Machine Reading of Historical Events

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7486–7497 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7486 Machine Reading of Historical Events Or Honovich1,* Lucas Torroba Hennigen2,* Omri Abend1 Shay B. Cohen3 1The Hebrew University of Jerusalem 2University of Cambridge 3University of Edinburgh or.honovich@gmail.com lft26@cam.ac.uk oabend@cs.huji.ac.il scohen@inf.ed.ac.uk Abstract Machine reading is an ambitious goal in NLP that subsumes a wide range of text understand- ing capabilities. Within this broad framework, we address the task of machine reading the time of historical events, compile datasets for the task, and develop a model for tackling it. Given a brief textual description of an event, we show that good performance can be achieved by extracting relevant sentences from Wikipedia, and applying a combination of task- specific and general-purpose feature embed- dings for the classification. Furthermore, we establish a link between the historical event ordering task and the event focus time task from the information retrieval literature, show- ing they also provide a challenging test case for machine reading algorithms.1 1 Introduction Machine reading concerns the extraction of entities and relations from text and the ability to use them meaningfully, for instance by answering questions based on them, inferring other relations from them, or using them to compile knowledge bases. Such an inclusive task definition necessarily builds on a wide range of NLP capabilities, from syntactic and semantic analysis, to the use of world knowledge and common sense. The inclusive nature of the task supports the development of general-purpose meth- ods, but also results in low performance in absolute terms, difficulty in defining widely agreed-upon evaluation protocols, and difficulties identifying the sources of prediction errors (Stanovsky and Dagan, 2016; Rajpurkar et al., 2016; Clark et al., 2018). This paper addresses a sub-task of machine read- ing, namely the task of estimating when a historical 1Code and data are available at https://github.com/ltorroba/ machine-reading-historical-events. *Equal contribution. Year Event text OTD 2005 107 die in Amagasaki rail crash in Japan. 1939 BMI (Broadcast Music Incorporated) formed. 1864 General Sherman’s armies reach Savannah & 12 day siege begins. WOTD 1887 Buffalo Bill Cody’s Wild West Show opens in Lon- don. 1399 Henry IV is proclaimed King of England. 1943 First Flight of the Gloster Meteor, Britain’s first combat jet aircraft. Table 1: Entries from the OTD and WOTD datasets. event took place. This distinguishes it from tradi- tional question answering (Rajpurkar et al., 2016) as the answer may not be given in the text but the models should still be able to place events in the correct period of time. In turn, this means that models trained for historical event ordering may have real-world applications such as to serve as a fallback for temporal question answering when the answers are not present in the text and to improve search engines that leverage the implicit time of queries (Gupta and Berberich, 2016). Concretely, given a short text description of a historical event, and an external data source (hence- forth contextual information …

Content truncated to 500 words