KOI at SemEval-2018 Task 5: Building Knowledge Graph of Incidents

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 81–87 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics KOI at SemEval-2018 Task 5: Building Knowledge Graph of Incidents Paramita Mirza,1 Fariz Darari,2∗ Rahmad Mahendra2∗ 1 Max Planck Institute for Informatics, Germany 2 Faculty of Computer Science, Universitas Indonesia, Indonesia {paramita}@mpi-inf.mpg.de {fariz,rahmad.mahendra}@cs.ui.ac.id Abstract We present KOI (Knowledge of Incidents), a system that given news articles as input, builds a knowledge graph (KOI-KG) of incidental events. KOI-KG can then be used to effi- ciently answer questions such as “How many killing incidents happened in 2017 that involve Sean?” The required steps in building the KG include: (i) document preprocessing involv- ing word sense disambiguation, named-entity recognition, temporal expression recognition and normalization, and semantic role labeling; (ii) incidental event extraction and coreference resolution via document clustering; and (iii) KG construction and population. 1 Introduction SemEval-20181 Task 5: Counting Events and Par- ticipants in the Long Tail (Postma et al., 2018) ad- dresses the problem of referential quantification that requires a system to answer numerical ques- tions about events such as (i) “How many killing incidents happened in June 2016 in San Antonio, Texas?” or (ii) “How many people were killed in June 2016 in San Antonio, Texas?” Subtasks S1 and S2 For questions of type (i), which are asked by the first two subtasks, partic- ipating systems must be able to identify the type (e.g., killing, injuring), time, location and partic- ipants of each event occurring in a given news article, and establish within- and cross-document event coreference links. Subtask S1 focuses on evaluating systems’ performances on identifying answer incidents, i.e., events whose properties fit the constraints of the questions, by making sure that there is only one answer incident per question. ∗ Both share the same amount of work. 1http://alt.qcri.org/semeval2018/ Subtask S3 In order to answer questions of type (ii), participating systems are also required to identify participant roles in each identified an- swer incident (e.g., victim, subject-suspect), and use such information along with victim-related nu- merals (“three people were killed”) mentioned in the corresponding answer documents, i.e., docu- ments that report on the answer incident, to deter- mine the total number of victims. Datasets The organizers released two datasets: (i) test data, stemming from three domains of gun violence, fire disasters and business, and (ii) trial data, covering only the gun violence domain. Each dataset contains (i) an input document (in CoNLL format) that comprises news articles, and (ii) a set of questions (in JSON format) to evaluate the participating systems.2 This paper describes the KOI (Knowledge of Incidents) system submitted to SemEval-2018 Task 5, which constructs and populates a knowl- edge graph of incidental events mentioned in news articles, to be used to retrieve answer incidents and answer documents given numerical questions about events. We propose a fully unsupervised approach to identify events and their properties in news texts, and to resolve within- and cross- document event coreference, which will be de- tailed in the …

Content truncated to 500 words