A Set of Rules for Function-Oriented Automatic Multi-Sentence Analysis in Patents

Abstract

This study proposes some rules for performing a function-oriented search (providing function and object) to extract technical systems from patents, using syntax and dependency patterns to analyse multiple sentences. Unlike the most common inter-sentence analysis methods, the proposed method does not use context information or distance to link the elements of several sentences, but generic terms from patent ontology. The content provided by the rules was entirely derived from a statistical analysis of many patents from different domains, in order to provide a general validity for the rules. The application of the method in two case studies, related to metal cutting and manure processing, highlighted its main advantages. Its degree of automation is such that the expert is almost exclusively excluded, except in the definition of the function on which to build the document pool. The precision and the recall of the results during the tests exceeded 90%. The current limitation concerns the manual control of some results, about 25%, which derive from an additional set of dependency patterns that are difficult to automate and deserve further investigation. The technical systems are many more in number and are more detailed with regard to structural aspects than those obtainable by analysing only single sentences and/or syntax.