NewsReader at SemEval-2018 Task 5: Counting events by reasoning over event-centric-knowledge-graphs

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 660–666 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics NewsReader at SemEval-2018 Task 5: Counting events by reasoning over event-centric-knowledge-graphs Piek Vossen VU University Amsterdam / De Boelelaan 1105, 1081HV Amsterdam, Netherlands piek.vossen@vu.nl Abstract In this paper, we describe the participation of the NewsReader system in the SemEval- 2018 Task 5 on Counting Events and Par- ticipants in the Long Tail. NewsReader is a generic unsupervised text processing sys- tem that detects events with participants, time and place to generate Event Centric Knowl- edge Graphs (ECKGs). We minimally adapted these ECKGs to establish a baseline perfor- mance for the task. We first use the ECKGs to establish which documents report on the same incident and what event mentions are coref- erential. Next, we aggregate ECKGs across coreferential mentions and use the aggregated knowledge to answer the questions of the task. Our participation tests the quality of News- Reader to create ECKGs, as well as the po- tential of ECKGs to establish event identity and reason over the result to answer the task queries. 1 Introduction This paper describes the NewsReader system par- ticipating in the SemEval 2018 Task 5 Counting Events and Participants in the Long Tail (Postma et al., 2018). Task 5 requires detection of certain events (cases of gun violence, dismissal of em- ployees, and burning fire incidents) with partici- pants, as well as extraction of their location and time in a set of documents and reasoning over their identity, in order to answer queries over the data set. A typical query in the task is How many peo- ple were killed in 2016 in Columbus, MS?. Partic- ipants were given a collection of news articles in CoNLL format to distill the answer to the queries. The task consists of 3 subtasks: subtask 1 asks for all documents from the data set that report on a sin- gle incident that fits the question constraints; sub- task 2 is to provide the number of incidents (zero or more) and the documents that report on these incidents given the constraints of a query; subtask 3 asks for the number of affected (injured, dead, or fired) people in the incidents that match the query constraints, and the supporting documents for these incidents. In addition to answering the subtask queries, participants were asked to mark the (cross-document) coreferential event mentions in the CoNLL file according to a specific event schema for gun violence. Since we also organized the task, we decided to participate out-of-competition. Our system is a version of the NewsReader system (Vossen et al., 2016) as it was delivered at the end of the project, with as little adaptation as possible to the pro- cessing of the text to answer the queries of the task. The generic NewsReader system created the semantic output by applying a deep reading ap- proach to the text, and the tasks were addressed by loading that output and reasoning over the …

Content truncated to 500 words