Enhanced Distant Supervision with State-Change Information for Relation Extraction

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5573–5579 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5573 Enhanced Distant Supervision with State-Change Information for Relation Extraction Jui Shah*,1, Dongxu Zhang*,1, Sam Brody2, Andrew McCallum1 1University of Massachusetts Amherst, 2Bloomberg {jbshah, dongxuzhang, mccallum}@cs.umass.edu sbrody18@bloomberg.net Abstract In this work, we introduce a method for enhancing distant supervision with state-change information for relation extraction. We provide a training dataset created via this process, along with manually-annotated development and test sets. We present an analysis of the curation process and data, and compare it to standard distant supervision. We demonstrate that the addition of state-change information reduces noise when used for static relation extraction, and can also be used to train a relation-extraction system that detects a change of state in relations. Keywords: Corpus, Information Extraction, Linked Data, Weakly-supervised Learning, Relation Extraction 1. Introduction Relation Extraction (RE) is an important task in nat- ural language processing, both as an end goal (e.g., for knowledge-base population), and as a component in downstream applications such as search engines, recommendation systems, and question-answering sys- tems. Most of the work in RE focuses on static rela- tions, where the question of when the relation held is ignored. However, there has also been significant inter- est in temporal aspects of RE (see Section 2). In this paper, we introduce a new task State-change Relation Extraction, which focuses specifically on the start and end of the relation, or state-change points. De- tecting a change of state as it occurs is of value in and of itself, for example in an algorithmic trading system, that scans news feeds as they are published. And cur- rent knowledge bases (such as Wikidata) usually have time labels of relations to indicate the valid duration of them. In addition, change-of-state information can benefit systems focused on static relations, by reduc- ing noise when using distant supervision (DS) (Mintz et al., 2009) since relationships mainly appear in news articles during the time they hold. In this work, we create a dataset labeled with change- of-state information, present an analysis of the curation process and data, and compare it to classic distant su- pervision without temporal information. We use the temporally-linked dataset to train a relation-extraction system to detect change of state in four relations, and also experiment with using the change-of-state infor- mation to reduce noise in distant supervision for static relation extraction. We conclude with an analysis of our results and directions for future work. Contributions of this paper mainly rely on three folds: • We introduce a new task State-change Relation Extraction and provide a dataset with four relation types on general domain new corpus. * Equal contribution • We propose a simple yet effective enhancement on traditional distant supervision to capture the state- change relations via time window alignment. • We also show that time window can de-noise the distant supervision signals for static relation ex- traction. Our annotated data and the …

Content truncated to 500 words