Using Computational Linguistics to Enhance Protest Event Analysis

Abstract

For now more than four decades, quantitative protest event analysis (PEA) has routinely contributed to the testing and refinement of theories on political processes from different perspectives. However, it is commonly agreed that PEA data face serious challenges regarding their data sources. Precisely, researchers applying PEA struggle with the fact that they cannot use multiple sources for large geographical areas and long time periods. As a consequence, most of the scholarship still focuses on a narrow set of European countries or the United States and does not cover the period since the early 2000s. We are bringing PEA and computational linguistics together to suggest and evaluate an approach that will enable political scientists to extend their research designs with a more efficient and at the same time reliable data collection. The approach relies on hidden topic models, word space models, and named entity recognition to identify and code protest events.

Black Protest Events In The Us 1994-2010: Issues, Campaigns And Trends

“…Machine-assisted approaches to political event data have been in use for nearly 30 years, since the inception of the Kansas Event Data System (KEDS) (Schrodt and Gerner 1994). More recently, there have been several approaches which incorporate machine learning methods into their pipelines (Croicu and Weidmann 2015, Marakov et al 2015, Nardulli, Althaus and Hayes 2015, Wueest, Rothenhäusler and Hutter 2013. MPEDS differs from other automated event data projects because it focuses on coding for protest events rather than for a wider range of political events and because it aims to collect rich information about each event.…”

Protest Event Analysis and Its Offspring

“…More advanced technologies rely on text classifiers that usually work on word frequency models. First tests show that such techniques perform quite well and clearly reduce the workload involved in the selection of articles (see Wüest et al 2013). While the half-automated selection of protest events from digital text sources works relatively well and can also be implemented quite easily in smaller research projects, the half-automated coding of events is still mainly restricted to English sources and to highly standardized types of texts (e.g., the titles of news agency reports, as used by some projects in the third generation of PEA, see Section 2).…”

A Pipeline for Multilingual Protest Event Selection and Annotation

“…[12] aggregate English articles ("The Guardian" archive) using a pre-defined keywords list via the Nexis engine, which yields poor results (only 68 out of 727 crawled articles address protest events), therefore, a postselection is needed. To this end, an active learning classifier is applied that uses as input the depency parsing output of the UIMA framework 10 , and hidden topic modelling (HTM) performed with the mallet toolkit 11 .…”

“…It started with the manual collection of event datasets in order to find reliable patterns of contentious collective behaviour and accumulate statistics for protest prediction and the analysis of its origins, dynamics and aftermath. However, human-based event collection turns out to be subjectivityprone [10] and insufficient in terms of source, time and location coverage [12]. Even for a single event, it is too time-consuming to perform cross-lingual analysis to complete the picture on its origins, consequences, actors involved, and parallels to other similar events.…”

“…The rest of the paper is organized as follows: Section 2 gives a short overview of CAMEO and SSP ontologies-based extraction systems, as well as of the two early prototypes by [5] and [12]. Section 3 gives details on our generic pipeline, and Section 4 presents the evaluation of event selection and annotation.…”