Semantic and Topological Graphs for Patent Retrieval

Information Retrieval often relies on the indexed terms that occur in the documents, but rarely is the available associated metadata processed to aid in the retrieval process. In addition to this, the associated metadata is often interlinked data that can be the basis for graphs, such as citation or coauthorship graphs. One of these types of graphs can be build from the classification systems. Classification systems aim to organize the knowledge in the documents and make retrieval easier. This work utilizes two types of graphs that were developed out of the classification systems in previous work: Topological and semantic graphs. The topological graph is based on the classification system assigned by experts and the co-classification that can be found with those. The semantic graph is based on the textual content of the documents and the relations that can be found when vectors are assigned to the classes based on this content. The structure and weights of those graphs can then be utilized to supplement the retrieval process and re-rank the result lists. In this work we show how those graphs are a valuable addition to improve performance in patent retrieval tasks and how different versions of this retrieval process compare to each other.