practice. On the one hand, this method assists patent holders to analyze their own patents and utilize their own patents for potential development. On the other hand, enterprises can use this method to exploit the undisclosed standard essential patents of competitive enterprises to warn potential patent hijacking from competitors, serving for market competition.

Key words: potential standard essential patent identification, Bert, CNN, semantic feature