Who is Killed by Police: Introducing Supervised Attention for Hierarchical LSTMs

Proceedings of the 27th International Conference on Computational Linguistics, pages 2277–2287 Santa Fe, New Mexico, USA, August 20-26, 2018. 2277 Who is Killed by Police: Introducing Supervised Attention for Hierarchical LSTMs Minh Nguyen† and Thien Huu Nguyen#‡ † Hanoi University of Science and Technology, Hanoi, Vietnam # Montreal Institute for Learning Algorithms, University of Montreal, Canada ‡ Department of Computer and Information Science, University of Oregon, USA minh.nv142950@sis.hust.edu.vn,thien@cs.uoregon.edu Abstract Finding names of people killed by police has become increasingly important as police shootings get more and more public attention (police killing detection). Unfortunately, there has been not much work in the literature addressing this problem. The early work in this field (Keith et al., 2017) proposed a distant supervision framework based on Expectation Maximization (EM) to deal with the multiple appearances of the names in documents. However, such EM-based framework cannot take full advantages of deep learning models, necessitating the use of hand- designed features to improve the detection performance. In this work, we present a novel deep learning method to solve the problem of police killing recognition. The proposed method relies on hierarchical LSTMs to model the multiple sentences that contain the person names of interests, and introduce supervised attention mechanisms based on semantical word lists and dependency trees to upweight the important contextual words. Our experiments demonstrate the benefits of the proposed model and yield the state-of-the-art performance for police killing detection. 1 Introduction We study the problem of police killing detection from text. The key challenge is to be able to take a per- son name in the pool of documents (corpus) and automatically decide whether the corresponding person is killed by police or not based on the textual evidences in the corpus. For instance, the following sen- tence describes the police-caused death of “Micah Jester”: “Old Micah Jester was fatally shot by APD officers.”. This problem has drawn much public attention recently; however, it has not been investigated adequately by the natural language processing (NLP) community. To our knowledge, the only NLP work for police killing recognition so far is by (Keith et al., 2017) who rely on machine learning models to perform the automatic detection. It is challenging to apply the machine learning models in this case as identifying police killings from text is a relatively new problem in machine learning research with no available training datasets to super- vise the models. The only sources of supervision on which we can rely for this problem are the current databases that record the names of the police-killed victims in the past. Among these databases, the Fatal Encounters1 (FE) database has emerged as the most comprehensive database with a relatively large number of recorded victims (over 23,000 victims). In order to take advantage of this database, (Keith et al., 2017) employs distant supervision (Craven et al., 1999; Mintz et al., 2009) that extracts person names from a corpus and aligns them with the victim names in the database. The matched names are considered as corresponding to people …

Content truncated to 500 words