Two-Stage Patent Retrieval Method Considering Claim Structure

Abstract This paper proposes a patent retrieval method that consists of two processing stages. In Stage 1, analysis and retrieval methods to improve recall are applied. In Stage 2, the top N documents retrieved in Stage 1 are re-arranged by applying analysis and retrieval methods that consider the claim structure to improve precision. This paper gives an overview of this retrieval method and evaluates its performance at the NTCIR4 Patent Retrieval Task. Keywords: Patent Retrieval, Claim Structure Analysis, Keyword Extraction, Allomorph Expansion, Related Term Expansion, Document Filtering, Score Merging. 1. Introduction Text retrieval methods using a natural language text as an input are becoming popular. These methods focus on a keyword set extracted from the input text and calculate the similarity between this keyword set and that extracted from each of the retrieval target documents. Keyword-based document retrieval methods have three technical issues: (a) How to extract appropriate keywords (b) How to assign weights to the keywords (c) How to treat allomorphs and synonyms This paper proposes a patent retrieval method to solve these problems and to improve retrieval accuracy. This method consists of two processing stages: in Stage 1, analysis and retrieval methods to improve recall are applied, and in Stage 2, the top N documents retrieved in Stage 1 are re-arranged by applying analysis and retrieval methods that consider the claim structure to improve precision. Section 2 overviews our two-stage retrieval method. Section 3 describes the analysis and retrieval methods used in each stage. Section 4 evaluates the feasibility of our method by using test data of the NTCIR4 Patent Retrieval Task.