EMBERS AutoGSR: Automated Coding of Civil Unrest Events

EMBERS AutoGSR:
Automated Coding of Civil Unrest Events
Parang Saraf
Discovery Analytics Center
Department of Computer Science
Virginia Tech
parang@cs.vt.edu
Naren Ramakrishnan
Discovery Analytics Center
Department of Computer Science
Virginia Tech
naren@cs.vt.edu
ABSTRACT
We describe the EMBERS AutoGSR system that conducts
automated coding of civil unrest events from news articles
published in multiple languages. The nuts and bolts of the
AutoGSR system constitute an ecosystem of filtering, rank-
ing, and recommendation models to determine if an article
reports a civil unrest event and, if so, proceed to identify and
encode specific characteristics of the civil unrest event such
as the when, where, who, and why of the protest. AutoGSR
is a deployed system for the past 6 months continually pro-
cessing data 24x7 in languages such as Spanish, Portuguese,
English and encoding civil unrest events in 10 countries of
Latin America: Argentina, Brazil, Chile, Colombia, Ecuador,
El Salvador, Mexico, Paraguay, Uruguay, and Venezuela. We
demonstrate the superiority of AutoGSR over both manual
approaches and other state-of-the-art encoding systems for
civil unrest.
Keywords
event extraction, event encoding, text mining
1. INTRODUCTION
The computational modeling and interpretation of soci-
etal events has been a holy grail in social science research.
Beginning in the 1980s, there have been significant efforts
in computational analysis of societal events supported by
government programs such as DARPA’s ICEWS (Integrated
Conflict Early Warning System) and CIA’s PITF (Political
Instability Task Force). Projects of similar (and more ambi-
tious) scope continues to this day, and offer greater specificity,
both spatially and temporally into modeling global events.
We are part of the EMBERS consortium [13] that aims to
forecast civil unrest phenomena (protests, strikes, and ‘oc-
cupy’ events) in multiple countries of Latin America, specif-
ically Argentina, Brazil, Chile, Colombia, Ecuador, El Sal-
vador, Mexico, Paraguay, Uruguay, and Venezuela. In our
earlier KDD 2014 paper [13] we demonstrated how we can
use open source indicators such as news, blogs, Twitter, food
prices, and economic data, to forecast civil unrest events.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
KDD ’16, August 13-17, 2016, San Francisco, CA, USA
c 2016 ACM. ISBN 978-1-4503-4232-2/16/08...$15.00
DOI: http://dx.doi.org/10.1145/2939672.2939737
EMBERS forecasts have been evaluated by a third party
(MITRE) wherein human analysts prepared a ground truth
dataset (called the GSR, or ‘Gold Standard report’) of re-
ported protests in newspapers of record. The GSR is com-
pared against EMBERS forecasts using metrics introduced
in [13].
As the EMBERS project matures, we realized that we
must pay attention to not just forecasting events but also to
coding ongoing events, i.e., the process of constructing the
GSR on a regular basis. For instance, see Fig. 2. Such coded
data serves two uses in EMBERS: to help evaluate EMBERS
forecasts, and to support the regular re-training and tuning
of the machine learning models. Accordingly, we launched a
parallel effort, referred to as the EMEBRS AutoGSR, that
conducts automated coding of civil unrest events from news
articles published in multiple languages. Like EMBERS, the
AutoGSR is also a deployed system continually processing
data 24x7 in languages such as Spanish, Portuguese, English.
Our key contributions can be summarized as follows:
1. To the best of our knowledge, the AutoGSR is the
only/first system to be able to automatically encode
civil unrest events across 10 countries in languages
local to these countries. This gives a significant advan-
tage over systems that are manual or systems that are
automatic but restricted to English.
2. The nuts and bolts of the AutoGSR system constitute
an ecosystem of filtering, ranking, and recommendation
models to determine if an article reports a civil unrest
event and, if so, proceed to identify and encode specific
characteristics of a civil unrest event such as the when,
where, who, and why of the protest. We present an
exhaustive evaluation of the performance of AutoGSR
using metrics in the large (e.g., does the system track
ongoing happenings in countries of interest?) as well
as metrics in the small (e.g., does the system identify
specific events of interest?).
3. AutoGSR is meant to be used in both a fully auto-
mated and an analyst assisted mode. Through detailed
analysis of hours logged in both the manual process and
in the AutoGSR system, we quantify the performance
gains of our approach.
2. RELATED WORK
The challenges associated with a system like AutoGSR
can be broadly classified in two categories: event encoders
and event databases. While the event encoders are not freely
available, the event databases constructed around them are
more available.

Table 1: Sample erroneous encodings by ICEWS and GDELT.
Source ID Representative Paragraph Reason for Error
GDELT 299144197
Pope Francis is hoping to demonstrate the power of prayer
next week when Israeli President Shimon Peres and Palestinian
President Mahmoud Abbas join the pontiff at the Vatican for
an exercise in peace building.
Presence of the word ‘demonstrate’
results in a false positive.
GDELT
256666928 and
256814375
U.N. Secretary-General Ban Ki-moon on Tuesday expressed
alarm at the violence in Turkey as confrontations between Turk-
ish security forces and protesters continued after three weeks of
demonstrations against Prime Minister Tayyip Erdogan.
No Duplicate Detection. Two
events extracted from two articles
covering the same story but
published on different days.
ICEWS 23909784
Since its inception, the Islamic State group has demonstrated
the firmness of its structure and the strength of its organizational
composition.
Presence of word ‘demonstrate’ re-
sults in a false positive.
ICEWS 19873295
Capriles has called off a march by his supporters in Caracas,
saying that his rivals were plotting to ”infiltrate” the rally to
trigger violence.
The protest was called off.
Figure 1: The EMBERS system forecasts civil unrest
events from open source indicators while the EMBERS
AutoGSR system encodes civil unrest events as they are re-
ported in newspapers. Output from the AutoGSR system
is used to both evaluate EMBERS forecasts and to retrain
EMBERS models on a periodic basis.
Table 2: Comparison of automated event encoders.
System
Protest
Encoding
Language
Agnostic
Tunability
Duplicate
Detection
ICEWS
GDELT
AutoGSR
2.1 Event Encoders
Hogenboom et. al.[3] provides an overview of different
extraction methodologies used by the current state-of-the-art
systems. The methodologies used here include statistical as
well as linguistic and lexicographical approaches for event
extraction. TABARI (Textual Analysis by Augmented Re-
placement Instructions) [15] and BBN’s SERIF (Statistical
Entity and Relation Information Finder) [1, 14] are two state-
of-the-art event encoders. These encoders not only extract
events but also encode them via a mapping to an event tax-
onomy. Such mappings add structure to the extracted event
thereby making it feasible to perform systematic studies.
CAMEO [16] is one such widely used event taxonomy and is
used by both TABARI and SERIF.
TABARI is one of the earlier open source event extraction
systems that uses sparse parsing to recognize patterns in
Figure 2: A sample event extracted from a news article.
text. These patterns are hand coded and identify three types
of information: actors, verbs, and actions. For a given text,
only a few initial sentences are used for event extraction,
to support high throughput applications. Several improved
versions of TABARI have been proposed. JABARI is one
such system, which is a Java implementation of TABARI and
uses a few advanced NLP techniques in addition to pattern
matching. More recently, PETRARCH has emerged as the
successor of TABARI. Instead of conducting a pattern based
extraction, PETRARCH uses the full parsed Penn TreeBank
as input to perform a parser-based encoding.
BBN’s SERIF is another state-of-the-art event encoder
that uses a series of NLP components to capture represen-
tations of type ‘who did what to whom’ in article text. The
encoder works at both the sentence and document level and is
able to identify and resolve entities between sentences. Once
the entities are resolved, the encoder detects and character-
izes the relationship between entities. Finally the encoder
maps these relationships to the CAMEO taxonomy using
an externally provided list of actor dictionaries and event
patterns.
2.2 Event Databases
The origins of automated event databases can be attributed
primarily to political scientists and intelligence agencies who
over the years have envisioned systems that can perform large
scale encoding of events by mining million of news articles.
ICEWS (Integrated Crisis Early Warning System) [11] and
GDELT (Global Database of Events, Language and Tone)[6]

are two such systems that analyze hundreds of news sources
from all over the world in order to generate a database of
events.
ICEWS, which began in 2007, is a DARPA funded project
that focuses primarily on monitoring, accessing and forecast-
ing events of interests for military commanders. Internally,
ICEWS employs TABARI and SERIF to encode news arti-
cles. ICEWS focuses primarily on generating high quality,
reliable events and uses several mechanisms to filter the raw
stream of reported stories into a unique stream of events.
Events are encoded in accordance to the CAMEO taxonomy.
GDELT, on the other hand focuses on capturing an exten-
sive set of events both in terms of categories and geographical
spread. By design, the goal of GDELT is to collect a large
number of events without worrying about false positives.
Internally, it uses an enhanced version of TABARI and maps
events to the CAMEO taxonomy.
Although both these systems are considered state-of-the-
art, in our experiments we have found that they perform
poorly in comparison to manually generated ground truth
data. See Table 1 for examples of erroneous encodings in
ICEWS and GDELT. This is primarily because of the fully
automated event generation process that yields false positives.
Hence, there is a need for a semi-automated system that can
generate validated event encodings with minimal human
effort. Table 2 contrasts ICEWS and GDELT against our
AutoGSR system.
3. SYSTEM ARCHITECTURE
EMBERS AutoGSR is a web based system that generates
validated civil unrest events extracted from news articles.
The system architecture is shown in Fig. 3.
3.1 Data Sources
News articles are collected every day from three data
sources: 1) Site specific Google Search 2) RSS feeds sub-
scribed to individual news websites, and 3) News databases.
Articles are filtered if they contain a protest related keyword.
With the help of subject matter experts, we created a list of
protest keywords for each language.
3.2 Data Processing
This stage of EMBERS AutoGSR is responsible for running
and managing the sophisticated models ecosystem, along with
performing standard tasks like data cleaning and enrichment.
3.2.1 Fetch and Clean
The fetch and clean component fetches the article from
the original web source and performs a boilerplate removal
in order to extract the full article text. On this extracted
full text, keyword based search is performed once again to
make sure that at least one protest related keyword is present
in the text. This component also fetches SEO meta tags,
if present for each article. (These meta tags are added by
news websites in order to improve their search ranking and
includes several key information about the article such as
publish date, keywords, summary, and description.)
3.2.2 Data Enrichment
This component enriches news articles by performing vari-
ous kinds of linguistic processing. The enrichment comprises
named entity extraction, lemmatization, location identifica-
tion, geo-reference resolution, and translation of article text
Figure 3: EMBERS AutoGSR System Architecture.
to English. This enriched information is used by models in
the models ecosystem.
Words or noun phrases identified as a location by the
named entity resolver are passed through a geo reference
resolver to resolve ambiguity, if any. Many a times, loca-
tions are referred by their alternate name(s), for example:
LA for City of Los Angeles or America/US/USA for The
United States of America. Alternatively, several times, fa-
mous landmarks like Times Square, White House, etc are
used to denote a location. The geo-reference resolver goes
through a world gazetteer1
to resolve these phrases and maps
them to the official name.
Please note that articles are translated into english only to
enhance the readability of an end user. The translated text
is not used by any of the models. The models are designed
to work with native language.
3.2.3 Models Ecosystem
This component, explained in much detail in a subsequent
section, applies several machine learning models on news
articles. A first set of models classify an incoming article
into protest or non protest. If the article is classified as a
protest article, then it passes through a second set of models
that recommend event encoding(s) for the article.
1
GeoNames: http://www.geonames.org/

Filtering Models. These models are based on hand coded
rules that perform a hard classification into protest or non
protest. The models are applied in a series, where if the
article fails one of the models, the article is classified as a
non protest article. When an article successfully passes all
the filtering models, it is processed through the subsequent
ranking and recommendation models.
Ranking Models. These models work independently of each
other and assign a probability score for classifying the article
as a protest article. The individual probability scores gener-
ated by each model are fed to an ensembler which generates
a representative ensembled probability score for the article.
The accuracy of individual models in the past is also used as
an input by the ensembler.
Recommendation Models. The recommendation models
assume that the article is a protest article and based on
this assumption extracts & recommends both full event en-
coding(s) and individual encoding components. Elaborating
on what was mentioned earlier, a full event encoding com-
prises of the following components: 1) protest location, 2)
protest date, 3) participating population group, 4) reason
for protest, 5) violent or peaceful protest, and 6) protest
reported date. These models work in tandem to generate the
full event encoding. If there are multiple models generating
recommendations for a particular encoding component then
an ensembler is used to determine a representative value.
3.3 Database
The output of the data processing component containing
clean, enriched news articles along with the individual and
ensembled values generated by the model ecosystem is stored
in a database. This database acts as the primary data
source for the AutoGSR Web Interface which displays news
articles along with the associated data. For each article, the
database also stores validated encoding results generated by
the manual validation process.
3.4 Web Interface
The web interface shown in Fig. 4 displays enriched news
articles; output from the models ecosystem; and controls
to validate recommended encodings. More specifically the
interface displays the following (the numerical labeling of the
components in the image corresponds with the list numbers
below):
1. Allows the user to specify the filtering criteria for dis-
playing news articles. The sytem also allows a user
to specify filtering criteria for displaying news articles.
An important criterion is the cutoff probability with
values between 0 and 1 and is used to classify an ar-
ticle as protest or non-protest. If the representative
ensembled probability of an article, generated by the
ranking models ensembler, is greater than this cutoff
probability, then the article is classified as a protest
article; otherwise it gets labeled as a non protest article.
Using this control, the user can also tune the precision
and recall of the system.
2. The articles satisfying the filtering criteria are clustered
in real time to generate news clusters. Each news
cluster brings together related stories. For each cluster,
a representative label is also generated. The real time
clustering is performed using the Lingo3G clustering
suite2
. A summary text is also generated for each
article in the cluster, using the description meta tag.
3. For each article, the detailed view shows the full article
text, article image and translated english text along
with recommended event encoding and the output of
the models ecosystem as described next.
4. Event encodings are recommended in following three
ways – a) Ensembled Recommendations: These full
encoding recommendations are generated by putting
together the recommendations for individual encod-
ing components as generated by the recommendation
models from the ecosystem. In the case of automated
event extraction, these recommended encodings are
considered as the extracted encoding for the article.
b) Clustering Based Recommendations: Full encoding
recommendations are generated using related articles
with validated encodings in the dynamically generated
news clusters. These recommendations assume that
similar/related articles will result in almost similar
encodings. c) Individual Recommendations: The in-
dividual components in the encoding validation form
show component specific recommendations.
5. All sentences used by the system to generate recommen-
dations are highlighted for the user’s reference. There
is an option to toggle the reading view, where only the
highlighted sentences of the article are shown. The
‘Highlighted Text’ view assumes that, in the ideal sce-
nario, the highlighted sentences will provide complete
information about the article. Hovering the mouse over
the article shows the kind of information that the sys-
tem extracted from a particular sentence. If the user
doesn’t agree with the highlighted sentence or the kind
of information identified by it, then he can click on that
particular sentence and modify the information type
through the popover. This active feedback helps the
system to learn.
3.5 Encoding Validation
For each article classified as a protest article, recommended
encodings are validated manually by subject matter experts.
During the validation process, the experts can either accept
the recommended encoding or modify it. The validation
process is performed using the validation controls of the
interface. Encodings for each article are validated by at least
two analysts. If the analysts agree on the encoding then that
encoding is considered as the final encoding. However, in
case of a conflict, the final encoding is decided by a quality
control analyst.
3.6 Event Generation
Based on the validated records, a final set of events is
generated by performing duplicate detection. As there can
be multiple articles reporting the same protest, duplicate
detection is performed. During duplicate detection, unique
event encoding tuples comprising protest location, protest
date, protest reason, participating population group,
violence/peaceful are identified and resolved.
4. MODELS ECOSYSTEM
In civil unrest encoding, the ratio of positive to negative
articles is highly skewed towards negative articles. Based on
2
Lingo3G: https://carrotsearch.com/

Figure 4: EMBERS AutoGSR Web Interface. 1) Filtering Controls 2) Real time Clustering of News Articles 3) Full Text,
Translated Text 4) Full and Partial Encoding Recommendations 5) Model Outputs.
our experience, given a set of articles containing at least one
protest keyword, this ratio is somewhere around 1:8. Due
to this skewness, an analyst has to typically skim through a
huge number of negative articles just to identify a few protest
articles. What makes this problem even more challenging
is the inherent inability to characterize negative articles.
While we know what constitutes a protest article, there is
no simple way to describe a non protest article. Given these
challenges, the aim of the models ecosystem is two fold:
1) reduce the number of negative articles while minimizing
the loss of positive articles, and 2) generate event encoding
recommendations for positive articles.
Both of these are non-trivial problems and therefore the
solution requires an ensemble of models, each with its own
selective superiorities, focus area, and performance. The
models as shown in Table 3 and categorized into three com-
partments: 1) Filtering Models, 2) Ranking Models, and 3)
Recommendation Models.
4.1 Filtering Models
These are rules based models that perform hard classi-
fication of news articles into protest and non protest. If
the article fails even one of these models, then the article is
classified as a non protest article.
4.1.1 Sub Domain Based Filtering
Many of the sub domains, for example entertainment,
technology, sports, etc. are tagged as non relevant. If an
article is published in any of these sub domains then the
article is not considered relevant and gets classified as non
protest. In this context, a sub domain worth mentioning
is sports. Sports articles use a lot of protest keywords like
attack, surrounded, etc. to describe competition between two
teams/players.
4.1.2 URL Based Filtering
There are several URL structures that list multiple articles
on a single page. For example, URLs listing top stories
of the day3
, or URLs listing stories by topic4
, or URLs
corresponding to search terms5
, etc. All such URLs are
considered irrelevant for the this task, and any article with
URL matching one of these types is discarded.
4.1.3 Negative Keywords Based Filtering
For many of the protest keywords, there exist words, called
Negative Keywords which when used in the vicinity of a
protest word can completely alter the meaning. A few such
negative keywords are described in Table 4. If any such nega-
tive keyword is present in the vicinity of the protest keyword
in the same sentence, then such matches are not considered
as a positive keyword match and are ignored. After ignoring
all such false matches, if the article still contains any other
protest keyword, only then the article is allowed to pass.
Otherwise, the article is reported as a non protest article.
4.2 Ranking Models
Each of these models assign a probability score of classi-
3
http://www.clarin.com/politica/
4
http://www.clarin.com/tema/manifestaciones.html
5
http://www.clarin.com/buscador?q=protesta

Table 3: EMBERS AutoGSR: Models Ecosystem
Filtering Models Ranking Models Recommendations Models
These are rules based models that
classify incoming news articles into
protest and non protest with a 0 or 1
certainty
These models use standard machine learning
algorithms on different components of an incom-
ing article for classification.
These models assume that the incoming
article is a protest article and tries to rec-
ommend complete or partial encoding(s)
for the article
1. Sub Domain Based Filtering
2. URL Based Filtering
3. Negative Keywords Based Filtering
1. Labeled Unlabeled Text Classifier
2. Entity Based Classifier
3. Distributed Representation Classifier
4. MetaTags Based Classifier
Protest Identification Models:
1. Geo Location Recommender
2. Temporal Recommender
Protest Characterization Models:
3. Entity Based Na¨ıve Bayes
Recommender
4. MetaTags Based Recommender
Usability Models:
5. Clustering Based Recommender
Approach: Articles are passed
through these models sequentially. If
any of these models classifies the
article as non protest then the article
is labeled as a non protest article in
the interface
Approach: Each of these models assign individ-
ual probabilities to an incoming article. An arti-
cle’s final probability is calculated using ‘model
ensemble’ approach. In the interface user can
specify a cutoff probability score. Articles that
have probability greater than the cutoff will
appear as protest articles.
Approach: These recommendations ap-
pear in the interface for each article. The
recommendations are generated for both
full encodings and individual encoding
components.
Table 4: Example Negative Keywords
Protest
Keyword
Negative
Keyword Phrase
Altered Meaning
marcha ponar en marcha
to start;
to set in motion
protesta tomar protesta
to swear in
(public official)
protesta rendir protesta
to swear in
(public official)
fying the article as a protest article. The models focus on
different indicators for classifying the article.
4.2.1 Labeled Unlabeled Text Classifier
This model implements a LU Classifier as proposed by
[9, 10] to work with a set of labeled and unlabeled articles.
The model is adapted to work in a binary class setting and
implements an Expectation Maximization procedure with a
Na¨ıve Bayes Classifier.
4.2.2 Entity Based Classifier
This model is designed to work with named entities and
protest keywords and can be considered as a special case
of the NB classifier built earlier. Using extracted entities
and protest keywords as evidences, this model employs a
multinomial na¨ıve Bayes method with Laplace smoothing to
determine the posterior probability of classifying an article
as a protest article. The entities include name, location and
organization along with the protest keywords found in the
article.
4.2.3 Distributed Representation Classifier
This model takes a step away from the bag of words models
and aims to capture the ordering of words while generating
distributed vector representations for the articles. The model
builds upon recent advancements in learning vector repre-
sentations of words using neural networks [7, 8]. Mikolov
et al. show that such vector representations not only cap-
ture syntactic and semantic information but is also aware of
’distance’ in the N dimensional representation space. The
model was adapted further by [5] to generate distributed
representations of documents and was shown to be distance
aware similar to word vectors.
These models have been designed primarily to work with
large datasets and have been shown in [7] to perform better
with larger corpora, as they allow the layer weights to stabi-
lize. Therefore, we train this model on an unlabeled corpus
of ∼250k Spanish and ∼100k Portuguese news articles. To
increase relevance, we made sure that these articles contain
at least one protest keyword. To this set, we added ∼20k
labeled articles. Each of the labeled articles contains two
labels: 1) Unique label identifying the article; and 2) Class
label identifying the article as protest or non protest. This
way, we not only identify vector representations for each
document but also vector representations for labels ‘protest’
and ‘non protest’.
Once the distributed vector representations for ‘protest’
and ‘non protest’ labels have been identified, we classify
an incoming article as follows: 1) Generate a distributed
vector representation of the article using the trained model;
2) Calculate the cosine distance of the article vector from
protest and non protest labels; 3) Classify the article based
on closeness and assign a posterior probability.
4.2.4 MetaTags based Classifier
In the literature, much of effort in classification of web
pages or news articles has focused only on full article text.
Almost little to no work exists in the domain of using HTML
meta tags as input. HTML meta tags provide a wide variety
of structured information about a web page, which is primar-
ily used by search engine crawlers. In our scenario, tags of
much importance are the ones that are used to generate rich
text snippets. Rich text snippet is a short summary descrip-
tion of a web page that is shown when the page appears in
search engine results and/or social media websites. Generally,
this short description is manually provided by the author of
the news article to succinctly describe the article. Google
uses meta tags description and title, along with several other
latent parameters to generate snippets. Facebook has cre-
ated the Open Graph Protocol that describes guidelines for
meta tags. These guidelines allow content creators to control
the display of their content on social networking websites.

Similarly, Twitter which supports the open graph protocol,
also has its own guidelines for content sharing.
Table 5: Metatags used by AutoGSR.
Targeted For Tag Name
Search Engines title, description
Facebook og:title, og:description
Twitter twitter:title, twitter:description
Table 5 lists the various standard meta tags that we ex-
tract information from. In this model, we work on the text
extracted from only these tags. We train a vanilla SVM
classifier [2] with a linear kernel that has been shown to work
exceptionally well with text classification tasks [4].
4.2.5 Ensembler
The goal of the ensembler is to take output probabilities
from the individual ranking models and fuse them together to
generate a final representative probability of the article. The
representative probability score should provide a better clas-
sification accuracy than any of the models individually. We
generate the representative probability score by performing
a weighted average of the individual outputs, where weights
correspond to the accuracy of each model:
P(lp) =
|M|
i=i cioi
|M|
i=i ci
where, P(lp) is the probability of classifying the article as
protest article, ci is the accuracy and oi is the output of
model i. The accuracy of a model is derived empirically
based on manual validation.
4.3 Recommendation Models
The goal of the recommendation models is to generate
partial or complete encoding suggestions for a given article.
Broadly speaking these models can be grouped in three cate-
gories: 1) Protest Identification, 2) Protest Characterization,
and 3) Usability. The Protest Identification models aim to
identify when and where a protest happened: Geo Location
Recommender along with Temporal Recommender identify
city and date of protest. Once a protest has been identified,
the Protest Characterization models work on characterizing a
protest event by identifying the reason behind protest and the
participating population group. Entity Based Na¨ıve Bayes
Recommender and MetaTags based recommender are two
such models. The output from these models is passed through
an ensembler, that combines these individual suggestions into
a complete encoding. The interface shows both complete as
well as individual encoding suggestions. Lastly, the usability
models aim to improve the ease of using the interface, by
generating clustering based full encoding recommendations.
4.3.1 Geo-location Recommender
This is a protest identification model, and aims to identify
protest location(s) from a news article. The model begins by
geo resolving the ‘location’ named entities found in the article.
Geo resolving involves mapping location named entities to
cities. These location entities can correspond to local points
of interest, alternative name for the city or the official city
name. The points of interest pose an interesting challenge as
they tend to match to multiple cities. For example: Main St.
can be found in almost every US city. The goal of the geo
resolution is to map these location entities to a minimum
number of cities with an assumption that generally a given
news article refers only to a focused set of locations. We use
the GeoNames database and perform the resolution in a top
down fashion. First, all the cities are identified and then
the ambiguous points of interest are mapped to these cities.
The cities are assigned a confidence rankings based on the
total number of mapping location entities. The case where
the same city name is present in multiple states is resolved
by selecting the most probable city based on past validated
encodings.
Frequently, articles report a statewide or a nationwide
protest. Such cases are identified by searching for various
language specific forms of the keywords: statewide and na-
tionwide, in the article. If such words are found, then we
add the top level state/country to our list of location recom-
mendations.
4.3.2 Temporal Recommender
This is another protest identification model that focuses
on identifying protest date(s) from a news article. Article
publish date is used as a date of reference and is identified
either by the publish date meta tag or is inferred by the
article fetch date. The reference date allows the model to
resolve words like yesterday or last Tuesday. The temporal
keywords are searched in following places: title, description
meta tag and the sentences containing protest keyword(s).
Temporal keywords found in title and description carry more
confidence score as compared to the ones found in article
text. Also, multiple mentions of the same date, increases the
confidence score of that date.
4.3.3 Entity Based Recommender
This is a protest characterization model that identifies the
salient features of a protest such as the reason of protest,
participating population group and whether the protest was
violent or peaceful. In context of AutoGSR, the reason
behind a protest can be defined in only following six cate-
gories: 1) Employment and Wages, 2) Housing, 3) Energy
and Resources, 4) Other Economic Policies, 5) Other Gov-
ernment Policies, and 6) Other. Similarly, the participating
population group can be one among the following eleven:
Agriculture, Business, Education, Ethnic, General Popula-
tion, Labor, Legal, Media, Medical, Refugees and Religious.
This model employs a multi class Na¨ıve Bayes model trained
on named entities extracted from article text and past protest
encodings to generate probability score for event Type and
population group categories.
4.3.4 MetaTags Based Recommender
This model is also a protest characterization model, and
works very similar to the Entity Based Recommender in
principle, but uses only the entities extracted from the title
and the description MetaTag to determine event type and
population group categories.
4.3.5 Ensembler
The task of the ensembler is to take the probabilistic
outputs of the protest identification and characterizations
models and generate complete encoding recommendations.
The complete encoding recommendation includes location,
date, event type and population group and is generated by
finding the most probable individual recommendations in

each category. Top two such complete recommendations are
shown in the interface.
4.3.6 Clustering Based Recommender
This is a usability model that aims to increase the ease of
using the system in two ways: first, by clustering similar news
articles together and second by generating full tuple encod-
ing recommendations based on validated encodings of other
articles in the article cluster. Article clusters are generated in
real time using the Lingo3G document clustering engine[12].
These clusters allow an analyst to work on related news
articles together. Encoding recommendations are generated
for un-validated documents using the validated encodings of
other documents in the cluster. These recommendations are
generated assuming that articles in the clusters are related
and hence will have similar encoding. Please note that these
recommendations are in addition to the recommendations
generated by the ensembler.
5. EVALUATION RESULTS
As stated earlier, the goal of the EMBERS AutoGSR is
to develop reliable ground truth civil unrest events while
minimizing the manual effort required to do so. With this
goal in mind, we evaluate our system alongside four aspects:
1. What is the reduction in number of articles realized
after each step of the models ecosystem? (Section 5.1)
2. How does AutoGSR compare w.r.t. manually generated
ground truth data? (Section 5.2)
3. What is the reduction in manual effort afforded by
AutoGSR? (Section 5.3)
4. How does AutoGSR compare with state-of-the-art sys-
tems like ICEWS and GDELT? (Section 5.4)
The evaluation is performed using a manually generated
list of civil unrest events for the same period, countries and
news sources. These events are hand coded by a team of 15-18
political scientists and subject matter experts working for an
independent agency (MITRE). As stated earlier, we refer to
the MITRE organized set of ground truth events as GSR, to
distinguish them from our AutoGSR system. We focus here
primarily on 3 recent months: October, November and De-
cember, 2015 for the 10 Latin American countries mentioned
earlier, viz. Argentina, Brazil, Chile, Colombia, Ecuador,
El Salvador, Mexico, Paraguay, Uruguay, and Venezuela.
The news articles are primarily reported in two languages:
Spanish and Portuguese, with some English sources as well.
5.1 Reduction in number of articles
Here we present the overall reduction in number of articles
for both protest and non protest articles after each stage of
the AutoGSR pipeline for both Spanish (ES) and Portuguese
(PT), fig. 5. For Spanish, the reduction in non protest articles
after filtering models is 60% and after ranking models is
64%, thereby leading to a net reduction of 86%. Similarly,
for Portuguese, the reduction in non protest articles after
filtering models is 34% and after ranking models is 60%,
thereby leading to a net reduction of 74%.
5.2 Quality and Coverage Evaluation
In order to determine quality, we need to identify how
similar GSR and AutoGSR encoding extractions are for a
given civil unrest event. As mentioned before, an event
encoding consists of the following four entities – protest
location (city level), protest date, reason of protest along with
Figure 5: Reduction in number of protest and non-protest
articles after each step.
violence identification (event type) and protesting population
group.
We define Quality Score (QS) on a 4 point scale where
each point correspond to how similar each of the encoding
entity is as explained below:
QualityScore(QS) = DS + LS + ES + PS
where DS, LS, ES and PS denote the date score, location
score, event type score and population score respectively.
Each of these scores are defined next:
DS = 1 − min(|GSReventdate − AutoGSReventdate|, 7)/7
In other words, if the date of the event identified by the GSR
is same as AutoGSR then DS is 1. Otherwise if they are
further apart by 7 days, then DS is 0.
Location score (LS) is defined as:
LS = 0.33 + 0.66(1 − min(dist, 300)/300)
where dist denotes the distance (in km) between the city
identified by GSR and AutoGSR city. Both GSR and Auto-
GSR use standardized city names from a World Gazetteer
that also provides latitude and longitude values, required for
the distance computation. Cities outside 300 km get a score
of 0.33; same cities get 1 and cities within 300km get scores
in the range [0.33, 1].
Event type score (ES) is defined in terms of triples (e1, e2, e3)
where e1 is the event granularity, e2 is the protest reason,
and e3 is the violence status. Each of these scores have a
value of 0 if they don’t match and 1 if they match.
ES =
1
3
e1 +
1
3
e2 +
1
3
e3
Population score (PS) is simply a binary (0/1) score de-
noting whether we identified the correct population group or
not. Finally, note that QS is designed to take values in the
range [0, 4].
Note that the above mentioned Quality Score is generated
only between a pair of event encodings for gsr and auto-
gsr. For a given month, there can be thousands of such
pairs and hence we need to find the most optimal mapping
pairs. We resolve this issue by first constructing a bipar-
tite graph between GSR and AutoGSR events where edge
weights correspond to Quality Score between each pair. Then
we construct a maximum weighted bipartite matching and
consider this the most optimal mapping between GSR and
AutoGSR events.
Table 6 shows the Average Quality Score, Precision and
Recall for each of the ten countries for three months: Oct

Table 6: Quality and Coverage Evaluation for AutoGSR vs Manual GSR*.
Country Quality Score Precision Recall
Oct Nov Dec Oct Nov Dec Oct Nov Dec
Argentina - 3.24 3.21 - 0.85 0.94 - 0.53 0.49
Brazil - 3.53 3.58 - 0.27 0.21 - 1.00 0.99
Chile 3.24 3.19 3.35 0.53 0.90 0.87 0.84 0.81 1.00
Colombia 3.17 3.24 2.71 0.95 0.97 0.94 0.81 0.63 0.92
Ecuador 3.09 3.20 3.20 0.60 1.00 0.86 1.00 0.84 1.00
El Salvador 3.33 3.09 3.10 1.00 0.89 0.80 0.91 1.00 1.00
Mexico 3.13 3.36 3.16 0.92 0.73 0.93 0.95 0.83 0.86
Paraguay 3.60 3.42 3.39 0.74 0.82 0.78 1.00 1.00 1.00
Uruguay 3.17 3.07 3.12 0.53 0.77 1.00 1.00 0.85 0.73
Venezuela - 3.48 3.39 - 1.00 0.86 - 0.59 0.93
*In October no manual GSR was generated for Argentina, Brazil and
Venezuela. Hence, we are unable to evaluate AutoGSR’s performance
Table 7: Daily Time Series Correlation Com-
parison of ICEWS, GDELT and AutoGSR with
Manual GSR.
Country
ICEWS
vs. GSR
GDELT
vs. GSR
AutoGSR
vs. GSR
Argentina 0.13 0.26 0.49
Brazil 0.35 0.29 0.70
Chile 0.21 0.28 0.51
Colombia 0.33 0.23 0.48
Ecuador 0.50 0.19 0.73
El Salvador nan 0.06 0.57
Mexico -0.09 0.35 0.26
Paraguay -0.12 0.10 0.86
Uruguay -0.11 0.00 0.41
Venezuela 0.08 -0.08 0.68
2015, Nov 2015, and Dec 2015. Here recall corresponds
to the number of events identified by the GSR that were
also identified by AutoGSR; and precision corresponds to
the number of events identified by AutoGSR that were also
found in the GSR. As the results in Table 6 demonstrate,
the AutoGSR system has a near perfect precision and recall,
thereby indicating that the system identified almost all civil
unrest events reported by the GSR while maintaining a really
low false positive rate. Further, for these identified events,
the quality of the extracted encodings is very close to the
true GSR encodings.
Brazil is the only country which appears to have a high
false positive rate which can be attributed to an incorrectly
added news source. Globo is a big media conglomerate in
Brazil that owns newspapers, news channels, and online
real time news portals. While MITRE focuses only on the
electronic version of the print newspaper (oglobo.globo.com),
the AutoGSR system was fetching articles from *.globo.com.
These also included articles from the real time news portal
g1.globo.com which reports events in real time without much
verification. Hence, we identified several more protests be-
cause the evolving news stories were considered as different
events. This is also corroborated by the fact that many of
the g1.globo.com links are dead or have a short life span as
links pointing to intermittent stories are removed once all
the details about an event are known. Interestingly, Brazil
has the highest Quality Score with a perfect recall pointing
to the fact that all the events reported by GSR were still
correctly identified by AutoGSR, and with very high quality.
These extraneous sub domains were removed from the system
in mid January 2016.
5.3 Reduction in Manual Effort
We evaluate both objectively and subjectively the reduc-
tion in manual effort. The objective evaluation is performed
by comparing total resource hours required to generate GSR
and AutoGSR, while subjective evaluation is done by sur-
veying MITRE’s analysts.
Tables 8 and 9 list the number of hours used by analysts
to generate GSR and AutoGSR events. Table 10 shows
minimum and maximum number of resource hours required
per week by GSR and AutoGSR teams. The table also lists
reduction figures. AutoGSR achieves an average reduction
of 71-72% in manual effort.
In order to develop intuition behind this reduction in man-
ual efforts, we surveyed the analysts from the MITRE team.
Figure 6: Time Distribution of MITRE’s Analysts across
various tasks of the GSR generation pipeline.
We requested them for an estimate of time that they spend in
performing various sub tasks of the GSR generation process.
Fig. 6 shows a rough sketch of the time distribution as re-
ported by them. The figure also shows individual reductions
provided for each of these sub tasks by AutoGSR. For each
pie, the missing color corresponds to the provided savings:
1) Gross Discovery: 100% reduction as articles are fetched
and loaded automatically into the system, 2) Skimming and
Targeting: 100% reduction as articles with cumulative prob-
ability score greater than equal to 0.5 are automatically
classified as protest articles, 3) Writing into GSR: almost
100% reduction as recording an extracted encoding is just a
matter of few clicks, and 4) Reading and Encoding Extrac-
tion: based on the left over reduction percentage (∼15%),
it appears that the automatically generated encoding rec-
ommendations are providing almost 33% reduction for this
task.
5.4 Comparison with ICEWS and GDELT
Finally, we evaluate AutoGSR’s performance against the
current state of the art systems: ICEWS and GDELT. For
each of these systems, for the given countries and time period,
we compare the daily events counts. Table 7 shows the
Pearson correlation values between daily time series for each
of the ten countries for ICEWS-GSR, GDELT-GSR and
AutoGSR-GSR combinations. While AutoGSR shows high
correlation with the GSR time series, the same is not true
for ICEWS and GDELT thereby presenting a strong case in
the favor of a system such as AutoGSR to generate reliable
ground truth data. Figure 7a compares the daily counts

Table 8: Resource Distribution for GSR
Resource
Type
#Resources
#Hours
per Week
Analysts 10-15 20
Q. C.
Analyst
2 24
Manager 1 10
Table 9: Resource Distribution for AutoGSR
Resource
Type
#Resources
#Hours
per Week
Analysts 4 15-20
Q. C.
Analyst
1 15-20
Table 10: Total Resource
Hours required per Week
Min. Max.
GSR 258 hrs 358 hrs
AutoGSR 75 hrs 100 hrs
Reduction 71% 72%
(a) Argentina Daily Counts
Comparison
(b) Weekly Time Series
Comparison for Paraguay
Figure 7: Daily and Weekly Time Series Comparison
between GSR and AutoGSR and Figure 7b compares the
weekly time series for all the four systems.
6. DISCUSSION
We have presented AutoGSR, an automated event coding
system for civil unrest events that is now in full continuous
production use. In addition to developing and deploying the
system, we have undertaken an entire life cycle analysis of
how such a system would fit in an analyst’s pipeline, with a
view to quantify benefits over a purely manual approach. The
results from our deployment indicate that the performance
measures obtained by AutoGSR are compelling to support
their continued use in an event modeling setting. Future
work is now focused on expanding the scope of AutoGSR
to new regions of the world, and to new classes of events,
beyond civil unrest.
Acknowledgments
Supported by the Intelligence Advanced Research Projects Ac-
tivity (IARPA) via DoI/NBC contract number D12PC000337,
the US Government is authorized to reproduce and distribute
reprints of this work for Governmental purposes notwithstanding
any copyright annotation thereon. Disclaimer: The views and
conclusions contained herein are those of the authors and should
not be interpreted as necessarily representing the official policies or
endorsements, either expressed or implied, of IARPA, DoI/NBC,
or the US Government.
7. REFERENCES
[1] E. Boschee, P. Natarajan, and R. Weischedel.
Automatic extraction of events from open source text
for predictive forecasting. In Handbook of
Computational Approaches to Counterterrorism, pages
51–67. Springer, 2013.
[2] C. Cortes and V. Vapnik. Support-vector networks.
Mach. Learn., 20(3):273–297, Sept. 1995.
[3] F. Hogenboom, F. Frasincar, U. Kaymak, and
F. De Jong. An overview of event extraction from text.
In DeRiVE Workshop at ISWC 2011, volume 779,
pages 48–57. Citeseer, 2011.
[4] T. Joachims. Text categorization with suport vector
machines: Learning with many relevant features. In
ECML ’98, ECML ’98, pages 137–142, London, UK,
UK, 1998. Springer-Verlag.
[5] Q. V. Le and T. Mikolov. Distributed representations
of sentences and documents. arXiv preprint
arXiv:1405.4053, 2014.
[6] K. Leetaru and P. A. Schrodt. Gdelt: Global data on
events, location, and tone, 1979–2012. In ISA Annual
Convention, volume 2. Citeseer, 2013.
[7] T. Mikolov, K. Chen, G. Corrado, and J. Dean.
Efficient estimation of word representations in vector
space. arXiv preprint arXiv:1301.3781, 2013.
[8] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and
J. Dean. Distributed representations of words and
phrases and their compositionality. In Advances in
neural information processing systems, pages
3111–3119, 2013.
[9] K. Nigam, A. McCallum, S. Thrun, and T. Mitchell.
Learning to classify text from labeled and unlabeled
documents. In AAAI ’98, AAAI ’98/IAAI ’98, pages
792–799, Menlo Park, CA, USA, 1998. American
Association for Artificial Intelligence.
[10] K. Nigam, A. K. McCallum, S. Thrun, and T. Mitchell.
Text classification from labeled and unlabeled
documents using em. Mach. Learn., 39(2-3):103–134,
May 2000.
[11] S. P. O’brien. Crisis early warning and decision
support: Contemporary approaches and thoughts on
future research. International Studies Review,
12(1):87–104, 2010.
[12] S. Osinski and D. Weiss. A concept-driven algorithm
for clustering search results. IEEE Intelligent Systems,
20(3):48–54, May 2005.
[13] N. Ramakrishnan and P. Butler et. al. ’beating the
news’ with embers: Forecasting civil unrest using open
source indicators. In KDD ’14, KDD ’14, pages
1799–1808, New York, NY, USA, 2014. ACM.
[14] L. Ramshaw, E. Boschee, M. Freedman, J. MacBride,
R. Weischedel, and A. Zamanian. Serif language
processing effective trainable language understanding.
Handbook of Natural Language Processing and Machine
Translation: DARPA Global Autonomous Language
Exploitation, pages 636–644, 2011.
[15] P. A. Schrodt. Tabari: Textual analysis by augmented
replacement instructions. Dept. of Political Science,
University of Kansas, Blake Hall, Version 0.7. 3B3,
pages 1–137, 2009.
[16] P. A. Schrodt. Cameo: Conflict and mediation event
observations event and actor codebook. Pennsylvania
State University, 2012.

EMBERS AutoGSR: Automated Coding of Civil Unrest Events

More Related Content

What's hot (18)

Viewers also liked (18)

Similar to EMBERS AutoGSR: Automated Coding of Civil Unrest Events (15)

More from Parang Saraf (15)

Recently uploaded (20)

EMBERS AutoGSR: Automated Coding of Civil Unrest Events