Effective Selectional Restrictions for Unsupervised Relation Extraction

Alan Akbik, Larysa Visengeriyeva, Johannes Kirschnick, and Alexander Löser

Effective Selectional Restrictions for Unsupervised Relation Extraction

Technische Universitat Berlin

6th International Joint Conference on Natural Language Processing (IJCNLP 2013) in Nagoya, Japan in October 14-18, 2013


Unsupervised Relation Extraction (URE) methods automatically discover relations in text corpora of unknown content by grouping pairs of entities that occur
in similar patterns. In this paper, we show that using a feature generation technique that incorporates selectional restrictions (SR) into patterns significantly
improves the ability of the URE method to discover relations. We investigate different methods of modeling SR and illustrate in detail how to incorporate them into
the feature generation technique. In particular, we propose modeling selection restrictions in the open domain using a Web-derived soft clustering of n-grams.
We conduct an experimental evaluation in which we comparatively evaluate all methods against a strong baseline. Our results show that SR increase the discriminative
power of patterns, leading to improvements in overall relation extraction quality, as well as impacting the granularity of discovered relations.

Download: Link