Weakly Supervised Multilingual Causality Extraction from Wikipedia

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2988–2999, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2988 Weakly Supervised Multilingual Causality Extraction from Wikipedia Chikara Hashimoto Yahoo Japan Corporation 1-3 Kioicho, Chiyoda-ku, Tokyo 102-8282, Japan chashimo@yahoo-corp.jp Abstract We present a method for extracting causality knowledge from Wikipedia, such as Protec- tionism → Trade war, where the cause and effect entities correspond to Wikipedia arti- cles. Such causality knowledge is easy to ver- ify by reading corresponding Wikipedia arti- cles, to translate to multiple languages through Wikidata, and to connect to knowledge bases derived from Wikipedia. Our method ex- ploits Wikipedia article sections that describe causality and the redundancy stemming from the multilinguality of Wikipedia. Experiments showed that our method achieved precision and recall above 98% and 64%, respectively. In particular, it could extract causalities whose cause and effect were written distantly in a Wikipedia article. We have released the code and data for further research. 1 Introduction Much of the world consists of entities that causally depend on each other. Therefore, causality knowl- edge, e.g., Protectionism → Trade war,1 is useful for many tasks such as why-QA (Oh et al., 2017), reading comprehension (Berant et al., 2014), and event prediction (Radinsky et al., 2012). Although many methods have been proposed for causality extraction from text (Ning et al., 2018; Gao et al., 2018; Kruengkrai et al., 2017; Rehbein and Ruppenhofer, 2017; Dunietz et al., 2017; Hidey and McKeown, 2016; Zhao et al., 2016; Hashimoto et al., 2014, 2012; Do et al., 2011; Riaz and Girju, 2010; Abe et al., 2008), they have rarely addressed three issues that are impor- tant for constructing a causality knowledge base (CKB). First, we should be able to verify extracted causalities, so that the CKB can sustain the credi- bility of its information.2 Second, it would be de- 1In this paper, A → B denotes that A causes B. 2The verifiability in this paper means that one can verify extracted causalities with credible sources of information. sirable to easily translate the CKB to multiple lan- guages, to avoid duplicating the construction effort for different languages. Third, it would also be de- sirable to automatically connect the CKB to other knowledge bases (KBs), to bring together KB con- struction efforts in various communities and thus maximize their synergistic effect. Therefore, we propose a method for extracting causalities from Wikipedia by using cause and ef- fect entities that correspond to Wikipedia articles; for example, for Tobacco → Lung cancer, English Wikipedia has articles titled Tobacco and Lung cancer. Such causalities satisfy the above three desiderata. First, we can easily verify such causal- ities, because Wikipedia articles tend to credibly attest them; for example, the Tobacco article states that inhaling its smoke can cause Lung cancer.3 In contrast, knowledge from other sources such as the web text tends to be difficult to verify, owing to a deluge of false …

Content truncated to 500 words