Abstract: The shortening of patent examination time and the increase of patent number bring great challenges to patent classification, and using patent automatic classification technology to improve the efficiency of patent classification and shorten the time of patent examination has become an important research topic. An automatic patent classification framework is proposed based on multi-dimensional features and graph convolutional networks. The framework extracts the patent features from the dimensions of patent abstract, citation patent and patent inventor according to document metrology and graph representation learning theory. Secondly, the patent-core word network is constructed by using the dimensionality features of patent abstracts, and the dimensionality features of citation patents and patent inventors are embedded into the patent-core word network as patent number features. The semi-supervised learning of graph convolutional network is used to determine the classification labels of patent nodes in the patent-core word co-occurrence network and complete the task of patent automatic classification. In order to verify the effect of the method, the patent data from the Incopat global patent database are used for experiments. The experimental results show that the patent text information and the patent structured information as the patent features can improve the patent classification accuracy, and the introduction of backward citation patent information can improve the patent classification accuracy. At the same time, the framework proposed in this paper also provides a new solution to the problem of patent automatic classification, and provides support for the implementation of the policy of shortening patent examination time.

Key words: patent; Graph Convolutional Network; multi-dimensional features; backward citation patent; automatic classification