Tag Archives: NLP

Enhanced Messaging for the Emergency Response Sector (EMERSE)

My colleague Andrea Tapia and her team at PennState University have developed an interesting iPhone application designed to support humanitarian response. This application is part of their EMERSE project: Enhanced Messaging for the Emergency Response Sector. The other components of EMERSE include a Twitter crawler, automatic classification and machine learning.

The rationale for this important, applied research? “Social media used around crises involves self-organizing behavior that can produce accurate results, often in advance of official communications. This allows affected population to send tweets or text messages, and hence, make them heard. The ability to classify tweets and text messages automatically, together with the ability to deliver the relevant information to the appropriate personnel are essential for enabling the personnel to timely and efficiently work to address the most urgent needs, and to understand the emergency situation better” (Caragea et al., 2011).

The iPhone application developed by PennState is designed to help humanitarian professionals collect information during a crisis. “In case of no service or Internet access, the application rolls over to local storage until access is available. However, the GPS still works via satellite and is able to geo-locate data being recorded.” The Twitter crawler component captures tweets referring to specific keywords “within a seven-day period as well as tweets that have been posted by specific users. Each API call returns at most 1000 tweets and auxiliary metadata [...].” The machine translation component uses Google Language API.

The more challenging aspect of EMERSE, however, is the automatic classification component. So the team made use of the Ushahidi Haiti data, which includes some 3,500 reports about half of which came from text messages. Each of these reports were tagged according to a specific (but not mutually exclusive category), e.g., Medical Emergency, Collapsed Structure, Shelter Needed, etc. The team at PennState experimented with various techniques from (NLP) and Machine Learning (ML) to automatically classify the Ushahidi Haiti data according to these pre-existing categories. The results demonstrate that “Feature Extraction” significantly outperforms other methods while Support Vector Machine (SVM) classifiers vary significantly depending on the category being coded. I wonder whether their approach is more or less effective than this one developed by the University of Colorado at Boulder.

In any event, PennState’s applied research was presented at the ISCRAM 2011 conference and the findings are written up in this paper (PDF): “Classifying Text Messages for the Haiti Earthquake.” The co-authors: Cornelia Caragea, Nathan McNeese, Anuj Jaiswal, Greg Traylor, Hyun-Woo Kim, Prasenjit Mitra, Dinghao Wu, Andrea H. Tapia, Lee Giles, Bernard J. Jansen, John Yen.

In conclusion, the team at PennState argue that the EMERSE system offers four important benefits not provided by Ushahidi.

“First, EMERSE will automatically classify tweets and text messages into topic, whereas Ushahidi collects reports with broad category information provided by the reporter. Second, EMERSE will also automatically geo-locate tweets and text messages, whereas Ushahidi relies on the reporter to provide the geo-location information. Third, in EMERSE, tweets and text messages are aggregated by topic and region to better understand how the needs of Haiti differ by regions and how they change over time. The automatic aggregation also helps to verify reports. A large number of similar reports by different people are more likely to be true. Finally, EMERSE will provide tweet broadcast and GeoRSS subscription by topics or region, whereas Ushahidi only allows reports to be downloaded.”

In terms of future research, the team may explore other types of abstraction based on semantically related words, and may also “design an emergency response ontology [...].” So I recently got in touch with Andrea to get an update on this since their ISCRAM paper was published 14 months ago. I’ll be sure to share any update if this information can be made public.

Crisis Tweets: Natural Language Processing to the Rescue?

My colleagues at the University of Colorado, Boulder, have been doing some very interesting applied research on automatically extracting “situational awareness” from tweets generated during crises. As is increasingly recognized by many in the humanitarian space, Twitter can at times be an important source of relevant information. The challenge is to make sense of a potentially massive number of crisis tweets in near real-time to turn this information into situational awareness.

Using Natural Language Processing (NLP) and Machine Learning (ML), Colorado colleagues have developed a “suite of classifiers to differentiate tweets across several dimensions: subjectivity, personal or impersonal style, and linguistic register (formal or informal style).” They suggest that tweets contributing to situational awareness are likely to be “written in a style that is objective, impersonal, and formal; therefore, the identification of subjectivity, personal style and formal register could provide useful features for extracting tweets that contain tactical information.” To explore this hypothesis, they studied the follow four crisis events: the North American Red River floods of 2009 and 2010, the 2009 Oklahoma grassfires, and the 2010 Haiti earthquake.

The findings of this study were presented at the Association for the Advancement of Artificial Intelligence. The team from Colorado demonstrated that their system, which automatically classifies Tweets that contribute to situational awareness, works particularly well when analyzing “low-level linguistic features,” i.e., word-frequencies and key-word search. Their analysis also showed that “linguistically-motivated features including subjectivity, personal/impersonal style, and register substantially improve system performance.” In sum, “these results suggest that identifying key features of user behavior can aid in predicting whether an individual tweet will contain tactical information. In demonstrating a link between situational awareness and other markable characteristics of Twitter communication, we not only enrich our classification model, we also enhance our perspective of the space of information disseminated during mass emergency.”

The paper, entitled: “Natural Language Processing to the Rescue? Extracting ‘Situational Awareness’ Tweets During Mass Emergency,” details the findings above and is available here. The study was authored by Sudha Verma, Sarah Vieweg, William J. Corvey, Leysia Palen, James H. Martin, Martha Palmer, Aaron Schram and Kenneth M. Anderson.

Situational Awareness in Mass Emergency: Behavioral & Linguistic Analysis of Disaster Tweets

Sarah Vieweg‘s doctoral dissertation from the University of Colorado is a must-read for anyone interested in the use of twitter during crises. I read the entire 300-page study because it provides important insights on how automated natural language processing (NLP) can be applied to the Twittersphere to provide situational awareness following a sudden-onset emergency. Big thanks to Sarah for sharing her dissertation with QCRI. I include some excerpts below to highlight the most important findings from her excellent research.


“In their research on human behavior in disaster, Fritz and Marks (1954) state: ‘[T]he immediate problem in a disaster situation is neither un-controlled behavior nor intense emotional reaction, but deficiencies of coordination and organization, complicated by people acting upon individual…definitions of the situation.’”

“Fritz and Marks’ assertion that people define disasters individually, which can lead to problematic outcomes, speaks to the need for common situational awareness among affected populations. Complete information is not attained during mass emergency, else it would not be a mass emergency. However, the more information people have and the better their situational awareness, and the better equipped they are to make tactical, strategic decisions.”

“[D]uring crises, people seek information from multiple sources in an attempt to make locally optimal decisions within given time constraints. The first objective, then, is to identify what tweets that contribute to situational awareness ‘look like’—i.e. what specific information do they contain? This leads to the next objective, which is to identify how information is communicated at a linguistic level. This process provides the foundation for tools that can automatically extract pertinent, valuable information—training machines to correctly ‘understand’ human language involves the identification of the words people use to communicate via Twitter when faced with a disaster situation.”

Research Design & Results

Just how much situational awareness can be extracted from twitter during a crisis? What constitutes situational awareness in the first place vis-a-vis emergency response? And can the answer to these questions yield a dedicated ontology that can be fed into automated natural language processing platforms to generate real-time, shared awareness? To answer these questions, Sarah analyzed four emergency events: Oklahoma Fires (2009), Red River Floods (2009 & 2010) and the Haiti Earthquake (2010).

She collected tweets generated during each of these emergencies and developed a three-step qualitative coding process to analyze what kinds of information on Twitter contribute to situational awareness during a major emergency. As a first step, each tweet was categorized as either:

O: Off-topic
“Tweets do not contain any information that mentions or relates to the emergency event.”

R: On-topic and Relevant to Situational Awareness
“Tweets contain information that provides tactical, actionable information that can aid people in making decisions, advise others on how to obtain specific information from various sources, or offer immediate post- impact help to those affected by the mass emergency.”

N: On-topic and Not Relevant to Situational Awareness
“Tweets are on-topic because they mention the emergency by including offers of prayer and support in relation to the emergency, solicitations for donations to charities, or casual reference to the emergency event. But these tweets do not meet the above criteria for situational relevance.”

The O, R, and N coding of the crisis datasets resulted in the following statistics for each of the four datasets:

For the second coding step, on-topic relevant tweets were annotated with more specific information based on the following coding rule:

S: Social Environment
“These tweets include information about how people and/or animals are affected by a hazard, questions asked in relation to the hazard, responses to the hazard and actions to take that directly relate to the hazard and the emergency situation it causes. These tweets all include description of a human element in that they explain or display human behavior.”

B: Built Environment
“Tweets that include information about the effect of the hazard on the built environment, including updates on the state of infrastructure, such as road closures or bridge outages, damage to property, lack of damage to property and the overall state or condition of structures.”

P: Physical Environment
“Tweets that contain specific information about the hazard including particular locations of the hazard agent or where the hazard agent is expected or predicted to travel or predicted states of the hazard agent going forward, notes about past hazards that compare to the current hazard, and how weather may affect hazard conditions. These tweets additionally include information about the type of hazard in general [...]. This category also subsumes any general information about the area under threat or in the midst of an emergency [...].”

The result of this coding for Haiti is depicted in the figures below.

According to the results, the social environment (‘S’) category is most common in each of the datasets. “Disasters are social events; in each disaster studied in this dissertation, the disaster occurred because a natural hazard impacted a large number of people.”

For the third coding step, Sarah created a comprehensive list of several dozen  “Information Types” for each “Environment” using inductive, data-driven analysis of twitter communications, which she combined with findings from the disaster literature and official government procedures for disaster response. In total, Sarah identified 32 specific types of information that contribute to situational awareness. The table below compares the Twitter Information Types for all three environments as related to government procedures, for example.

“Based on the discourse analysis of Twitter communications broadcast during four mass emergency events,” Sarah identified 32 specific types of information that “contribute to situational awareness. Subsequent analysis of the sociology of disaster literature, government documents and additional research on the use of Twitter in mass emergency uncovered three additional types of information.”

In sum, “[t]he comparison of the information types [she] uncovered in [her] analysis of Twitter communications to sociological research on disaster situations, and to governmental procedures, serves as a way to gauge the validity of [her] ground-up, inductive analysis.” Indeed, this enabled Sarah to identify areas of overlap as well as gaps that needed to be filled. The final Information Type framework is listed below:

And here are the results of this coding framework when applied to the Haiti data:

“Across all four datasets, the top three types of information Twitter users communicated comprise between 36.7-52.8% of the entire dataset. This is an indication that though Twitter users communicate about a variety of informa-tion, a large portion of their attention is focused on only a few types of in-formation, which differ across each emergency event. The maximum number of information types communicated during an event is twenty-nine, which was during the Haiti earthquake.”

Natural Language Processing & Findings

The coding described above was all done manually by Sarah and research colleagues. But could the ontology she has developed (Information Types) be used to automatically identify tweets that are both on-topic and relevant for situational awareness? To find out, she carried out a study using VerbNet.

“The goal of identifying verbs used in tweets that convey information relevant to situational awareness is to provide a resource that demonstrates which VerbNet classes indicate information relevant to situational awareness. The VerbNet class information can serve as a linguistic feature that provides a classifier with information to identify tweets that contain situational awareness information. VerbNet classes are useful because the classes provide a list of verbs that may not be present in any of the Twitter data I examined, but which may be used to describe similar information in unseen data. In other words, if a particular VerbNet class is relevant to situational awareness, and a classifier identifies a verb in that class that is used in a previously unseen tweet, then that tweet is more likely to be identified as containing situational awareness information.”

Sarah identified 195 verbs that mapped to her Information Types described earlier. The results of using this verb-based ontology are mixed, however. “A majority of tweets do not contain one of the verbs in the identified VerbNet classes, which indicates that additional features are necessary to classify tweets according to the social, built or physical environment.”

However, when applying the 195 verbs to identify on-topic tweets relevant to situational awareness to previously unused Haiti data, Sarah found that using her customized VerbNet ontology resulted in finding 9% more tweets than when using her “Information Types” ontology. In sum, the results show that “using VerbNet classes as a feature is encouraging, but other features are needed to identify tweets that contain situational awareness information, as not all tweets that contain situational awareness information use one of the verb members in the […] identified VerbNet classes. In addition, more research in this area will involve using the semantic and syntactic information contained in each VerbNet class to identify event participants, which can lead to more fine-grained categorization of tweets.”


“Many tweets that communicate situational awareness information do not contain one of the verbs in the identified VerbNet classes, [but] the information provided with named entities and semantic roles can serve as features that classifiers can use to identify situational awareness information in the absence of such a verb. In addition, for tweets correctly identified as containing information relevant to situational awareness, named entities and semantic roles can provide classifiers with additional information to classify these tweets into the social, built and physical environment categories, and into specific information type categories.”

“Finding the best approach toward the automatic identification of situational awareness information communicated in tweets is a task that will involve further training and testing of classifiers.”

Twitcident: Filtering Tweets in Real-Time for Crisis Response

The most recent newcomer to the “tweetsourcing” space comes to us from Delft University of Technology in the Netherlands. Twitcident is a web-based filtering system that extracts crisis information from Twitter in real-time to support emergency response efforts. Dutch emergency services have been testing the platform over the past 10 months and results “show the system to be far more useful than simple keyword searching of a twitter feed” (NewScientist).

Here’s how it works. First the dashboard, which shows current events-of-interest being monitored.

Lets click on “Texas”, which produces the following page. More than 22,000 tweets potentially relate to the actual fire of interest.

This is where the filtering comes in:

The number of relevant tweets is reduced with every applied filter.

Naturally, geo-location is also an optional filter.

Twitcident also allows for various visualization options, including timelines, word clouds and charts.

The system also allows the user to view the filtered tweets on a map. The pictures and videos shared via twitter are also aggregated and viewable on dedicated tabs.

The developers of the platform have not revealed how their algorithms work but will demo the tool at the World Wide Web 2012 conference in France next week. In the meantime, here’s a graphic that summarizes the platform workflow.

I look forward to following Twitcident’s developments. I’d be particularly interested in learning more about how Dutch emergency services have been using the tool and what features they think would improve the platform’s added value.

Some Thoughts on Real-Time Awareness for Tech@State

I’ve been invited to present at Tech@State in Washington DC to share some thoughts on the future of real-time awareness. So I thought I’d use my blog to brainstorm and invite feedback from iRevolution readers. The organizers of the event have shared the following questions with me as a way to guide the conver-sation: Where is all of this headed?  What will social media look like in five to ten years and what will we do with all of the data? Knowing that the data stream can only increase in size, what can we do now to prepare and prevent being over-whelmed by the sheer volume of data?

These are big, open-ended questions, and I will only have 5 minutes to share some preliminary thoughts. I shall thus focus on how time-critical crowdsourcing can yield real-time awareness and expand from there.

Two years ago, my good friend and colleague Riley Crane won DARPA’s $40,000 Red Balloon Competition. His team at MIT found the location of 10 weather balloons hidden across the continental US in under 9 hours. The US covers more than 3.7 million square miles and the balloons were barely 8 feet wide. This was truly a needle-in-the-haystack kind of challenge. So how did they do it? They used crowdsourcing and leveraged social media—Twitter in particular—by using a “recursive incentive mechanism” to recruit thousands of volunteers to the cause. This mechanism would basically reward individual participants financially based on how important their contributions were to the location of one or more balloons. The result? Real-time, networked awareness.

Around the same time that Riley and his team celebrated their victory at MIT, another novel crowdsourcing initiative was taking place just a few miles away at The Fletcher School. Hundreds of students were busy combing through social and mainstream media channels for actionable and mappable information on Haiti following the devastating earthquake that had struck Port-au-Prince. This content was then mapped on the Ushahidi-Haiti Crisis Map, providing real-time situational awareness to first responders like the US Coast Guard and US Marine Corps. At the same time, hundreds of volunteers from the Haitian Diaspora were busy translating and geo-coding tens of thousands of text messages from disaster-affected communities in Haiti who were texting in their location & most urgent needs to a dedicated SMS short code. Fletcher School students filtered and mapped the most urgent and actionable of these text messages as well.

One year after Haiti, the United Nation’s Office for the Coordination of Humanitarian Affairs (OCHA) asked the Standby Volunteer Task Force (SBTF) , a global network of 700+ volunteers, for a real-time map of crowdsourced social media information on Libya in order to improve their own situational awareness. Thus was born the Libya Crisis Map.

The result? The Head of OCHA’s Information Services Section at the time sent an email to SBTF volunteers to commend them for their novel efforts. In this email, he wrote:

“Your efforts at tackling a difficult problem have definitely reduced the information overload; sorting through the multitude of signals on the crisis is no easy task. The Task Force has given us an output that is manageable and digestible, which in turn contributes to better situational awareness and decision making.”

These three examples from the US, Haiti and Libya demonstrate what is already possible with time-critical crowdsourcing and social media. So where is all this headed? You may have noted from each of these examples that their success relied on the individual actions of hundreds and sometimes thousands of volunteers. This is primarily because automated solutions to filter and curate the data stream are not yet available (or rather accessible) to the wider public. Indeed, these solutions tend to be proprietary, expensive and/or classified. I thus expect to see free and open source solutions crop up in the near future; solutions that will radically democratize the tools needed to gain shared, real-time awareness.

But automated natural language processing (NLP) and machine learning alone are not likely to succeed, in my opinion. The data stream is actually not a stream, it is a massive torent of non-indexed information, a 24-hour global firehose of real-time, distributed multi-media data that continues to outpace our ability to produce actionable intelligence from this torrential downpour of 0′s and 1′s. To turn this data tsunami into real-time shared awareness will require that our filtering and curation platforms become more automated and collaborative. I believe the key is thus to combine automated solutions with real-time collabora-tive crowdsourcing tools—that is, platforms that enable crowds to collaboratively filter and curate real-time information, in real-time.

Right now, when we comb through Twitter, for example, we do so on our own, sitting behind our laptop, isolated from others who may be seeking to filter the exact same type of content. We need to develop free and open source platforms that allow for the distributed-but-networked, crowdsourced filtering and curation of information in order to democratize the sense-making of the firehose. Only then will the wider public be able to win the equivalent of Red Balloon competitions without needing $40,000 or a degree from MIT.

I’d love to get feedback from readers about what other compelling cases or arguments I should bring up in my presentation tomorrow. So feel free to post some suggestions in the comments section below. Thank you!

Detecting Emerging Conflicts with Web Mining and Crisis Mapping

My colleague Christopher Ahlberg, CEO of Recorded Future, recently got in touch to share some exciting news. We had discussed our shared interests a while back at Harvard University. It was clear then that his ideas and existing technologies were very closely aligned to those we were pursuing with Ushahidi’s Swift River platform. I’m thrilled that he has been able to accomplish a lot since we last spoke. His exciting update is captured in this excellent co-authored study entitled “Detecting Emergent Conflicts Through Web Mining and Visualization” which is available here as a PDF.

The study combines almost all of my core interests: crisis mapping, conflict early warning, conflict analysis, digital activism, pattern recognition, natural language processing, machine learning, data visualization, etc. The study describes a semi-automatic system which automatically collects information from pre-specified sources and then applies linguistic analysis to user-specified extract events and entities, i.e., structured data for quantitative analysis.

Natural Language Processing (NLP) and event-data extraction applied to crisis monitoring and analysis is of course nothing new. Back in 2004-2005, I worked for a company that was at the cutting edge of this field vis-a-vis conflict early warning. (The company subsequently joined the Integrated Conflict Early Warning System (ICEWS) consortium supported by DARPA). Just a year later, Larry Brilliant told TED 2006 how the Global Public Health Information Net-work (GPHIN) had leveraged NLP and machine learning to detect an outbreak of SARS 3 months before the WHO. I blogged about this, Global Incident Map, European Media Monitor (EMM), HavariaHealthMap and Crimson Hexagon back in 2008. Most recently, my colleague Kalev Leetaru showed how applying NLP to historical data could have predicted the Arab Spring. Each of these initiatives represents an important effort in leveraging NLP and machine learning for early detection of events of interest.

The RecordedFuture system works as follows. A user first selects a set of data sources (websites, RSS feeds, etc) and determines the rate at which to update the data. Next, the user chooses one or several existing “extractors” to find specific entities and events (or constructs a new type). Finally, a taxonomy is selected to specify exactly how the data is to be grouped. The data is then automatically harvested and passed through a linguistics analyzer which extracts useful information such as event types, names, dates, and places. Finally, the reports are clustered and visualized on a crisis map, in this case using an Ushahidi platform. This allows for all kinds of other datasets to be imported, compared and analyzed, such as high resolution satellite imagery and crowdsourced data.

A key feature of the RecordedFuture system is that extracts and estimates the time for the event described rather than the publication time of the newspaper article parsed, for example. As such, the harvested data can include both historic and future events.

In sum, the RecordedFuture system is composed of the following five features as described in the study:

1. Harvesting: a process in which text documents are retrieved from various sources and stored in the database. The documents are stored for long-term if permitted by terms of use and IPR legislation, otherwise they are only stored temporarily for the needed analysis.

2. Linguistic analysis: the process in which the retrieved texts are analyzed in order to extract entities, events, time and location, etc. In contrast to other components, the linguistic analysis is language dependent.

3. Refinement: additional information can be obtained in this process by synonym detection, ontology analysis, and sentiment analysis.

4. Data analysis: application of statistical and AI-based models such as Hidden Markov Models (HMMs) and Artificial Neural Networks (ANNs) to generate predictions about the future and detect anomalies in the data.

5. User experience: a web interface for ordinary users to interact with, and an API for interfacing to other systems.

The authors ran a pilot that “manually” integrated the RecordedFuture system with the Ushahidi platform. The result is depicted in the figure below. In the future, the authors plan to automate the creation of reports on the Ushahidi platform via the RecordedFuture system. Intriguingly, the authors chose to focus on protest events to demo their Ushahidi-coupled system. Why is this intriguing? Because my dissertation analyzed whether access to new information and communication technologies (ICTs) are statistically significant predictors of protest events in repressive states. Moreover, the protest data I used in my econometric analysis came from an automated NLP algorithm that parsed Reuters Newswires.

Using RecordedFuture, the authors extracted some 6,000 protest event-data for Quarter 1 of 2011. These events were identified and harvested using a “trained protest extractor” constructed using the system’s event extractor frame-work. Note that many of the 6,000 events are duplicates because they are the same events but reported by different forces. Not surprisingly, Christopher and team plan to develop a duplicate detection algorithm that will also double as a triangulation & veracity scoring feature. I would be particularly interested to see them do this kind of triangulation and validation of crowdsourced data on the fly.

Below are the protest events picked up by RecordedFuture for both Tunisia and Egypt. From these two figures, it is possible to see how the Tunisian protests preceded those in Egypt.

The authors argue that if the platform had been set up earlier this year, a user would have seen the sudden rise in the number of protests in Egypt. However, the authors acknowledge that their data is a function of media interest and attention—the same issue I had with my dissertation. One way to overcome this challenge might be by complementing the harvested reports with crowdsourced data from social media and Crowdmap.

In the future, the authors plan to have the system auto-detect major changes in trends and to add support for the analysis of media in languages beyond English. They also plan to test the reliability and accuracy of their conflict early warning algorithm by comparing their forecasts of historical data with existing conflict data sets. I have several ideas of my own about next steps and look forward to speaking with Christopher’s team about ways to collaborate.