Reporting the Unreported: Event Extraction for Analyzing the Local Representation of Hate Crimes

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5753–5757, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5753 Reporting the Unreported: Event Extraction for Analyzing the Local Representation of Hate Crimes Aida Mostafazadeh Davani, Leigh Yeh, Mohammad Atari, Brendan Kennedy Gwenyth Portillo-Wightman, Elaine Gonzalez, Natalie Delong, Rhea Bhatia, Arineh Mirinjian, Xiang Ren, Morteza Dehghani University of Southern California Abstract Official reports of hate crimes in the US are under-reported relative to the actual number of such incidents. Further, despite statistical ap- proximations, there are no official reports from a large number of US cities regarding incidents of hate. Here, we first demonstrate that event extraction and multi-instance learning, applied to a corpus of local news articles, can be used to predict instances of hate crime. We then use the trained model to detect incidents of hate in cities for which the FBI lacks statistics. Lastly, we train models on predicting homi- cide and kidnapping, compare the predictions to FBI reports, and establish that incidents of hate are indeed under-reported, compared to other types of crimes, in local press. 1 Introduction Hate crimes are defined as crimes of violence ei- ther against a person or their property that dis- play evidence of prejudice based on the victims’ race, gender or gender identity, religion, disabil- ity, sexual orientation, or ethnicity (Jacobs et al., 1998). According to the results of a new Depart- ment of Justice hate crime report released in 2017 (Masucci and Langton, 2017), approximately 54% of hate crime victimizations were not reported to police during 2011-2015. Despite the recent ef- forts of advocacy groups, policy makers, and re- searchers to create reliable, national data to under- stand the extent and severity of hate crime victim- ization, the existing estimates continue to fall short (Pezzella et al., 2019). It stands to reason that hate crimes induce lo- cal disturbance, and as a result, might be likely to get local coverage. Therefore, local news agen- cies can be considered a unique source of infor- mation for detecting these incidents. Here, we use a corpus of local news articles, collected from the Patch1 website. The Patch data contain indepen- dent, hyper-local news articles compiled from lo- cal news sites. We apply event extraction methods to identify incidents of hate crime reported in the Patch cor- pus for cities with no representation in FBI reports, and analyze the frequency of the extracted events compared to the number of incidents reported by the FBI. The task of labeling each article as a hate crime or not is defined as a Multi-Instance Learn- ing (MIL) problem since each article is modeled as a sequence of sentences. Instead of predicting a la- bel for each sentence, we use the information em- bedded in all the sentences of an article to deter- mine whether the article is reporting a hate crime. After testing the model on a set of annotated …

Content truncated to 500 words