Patrice Lopez (INRIA, FR)

Abstract
 

This paper describes experiments with Conditional Random Fields (CRF) for extracting bibliographical references in patent documents. CRF are used for performing extraction and parsing tasks which are expressed as sequence tagging problems. The automatic recognition covers references to other patent documents and to scholarship publications which are both characterized by a strong variability of contexts and patterns. Our work is not limited to the extraction of reference blocks but also includes fine-grained parsing and the resolution of the bibliographical references based on data normalization and the access to different online bibliographical services. For these different tasks, CRF models surpass significantly existing rule-based algorithms and other machine learning techniques, resulting more particularly in a very high performance for patent reference extractions with a reduction of approx. 75% of the error rate compared to previous works.

INRIA