ABSTRACT
Patent citation recommendation, as one of the prior-art search, is critical in the patent examination. Other than paper citation, the content scope is important for patent recommendation. A thorough prior-art search for patent will both help applicants to properly scale the proposal and USPTO examiners to evaluate the originality of the invention and accept how the invention can be distinguished from the prior art. Integrating technical terminology will help determine the scale of technical paper in a more precise way. MeSH ontology is a controlled vocabulary with hierarchical semantics developed by the National Library of Medicine to index the biomedical articles.
Here we developed a method to assign and evaluate the MeSH semantic similarity for the patent document. The experimental results generate a three-step “best measure set” to integrate the MeSH descriptor semantic similarity into the document similarity measure. We also further improved the result by using the Medical Text Indexer (MTI) to assign the MeSH descriptors.
We used 28438 Biotechnology patents as patent pool. Results evaluated by the average recall rate suggest though the MeSH semantic similarity measure may not exceed some of the sophisticated content-based measure, its outperforming the basic cosine similarity measure suggests its potential to act as one feature to be integrate into other complex search engine.
Chapter 1
INTRODUCTION
The rapid growing in the information technology development is impacting Intellectual Property field. Patent citation recommendation, one of the prior-art search, is critical in the patent application both for the applicants and United States Patent and Trademark Office (USPTO) examiners. A thorough prior-art search will help the patent applicant to narrow down the candidate patent into a proper scope that are both not too wide to be patented and innovated enough to be applied as a new patent. On the other hand, a precise and comprehensive prior-art search will help the USPTO examiners to decide whether the application is patentable and scale down the application into proper scale with supplementary references.
Patent documents are viewed as a kind of technical document, thus, integrating the technical terminology may help describe the patent more precisely and improve the recommendation. MeSH ontology is a controlled vocabulary with hierarchical semantics developed by the National Library of Medicine (NLM) to index the articles in the MEDLINE database. In the MEDLINE database, MeSH descriptors are pre-assigned to each article by the indexing experts. Usually each article is assigned 5~15 descriptors. The MeSH ontology is originally developed for article indexing and cataloging only in the PubMed, which is the major portal in NLM for the search of a large number of biomedical articles including MEDLINE database. However, most of the studies have focused on applying the MeSH semantics into MEDLINE articles retrieval. But few studies have been found to applying the MeSH ontology outside MEDLINE.
In this work we propose to apply the MeSH ontology into the patent citation recommendation. Since one MeSH descriptor usually contains more than one MeSH tree node in 2 the hierarchical structure, and one document is most likely to be assigned with several MeSH descriptors, we proposed a three-step semantic similarity measure.
First we compared four methods of computing semantic similarity between two tree nodes; for the second step, we proposed to use the Average Maximum Measure (AMM) to measure the similarity between two MeSH descriptors; and the final step, we propose both the AMM and a modified tf-idf related AMM measure. The MeSH descriptors are originally assigned by directly extracted from the claim of each patent document. To further improve the similarity measure, we also apply the Medical Text Indexer (MTI) for MeSH descriptor auto assignment.
We apply this similarity measure into 28438 Biotechnology patents, and randomly selected 100 patents from this patent pool as the testing set. For evaluation purpose, all the patents in the testing set have at least four reference patents. Since in this task, it is more important to not miss any patents, the patent citation recommendation can be viewed as recalloriented. We use the average recall rate as the evaluation matrix. Average precision rate is also listed for reference.
The rest of the thesis is organized as below:
Chapter 2 will provide the necessary background and related work both about the citation recommendation and the existing application of MeSH ontology into MEDLINE and other databases.
Chapter 3 will describe and specify the MeSH semantic similarity in patent documents, introduce our three-step similarity measure and our candidate methods in each step. This chapter will also describe the evaluation matrices and two baseline methods for comparison purpose. Chapter 4 will describe the dataset for the experiment and compare the results using different methods.
Chapter 5 will describe the differences in results from the experiments of Chapter 4. It also will draw the underline issues related with each method.
Chapter 6 will highlight the conclusion from the study and propose the future work. Appendix will describe some technical issues and problems in details during the implementation of the method and evaluation.
EXPLORING APPLYING MESH ONTOLOGY FOR BIOMEDICAL PATENT CITATION RECOMMENDATION