Towards automatic building of term hierarchies from large patent datasets

Term Hierarchies are structures that represent semantic relations among terms, usually of the type hyperonym and hyponym (generality and specificity, respectively). There are many scenarios that may benefit from the knowledge within Term Hierarchies. Particularly, in the patents genre there is an important demand of knowledge representation for Information Retrieval purposes, and few works have approached this demand. In this work we proposed a three stage strategy for term hierarchy building from patents. In the first stage, terms were extracted through non-phrase identification; in the second stage terms were organized hierarchically through n-gram decomposition; linally, in the third stage, the term hierarchy was eoriched with term embeddings information, particularly from Word2Vec model. This strategy was applied over patents from the United States Patent and Trademark Office (USPTO) collection. Each term in the hierarchy generated a set of associated patents. For the evaluation task we applied two strategies over the sets of associated documents, one based on the clustering degree of the sets, and the other one based on the IPC (International Patent Classification) categories proportions within the sets. Results show tbat the produced term hierarchy efficiently captures generic and specific concepts.