Relation Extraction with Temporal Reasoning Based on Memory Augmented Distant Supervision

Proceedings of NAACL-HLT 2019, pages 1019–1030 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1019 Relation Extraction with Temporal Reasoning Based on Memory Augmented Distant Supervision Jianhao Yan1, Lin He1, Ruqin Huang2, Jian Li3∗, and Ying Liu1 1Institute for Network Sciences and Cyberspace, Tsinghua University 2Department of Computer Science and Technology, Tsinghua University 3Institute for Interdisciplinary Information Sciences, Tsinghua University {yjh16, he-l14, hrq16}@mails.tsinghua.edu.cn lijian83@tsinghua.edu.cn, liuying@cernet.edu.cn Abstract Distant supervision (DS) is an important paradigm for automatically extracting rela- tions. It utilizes existing knowledge base to collect examples for the relation we intend to extract, and then uses these examples to au- tomatically generate the training data. How- ever, the examples collected can be very noisy, and pose significant challenge for obtaining high quality labels. Previous work has made remarkable progress in predicting the rela- tion from distant supervision, but typically ig- nores the temporal relations among those su- pervising instances. This paper formulates the problem of relation extraction with tem- poral reasoning and proposes a solution to predict whether two given entities participate in a relation at a given time spot. For this purpose, we construct a dataset called WIKI- TIME1 which additionally includes the valid period of a certain relation of two entities in the knowledge base. We propose a novel neu- ral model to incorporate both the temporal in- formation encoding and sequential reasoning. The experimental results show that, compared with the best of existing models, our model achieves better performance in both WIKI- TIME dataset and the well-studied NYT-10 dataset. 1 Introduction As an important technique to automatically com- plete the knowledge base and reduce labeling ef- forts, distant supervision (DS) for relation ex- traction has drawn much attention. In DS, we align the entity pair (head, tail) from a triple ⟨head, rel, tail⟩ extracted from a huge knowledge base (e.g., Freebase, Wikidata) with sentences from free texts (e.g., Wikipedia, New York Times) ∗Jian Li is the corresponding author. 1https://github.com/ElliottYan/DS_ Temporal to obtain the training examples, and the label of such an example is the corresponding relation rel. Therefore, DS can automatically create a set of training data for each entity pair. However, the noisy training data problem (Riedel et al., 2010) significantly affects the per- formance of DS. Therefore, most of the recent approaches (Riedel et al., 2010; Hoffmann et al., 2011; Zeng et al., 2015; Lin et al., 2016) follow a common assumption called the at-least-once as- sumption, which treats all aligned sentences of each entity pair as one training sample. We re- fer to a sentence as an instance and all sentences aligned to one entity pair as a mention set in the following, respectively. The models in previous work (Zeng et al., 2015; Lin et al., 2016; Luo et al., 2017) generally include two parts, encoding and fusion. The former en- codes each instance into a low-dimensional repre- sentation. The latter combines representation of each instance. Then, their combination is used to predict the relation. Although the approaches mentioned …

Content truncated to 500 words