Event-Guided Denoising for Multilingual Relation Learning

Proceedings of the 28th International Conference on Computational Linguistics, pages 1505–1512 Barcelona, Spain (Online), December 8-13, 2020 1505 Event-Guided Denoising for Multilingual Relation Learning Amith Ananthram Emily Allaway Department of Computer Science, Columbia University, New York, NY amith.ananthram@columbia.edu, {eallaway,kathy}@cs.columbia.edu Kathleen McKeown Abstract General purpose relation extraction has recently seen considerable gains in part due to a mas- sively data-intensive distant supervision technique from Soares et al. (2019) that produces state- of-the-art results across many benchmarks. In this work, we present a methodology for collect- ing high quality training data for relation extraction from unlabeled text that achieves a near- recreation of their zero-shot and few-shot results at a fraction of the training cost. Our approach exploits the predictable distributional structure of date-marked news articles to build a denoised corpus – the extraction process filters out low quality examples. We show that a smaller multi- lingual encoder trained on this corpus performs comparably to the current state-of-the-art (when both receive little to no fine-tuning) on few-shot and standard relation benchmarks in English and Spanish despite using many fewer examples (50k vs. 300mil+). 1 Introduction Multilingual relation extraction is an important problem in NLP, facilitating a diverse set of downstream tasks from the autopopulation of knowledge graphs (e.g., Ji and Grishman (2011)) to question answering (e.g., Xu et al. (2016)). While early efforts in relation extraction used supervised methods that rely on a fixed set of predetermined relations, research has since shifted to the identification of arbitrary unseen relations in any language. In this paper, we present a method for extracting high quality relation training examples from date-marked news articles. This technique leverages the predictable distributional struc- ture of such articles to build a corpus that is denoised (i.e., where sentences with the same entities only express the same relation, see Figure 1). We use this corpus to learn general purpose relation representa- tions and evaluate their quality on few-shot and standard relation extraction benchmarks in English and Spanish with little to no task-specific fine-tuning, achieving comparable results to a significantly more data-intensive approach that is the current state-of-the-art. The current state-of-the-art model, “Matching the Blanks” or MTB (Soares et al., 2019), is a distant su- pervision (Mintz et al., 2009) technique that provides large gains on many relation extraction benchmarks and builds on Harris’ distributional hypothesis (Harris, 1954) and its extensions (Lin and Pantel, 2001). Soares et al. (2019) assume that the informational redundancy of very large text corpora (e.g., Wikipedia) results in sentences that contain the same pair of entities generally expressing the same relation. Thus, an encoder trained to collocate such sentences can be used to identify the relation between entities in any sentence s by finding the labeled relation example whose embedding is closest to s. While Soares et al. (2019) achieve state-of-the-art on FewRel (Han et al., 2018) and SemEval 2010 Task 8 (Hendrickx et al., 2019), their approach relies on a huge amount of data, making it difficult to retrain in English or any other language …

Content truncated to 500 words