Matching the Blanks: Distributional Similarity for Relation Learning

Abstract

General purpose relation extractors, which can model arbitrary relations, are a core aspiration in information extraction. Efforts have been made to build general purpose extractors that represent relations with their surface forms, or which jointly embed surface forms with relations from an existing knowledge graph. However, both of these approaches are limited in their ability to generalize. In this paper, we build on extensions of Harris' distributional hypothesis to relations, as well as recent advances in learning text representations (specifically, BERT), to build task agnostic relation representations solely from entity-linked text. We show that these representations significantly outperform previous work on exemplar based relation extraction (FewRel) even without using any of that task's training data. We also show that models initialized with our task agnostic representations, and then tuned on supervised relation extraction datasets, significantly outperform the previous methods on Se-mEval 2010 Task 8, KBP37, and TACRED.

Superbloom: Bloom filter meets Transformer

“…This vocabulary size is two orders of magnitude larger than in previous work that applies a Transformer model with full softmax loss [8,23,18]. Other works, such as [24] and [17], train a Transformer model with a large number of entities using sampled softmax, with either in-batch or in-example negative sampling. But as we shall show, sampled softmax, even with a large number of 128K negative samples, results in much worse quality.…”

Semi-supervised Relation Extraction via Incremental Meta Self-Training

“…For an input sentence containing a relation mention, two entities Entity 1 and Entity 2 are marked in advance. We follow the labeling mechanism adopted by Soares et al (2019) and Zhang and Wang (2015) to enhance the position information of entities. For each sentence X = [x 1 , .., x T ], four reserved tokens…”

Inconsistent Few-Shot Relation Classification via Cross-Attentional Prototype Networks with Contrastive Learning

“…FewRel 2.0 (Gao et al, 2019b) extends the dataset with few-shot domain adaption and few-shot none-of-the-above detection. Many works on FewRel datasets focus on improvement of methods, including modeling the distance distribution (Gao et al, 2019a;Ding et al, 2021), utilizing external knowledge such as knowledge graph (Qu et al, 2020), learning different level of features (Sun et al, 2019;Ye and Ling, 2019), and using pre-trained language models (Soares et al, 2019). Apart from method innovation on the standard setting of consistent few-shot RC, the investigation for inconsistent few-shot RC still in demand.…”