Research On Patent Citation Recommendation

The patent aims to encourage the disclosure of new technologies and new ideas by granting inventors a proprietary right to use the invention within a limited time.There is a consensus that patent examiners,who need to determine the prior art relevant to the adjudication of patentability of patent applications,often lake the time or experience necessary to conduct adequate prior art search or evaluate the information they find.The traditional patent citation recommendation is very time-consuming.Thus,this dissertation mainly builds an automatic and efficient system of patent citation recommendation.The main research work of this dissertation includes:1.It firstly introduces the background knowledge of patent citation recommendation and fully studies the research status and basic method theory of patent citation recommendation.Also,the dissertation mainly analyzes the difficulties and challenges faced by this research and proposes corresponding research methods for these challenges.To eliminate the semantic gap between search terms and patent terms,make full use of existing information to retrieve related patents,use paper information to improve the effect of patent recommendation,propose two methods: a two-stage patent citation recommendation method based on learning to rank and a method for inventor-author alignment between the thesis network,and a method for inventor-author alignment between the paper network and the patent network.A patent citation recommendation system is implemented using the above two methods.2.To solve the problem that it is difficult to fully couple patents semantic and structural features and consider diversified information,this dissertation proposes a new two-stage patent citation recommendation framework based on learning to rank.To improve the usability of the algorithm,this method refers to the examination process of the patent examiner to conduct a preliminary screening in the patent database,and minimize the range of patent citation recommendations without losing many related patents,select the patent candidate set as the first stage.In the second stage,the method uses learning to rank methods to rank the candidate set of the patent.Learning to rank can refer to multidimensional features to rank documents,this method designs different feature extraction methods for patent data.To solve the term mismatch problem in patent documents,this method uses patent classification as a supervised learning semantic vector of patents and calculates document similarity through vector distance.To utilize the text information of the patent document,this method extracts keywords for the patent text and calculates the keyword similarity.To utilize the semi-structure information of patent documents,this method designs features for patent attributes,such as classification similarity.Using learning to rank,comprehensively considering the relationship of these similarities,a ranking model is learned.Finally,use the model to rank the candidate set.Experiments were performed on several datasets,and the result suggests that this study has certain applicability in the patent citation recommendation.3.Papers,as an important non-patent literature,play an important role in confirming the novelty,creativity,patent value and inventor value of patents.If the paper network and the patent network can be merged,it has an important guidance for recommending patent citations.Based on the text content and network structure,this dissertation proposes a method to align the paper author and the patent inventor between the paper network and the patent network.To align the author entities between two multi-source heterogeneous networks,we first need to select the anchor point that may be the same entity.This method first selects the patent inventor candidate set that may be the same as the author of the given paper according to the rules.To determine whether the inventor in the candidate set is the same as the author of the given paper,this method judges each candidate from the perspectives of collaborator network,text content and author company,etc.,all confidences are greater than the given threshold are used as alignment results.Experiments on the constructed data set show that this study can align the author and inventor between two heterogeneous information networks.Finally,this method is used to align a group of selected high-index scholars in the patent network of the USPTO.4.Combining the above methods,we implement a patent citation recommendation system.The system has basic patent full-text search,field search functions,a patent citation recommendation module,and an author inventor alignment module.The patent citation recommendation module takes the full text of the patent as input and recommends related patent citations by the ranking model.The author inventor alignment module can start from a specific author and obtain all his papers and patents.