Twitter Speaks: A Case of National Disaster Situational Awareness

TL;DR

This paper introduces TwiSA, a social media analysis framework that leverages Twitter data with text mining techniques to enhance situational awareness during natural disasters, demonstrated through the 2015 South Carolina flood.

Contribution

It presents a novel analytical framework combining sentiment analysis and topic modeling for disaster situational awareness using Twitter data.

Findings

TwiSA effectively tracks public concerns during disasters.

Social media analysis provides timely insights compared to traditional surveys.

Framework improves disaster response and recovery efforts.

Abstract

In recent years, we have been faced with a series of natural disasters causing a tremendous amount of financial, environmental, and human losses. The unpredictable nature of natural disasters' behavior makes it hard to have a comprehensive situational awareness (SA) to support disaster management. Using opinion surveys is a traditional approach to analyze public concerns during natural disasters; however, this approach is limited, expensive, and time-consuming. Luckily the advent of social media has provided scholars with an alternative means of analyzing public concerns. Social media enable users (people) to freely communicate their opinions and disperse information regarding current events including natural disasters. This research emphasizes the value of social media analysis and proposes an analytical framework: Twitter Situational Awareness (TwiSA). This framework uses text mining methods including sentiment analysis and topic modeling to create a better SA for disaster preparedness, response, and recovery. TwiSA has also effectively deployed on a large number of tweets and tracks the negative concerns of people during the 2015 South Carolina flood.

Tables

Table 1. Table 1: Queries for Twitter Data Collection

#floodsc OR #scflood2015 OR #SCFloodRelief OR #southcarolinastrong OR #prayforsc OR #scflood OR #scflooding OR #FloodGSSCMMwithlove OR #floodingsc OR #flood OR flood

Table 2. Table 2: A Sample of Tweets

Tweet 1	How many South Carolinians have flood insurance? Few.They’ll be looking to federal gov’t or their tight-fisted governor for help. #SCFlood
Tweet 2	Damage inside Student Activities room at Westwood. #SCFlood

Table 3. Table 3: Flood Topics Examples

Victims	Damage & Costs	Drinking Water	Insurance	Road Damage	Roof Damage
scflood	damage	water	families	road	roof
victims	property	drink	insurance	damage	damage
flooding	loss	boil	homeowners	roadway	repair
disaster	roadway	bottle	destroyed	st	danger
death	construction	clean	support	exit	home
Bridge Damage	Flood Report	Homelessness	Power Lost	Animal
disaster	alert	flood	lost	Animal
bridge	warning	homeless	flood	Shelter
flooded	effect	lives	home	dog
natural	flood	woman	power	poor
flood	remain	boy	family	find

Table 4. Table 4: Total Frequency of Topics

Topic	Percentage	Topic	Percentage
Victims	18.75%	Roof Damage	9.37%
Damage and Costs	15.62%	Bridge Damage	6.25%
Drinking Water	9.37%	Flood Report	6.25%
Insurance	9.37%	Power Lost	4.68%
Homelessness	9.37%	Animal	1.6%
Road Damage	9.37%

Table 5. Table 5: Negative Topics from 10/03/2015 to 10/15/2015

	Day 1: 10/03/15	Day 2: 10/04/15	Day 3: 10/05/15	Day 4: 10/06/15	Day 5: 10/07/15	Day 6: 10/08/15	Day 7: 10/09/15	Day 8: 10/10/15	Day 9: 10/11/15	Day 10: 10/12/15	Day 11: 10/13/15	Day 12: 10/14/15	Day 13: 10/15/15
Animal	0	1	0	0	0	0	0	0	0	0	0	0	0
Bridge Damage	0	0	1	0	0	1	1	0	0	0	0	1	0
Damage and Costs	1	1	1	1	1	0	1	1	1	0	0	1	1
Drinking Water	0	0	1	1	1	1	1	1	0	0	0	0	0
Flood Report	1	1	0	1	1	0	0	0	0	0	1	0	0
Homelessness	1	0	1	1	1	1	1	0	0	0	0	0	0
Insurance	0	0	0	1	1	1	0	0	1	1	1	0	0
Power Loss	0	1	0	0	0	1	0	1	0	0	0	0	0
Road Damage	0	0	1	1	1	1	1	0	0	1	0	0	0
Roof Damage	0	0	1	1	1	1	1	1	0	0	0	0	0
Victims	0	1	1	1	1	1	1	1	1	1	1	1	1

Full text

1 Introduction

The growth of social media has provided an opportunity to track people’s concerns in a new way. Twitter has 316 million users techrunch , provides capability of real-time feedback, and utilizes time stamp to provide conversation updates to users. Therefore, Twitter’s potential as a reliable and relevant as a data source is evident and provides a unique opportunity to understand users’ concerns mejova2015twitter ; paul2011you . In contrast to surveys and surveillance networks that can take weeks or even years to collect data, Twitter is publicly accessible with no waiting time.

SA is “all knowledge that is accessible and can be integrated into a coherent picture, when required, to assess and cope with a situation” sarter1991situation . Obtaining reliable and accurate information and providing access to knowledge derived from the collected data is the main concern in situational awareness. Thus, SA and its inherent processes play an essential role in helping people facing a natural disaster. During a natural disaster, people immediately try to collect or share information to keep them away from possible dangerous situations castillo2016big . Developing high level of Situational Awareness (SA) by analyzing real-time big data helps to become more effective in disaster management kitchin2014real ; madey2007enhanced .

In October 2015 hurricane Joaquin led to an unusually high amount of rainfall in the state of South Carolina (SC). The precipitation ranged from 11 inches to 26 inches in different areas and caused server flooding in the state. This flood caused more than $12 billion damage to roads, homes, and infrastructures with 19 dead zurich ; scemd . Twitter users expressed their feelings and opinions during the 2015 SC flood through their Tweets. This valuable time-sensitive data can improve SA development and give all involved parties a more accurate picture of the situation during the flood.

Crowdsourced data such as social media data is a rapid and accessible source for SA development vieweg2010microblogging . In other words, understanding users’ concerns and their emotions can help disaster managers develop a better real-time SA and decision making plans neppalli2017sentiment ; callaghan2016disaster . During the height of the 2015 SC flood, thousands of people were twitting about the flood on an average of 3000 tweets per hour. One can imagine that manually analyzing this many tweets is not easily possible; therefore, new customized approaches are necessary to disclose hidden semantic features from social media. This study proposes an analytical framework, Twitter Situational Awareness (TwiSA), to analyze unstructured flood-related tweets to better understand temporal concerns of people affected by this natural disaster. The proposed framework employs both sentiment analysis and topic modeling to uncover temporal patterns of public concerns to provide a better SA. We apply TwiSA on a large number of tweets and tracks the negative concerns of people during the 2015 SC flood.

2 Related Work

Twitter data has been used for a broad range of applications such as business bollen2011twitter ; collins2018social ; karami2018us ; kordzadeh2019exploring ; kordzadeh2018investigating , politics tumasjan2010predicting ; karami2018mining ; alperin2018politicians ; kitzie2018life ; najafabadi2018hacktivism ; karami2019political , and health szomszor2010swineflu ; webb2018Characterizing ; karami2018characterizing ; karami2018characterizingtrans ; shaw2017computational ; shaw2019exploratory . In this section, we review the applications of Twitter data in natural disasters including fire, flood, earthquake, hurricane and typhoon, and volcano eruption.

Abel et al. (2012) developed a tool, called Twitcident, to collect fire-related tweets and search tweets using queries abel2012twitcident . Sinnappan et al. (2010) collected tweets and used qualitative approach to find meaningful categorization of fire-related tweets sinnappan2010priceless

Earthquake-related studies proposed frameworks to improve disaster management process. These included using data mining and natural language processing for damage detection and assessment of earthquakes avvenuti2014ears , proposing a probabilistic spatiotemporal model for reporting earthquake related events sakaki2013tweet , developing a detection algorithm based on the frequency of tweets to detect earthquake earle2012twitter , applying classifier methods on tweets to detect earthquake sakaki2010earthquake ; robinson2013sensitive , using qualitative approach to analyze people’s behavior after an earthquake miyabe2012use , applying a keyword level analysis to track social attitudes during and after an earthquake doan2011analysis , and analyzing dynamic of rumor mill in tweets oh2010exploration .

Literature also contains methods for disaster management during hurricanes and typhoon. This research area analyzed online public communications by police and fire departments in Twitter and Facebook hughes2014online , used classification methods to detect fake images during the hurricane gupta2013faking , and analyzed external factors such as geolocation and internal factors (e.g. geolocation, and stakeholders power) in a small sample of tweets to find patterns of information dissemination takahashi2015communicating . Recently, a study mapped online users’ sentiments to track their changes during a hurricane neppalli2017sentiment . Twitter data was also used to help disaster management using qualitative content analysis to categorize information inside the Twitter data during a volcano eruption sreenivasan2011tweet .

Flood-related studies developed numerous frameworks to improve disaster management during floods. These studies tracked users’ behavior during a flood palen2010twitter , combined georeferenced social media messages and geographic features of flood phenomena de2015geographic , used machine learning techniques to find tweets reporting damage ashktorab2014tweedr , mapped geo-tagged tweets to directly support geotechnical experts for reconnaissance purposes dashti2014supporting , analyzed crisis relevant information from Twitter using related keywords and geo-tagged tweets herfort2014does , investigated the correlation between location and the types of URLs inside tweets murthy2013twitter , explored the discursive aspects of Twitter communication with qualitative analysis shaw2013sharing , analyzed the related tweets with qualitative data coding to find useful information vieweg2010microblogging , and used keyword analysis to determine the types of information in the flood-related tweets kongthon2012role .

Although it is evident from the literature that past research on the use of social media during a natural disaster provides useful insight into disaster management, unpredictable nature of natural disasters provides a great motivation for disaster responders to constantly improve disaster management by exploring new perspectives. Thus, we believe we can improve disaster management through better integration of SA into the management practices specially when it comes to people concerns and negative feelings. To our knowledge no other research studied the concerns of people behind their negative feelings for having a better situational awareness. This study proposes an analytical framework to explore a large number of tweets using sentiment analysis and topic modeling for tracking temporal patterns of public negative concerns during the 2015 SC flood.

3 Methodology and Results

This paper proposes a framework, Twitter Situational Awareness (TwiSA), with four components: data collection, sentiment analysis, temporal topic discovery, and topic content analysis.

3.1 Data Collection

There are two methods to collect a large number of tweets in TwiSA. One method is to retrieve data using Twitter APIs (Application Programming Interfaces) twitapi , and the other method is to ask an independent service to provide the data. The first method, despite being free, is limited in that one can only retrieve a small portion of relevant tweets. Therefore, we asked an independent company, Gnip gnip , to extract 100% of tweets related to our research. Similar to Palen et al. (2010) palen2010twitter , we ordered the queries in Table 1 to collect tweets for the days having flood danger or flood side effect. One million tweets were filtered based on 13 days in October 2015. We chose this time frame because it represents a period that starts with the highest amount of rainfall (October 3, 2015) and ends with the last day for boil advisory (October 15, 2015) zurich . It is worth mentioning that we did not remove retweets and the tweets containing URLs from our dataset but the word “rt” and URLs were removed from the collected tweets. Table 2 shows a sample of tweets that are related to insurance and damage issues using the #scflood query.

3.2 Sentiment Analysis

Sentiment analysis discloses the overall feelings inside text data taboada2011lexicon . Learning-based and lexicon-based methods are two main approaches for sentiment analysis khan2015combining . The first approach uses machine learning classifiers when there is prior knowledge about data categories. In this case, a sample of the data is first labeled by human raters such as assigning spam and non-spam labels to a sample of emails taboada2011lexicon . The second approach, a cost-effective one, finds the frequency of a predefined dictionary of positive and negative terms to disclose sentiment in the data when there is not any prior knowledge about its categories medhat2014sentiment . We did not have any prior knowledge about the categories of the tweets in this research; therefore, we applied the second approach to find positive, negative, and neutral tweets.

Linguistic Inquiry and Word Count (LIWC) liwcco is a linguistics analysis tool that reveals thoughts, feelings, personality, and motivations in a corpus based on lexicon-based approach. This software assumes that the frequency of words can reveal overall tone in the corpus karami2014improving ; karami2014exploiting ; karami2015online . This tool applies a simple word-count process in text documents and maps each word in the documents on the already developed internal LIWC dictionaries in different categories such as negative emotion franklin2015some .

LIWC has good sensitivity value, specificity value, and English proficiency measure golder2011diurnal . Comparing to competitors such as deep learning, this tool can be used alone or be combined with other methods gonccalves2013comparing . Applying LIWC on the collected tweets shows that there are 217,074 negative tweets, 529,150 neutral tweets, and 217,183 positive tweets. When people are exposed to unpredictable events, they show unpleasant negative feelings that come from the presence of threats with undesirable outcomes nolen1991prospective ; yin2014anxious . We used the negative tweets for the next step to identify unpleasant elements of the SC flood through temporal topic discovery analysis (Figure 1). Figure 2 shows that the number of negative tweets per day has changed from $\sim$12,700 tweets to $\sim$30,000 tweets with average $\sim$16,700 tweets per day.

3.3 Temporal Topic Discovery

Next we turned our attention to detect topics among the pool of the 217,074 negative tweets to showcase the proposed framework. Different approaches for detecting the topics in a corpus have been developed based on the neural network and statistical distributions. During the last decade, the neural network has been considered for improving data analysis methods such as the deep neural network (DNN) zhang2016learning . The models based on neural network are costly and time-consuming methods mohammadi2018deep ; zhao2017research ; chilimbi2014project that require a lot of training data along with having a chance of overfitting problem erhan2010does . For example, one study needed to train millions of tweets for the pre-training step severyn2015twitter that is also a barrier for having an efficient real-time streaming data mahmud2018applications . In addition, the deep learning methods haven’t shown significant performance over other methods in small or medium size corpora mahmud2018applications ; zhang2016learning .

According to the literature, Latent Dirichlet Allocation (LDA) blei2003latent is a valid mcauliffe2008supervised ; hong2010empirical ; karami2014fftm ; karami2015fuzzyiconf and widely used lu2011investigating ; paul2011you model for discovering topics in a corpus karami2018computational . LDA assumes that there are multiple topics in a corpus and each topic is a distribution over corpus’s vocabulary and can be inferred from word-document co-occurrences blei2003latent ; karami2015fuzzy ; karami2018fuzzy .

LDA assigns each word in a document to each of topics with a different degree of membership using a statistical distribution karami2015flatm . For example, in a given corpus, LDA assigns “flood,” “earthquake,” and “hurricane” into a topic with “natural disaster” theme. TwiSA uses LDA to identify topics in the negative tweets. This research uses Mallet implementation of LDA mallet that uses Gibbs sampling that is a specific form of Markov Chain Monte Carlo (MCMC) steyvers2007probabilistic with removing standard stopwords list and having 1000 iterations to detect top 25 topics per day (325 topics in total) in the negative tweets per each of the 13 days.

3.4 Topic Content Analysis

The last component of TwiSA analyzes extracted topics. This component labels and categorizes 25 topics per day, analyzes frequency of them, and explains possible reasons behind the detected topics comparing with official and published reports.

We manually labeled the negative topics (concerns) for each of the 13 days between 10/03/2015 and 10/15/2015. For example, if “road”, “damage”, “roadway”, “st”, and “exit” are in a topic, we labeled the topic “Road Damage”. Table 3 shows some examples of the detected topics. It is worth mentioning that the quantity of topics/day were more than 11 topics but we assigned related topics to one category. For example, if some topics represent injured people and victims, we assigned just “Victims” label to each of those topics.

The analyzed topics were categorized into 11 unique ones with different frequencies (Table 4). This analysis indicates that damages and animal support are the most and the least discussed negative topics, respectively. Moreover, figure 3 shows that the number of unique negative topics per day has changed from 2 to 8 during the time period considered for this research. The highest number of topic diversity is on the fourth, the fifth, and the sixth days, and the lowest number of topic diversity is on the last day. This figure indicates growing rate of topic diversity between the first day and the sixth day, and declining rate of topic diversity between the seventh day and the last day. It seems that SC disaster responders have tried to reduce people’s negative concerns between the sixth day and the thirteenth day.

Table 5 shows the topics’ distribution for each of the 13 days. We compare our results with official and published reports to explain possible reasons behind the detected topics.

“Victims” is the most frequent negative topic that is about the people who were harmed, injured, or killed during the SC flood. South Carolina Emergency Management Department (SCEMD) develops and coordinates the South Carlina emergency management program for effective response to disasters. SCEMD reported that 19 people were killed and more than 1,500 were rescued from water scemd .

“Damage & Costs” is the next high frequent topic that describes financial damage. SCEMD reported that 28,162 people received $2.2 billion aid through FEMA (Federal Emergency Management Agency) and SCEMD. In addition, more than 3000 collisions scemd and more than 73,000 damaged structures zurich were recorded during the SC flood.

The next topic is “Drinking Water”. During the 2015 SC flood, the drinking water system was collapsed in some area such as Columbia (South Carolina’s capital city) pjstar and thousands of SC residents did not have water for days state1 . In addition, boil water advisory was issued for some days in some parts of South Carolina cola . With respect to “Insurance” topic, most SC residents did not have flood insurance in 2015 and FEMA allocated more than $89 million for individual assistance prnews . In addition, it was suggested to increase insurance penetration and accessibility in South Carolina zurich .

The next topic is “Homelessness.” Thousands of SC residents were displaced and hundreds of people were staying in shelters state1 . This topic covers two groups of people. First people who were forced to leave their home and the second people who did not have a permanent place to live wbtv . This topic is an interesting one because SCEMD report did not mention this topic scemd .

“Road”, “Roof”, and “Bridge” damages are the next three issues. South Carolina has the fourth largest state-owned highway system with 41,000 miles of road and 8,400 bridges in U.S.A statelib . More than 500 roads and bridges were damaged during the flood scemd , and 8 out of the 19 deaths occurred on a flooded road weather . The SC flood caused expensive roof damage. The estimates placed the repair cost close to $137 million dollars greenville ; postandcourier ; zurich . Hundreds of bridges were damaged and closed in October 2015 scemd .

“Flood Report” topic shows the reports and warnings that were retweeted (rt). 41 warning were issued during the flood, and different contents such as images from the floods, local storm reports, and areas of major flooding were shared weather . The next topic is “Power Loss”. Thousands of people did not have power for some days during the flood state1 . Finally, “Animals” was the last and the least discussed topic in the negative tweets. This topic shows main public concerns for animals. During the 2015 SC flood, hundreds of pets were rescued cnbc and some of them lost their homes masslive . Supporting animals was one of the topics mentioned in tweets on the second day. This is the second interesting topic that was not addressed in the SCEMD’s report. On the other side, we saw that food was the only topic that was in the SCEMD’s report but not in our analysis. More than 2 million meals were served scemd and it seems that this topic wasn’t among main concerns during the flood.

The findings show that this research helps situational awareness in some directions. The first one is providing a big picture of damages and public concerns. For example, detecting damage reports by our framework shows show the large size of damages because LDA identifies topics discussed by a large number of users. That big picture can aid the state and federal agencies to have a better-cost estimation and resource allocation. This study can give the agencies a list of priorities for each day. For example, we think that the considering and addressing the insurance concern in the first five days were an excellent strategy to calm the people without an immediate cost.

Comparing to other studies, this study has several benefits. First, TwiSA is a fast, real-time, and cost effective framework. The second benefit is combining sentiment analysis and topic modeling methods. While other studies have used the two methods separately without a connection between them in SA analysis, this research provides a flexible framework that has a potential to embed other techniques such as utilizing deep learning for sentiment analysis and LDA. The next benefit is focusing on disagreeable or negative experiences, instead of all experiences. The third advantage is introducing new issues such as insurance in a disaster, which were not considered in other studies and confirmed by official reports. The next asset is to have a dynamic perspective exploring the disaster issues during a time frame.

Accessing high quality and relevant real-time data during a natural disaster can increase safety and reduce damage and social effects kitchin2014real . Every user is a sensor and contributor in social media and can generate valuable and immediate real-time data to develop better situational awareness. However, social media users generate a huge amount of data that needs to be summarized to provide a big picture in a disastrous situation.

The opinion survey is a traditional method to analyze public opinion; however, this expensive method should be implemented after natural disasters. In addition, a low number of people participate in the data collection and data analysis steps takes a considerable amount of time urcan2012flood .

The growth of social media has provided a great opportunity to track public opinion. Big real-time social media data can help disaster managers to develop a better SA during a natural disaster. This research proposes an analytical framework to detect and track leading public concerns on social networks such as Twitter to provide a better picture of a disaster. TwiSA combines sentiment analysis and topic modeling to handle a hug number of tweets and to disclose negative public concerns.

This paper found that “victims”, “damages”, “drinking water”, “insurance”, “flood report”, “power loss”, ‘homelessness”, and “animals” are the topics that were discussed in negative tweets during the flood. We also found two topics, “homelessness”, and “animals”, were not mentioned in the official reports after the flood. It seems that human and non-human damages, drinking water, and insurance were the most discussed topics during the SC flood. Even though TwiSA was applied to the 2015 SC flood twitter data, it is not limited to this context and can be applied to other natural disasters and emergency situations.

This research shows that social media are valuable data sources to explore people concerns during a natural disaster. TwiSA can help disaster management teams to find public’s concerns in the fastest way to develop a better real-time crisis management plan. The proposed framework can be used not only during natural disasters but also after natural disasters as a post-event methodology review capability (PERC) for future disaster risk reduction. This research has applications for policymakers and disaster risk management experts in developing a better strategy for analyzing social media contents and responding to it. It also helps social scientific research in developing Public Service Announcement (PSA) and unveiling public opinions weather .

Our future directions center on minimizing the impact of noises and misinformation, and analyzing positive and neutral tweets. A potential limitation of our approach is that LIWC tool doesn’t usually detect emotions, misspellings, colloquialisms, foreign words, sarcasm, and abbreviations pennebaker2015development . Another limitation is related to the tweets containing wrong or inaccurate information and spams najafabadi2017research . LDA detects major topics, not every single of topics, in a corpus. Therefore, we assumed that the massive number of tweets in a disaster buries noises. While we believe that the noises don’t have a significant impact on the results, we will utilize some approaches to reduce the possible effects of the noises. The first approach is removing retweets and the tweets containing URL to avoid spams. The second approach is detecting unrelated or inaccurate topics and then eliminating the tweets whose primary topic is among the irrelevant or inaccurate topics. The primary topic is the topic that has the highest probability for a tweet. The next approach is using bot detection tools such as Botometer ferrara2016rise to remove social bots spreading misinformation. We believe that the future plan will address these limitations to provide higher quality data and new insights to SA.

{acks}

This research is supported in part by the University of South Carolina Office of the Vice President for Research through the 2015 SC Floods Research Initiative. The authors are grateful for this support and all opinions, findings, conclusions and recommendations in this article are those of the authors and do not necessarily reflect the views of the funding agency.

swineflu: Twitter predicts swine flu outbreak in 2009.

In International Conference on Electronic Healthcare. Springer, pp. 18–26.

• (25)

Frank W, Karami A and Vanessa K.

Characterizing diseases and disorders in gay users’ tweets.

In Proceedings of the Southern Association for Information Systems (SAIS). Atlanta, Georgia.

• (26)

Karami A, Dahl AA, Turner-McGrievy G et al.

Characterizing diabetes, diet, exercise, and obesity comments on twitter.

International Journal of Information Management 2018; 38(1): 1–6.

• (27)

Karami A, Webb F and Kitzie VL.

Characterizing transgender health issues in twitter.

Proceedings of the Association for Information Science and Technology 2018; 55(1): 207–215.

• (28)

Shaw Jr G and Karami A.

Computational content analysis of negative tweets for obesity, diet, diabetes, and exercise.

Proceedings of the Association for Information Science and Technology 2017; 54(1): 357–365.

• (29)

Shaw G and Karami A.

An exploratory study of (#) exercise in the twittersphere.

In Proceedings of the iConference. Washington, DC.

• (30)

Abel F, Hauff C, Houben GJ et al.

Twitcident: fighting fire with information from social web streams.

In Proceedings of the 21st International Conference on World Wide Web. ACM, pp. 305–308.

• (31)

Sinnappan S, Farrell C and Stewart E.

Priceless tweets! a study on twitter messages posted during crisis: Black saturday.

ACIS 2010 Proceedings 2010; 39.

• (32)

Avvenuti M, Cresci S, Marchetti A et al.

Ears (earthquake alert and report system): a real time decision support system for earthquake crisis management.

In Proceedings of the 20th ACM SIGKDD international conference on knowledge discovery and data mining. ACM, pp. 1749–1758.

• (33)

Sakaki T, Okazaki M and Matsuo Y.

Tweet analysis for real-time event detection and earthquake reporting system development.

IEEE Transactions on Knowledge and Data Engineering 2013; 25(4): 919–931.

• (34)

Earle PS, Bowden DC and Guy M.

Twitter earthquake detection: earthquake monitoring in a social world.

Annals of Geophysics 2012; 54(6).

• (35)

Sakaki T, Okazaki M and Matsuo Y.

Earthquake shakes twitter users: real-time event detection by social sensors.

In Proceedings of the 19th international conference on World wide web. ACM, pp. 851–860.

• (36)

Robinson B, Power R and Cameron M.

A sensitive twitter earthquake detector.

In Proceedings of the 22nd International Conference on World Wide Web. ACM, pp. 999–1002.

• (37)

Miyabe M, Miura A and Aramaki E.

Use trend analysis of twitter after the great east japan earthquake.

In Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work Companion. ACM, pp. 175–178.

• (38)

Doan S, Vo BKH and Collier N.

An analysis of twitter messages in the 2011 tohoku earthquake.

In International Conference on Electronic Healthcare. Springer, pp. 58–66.

• (39)

Oh O, Kwon KH and Rao HR.

An exploration of social media in extreme events: Rumor theory and twitter during the haiti earthquake 2010.

In International Conference on Information Systems.

• (40)

Hughes AL, St Denis LA, Palen L et al.

Online public communications by police & fire services during the 2012 hurricane sandy.

In Proceedings of the 32nd annual ACM conference on Human factors in computing systems. ACM, pp. 1505–1514.

• (41)

Gupta A, Lamba H, Kumaraguru P et al.

Faking sandy: characterizing and identifying fake images on twitter during hurricane sandy.

In Proceedings of the 22nd international conference on World Wide Web. ACM, pp. 729–736.

• (42)

Takahashi B, Tandoc EC and Carmichael C.

Communicating on twitter during a disaster: An analysis of tweets during typhoon haiyan in the philippines.

Computers in Human Behavior 2015; 50: 392–398.

• (43)

Sreenivasan ND, Lee CS and Goh DHL.

Tweet me home: Exploring information use on twitter in crisis situations.

In International Conference on Online Communities and Social Computing. Springer, pp. 120–129.

• (44)

Palen L, Starbird K, Vieweg S et al.

Twitter-based information distribution during the 2009 red river valley flood threat.

Bulletin of the American Society for Information Science and Technology 2010; 36(5): 13–17.

• (45)

de Albuquerque JP, Herfort B, Brenning A et al.

A geographic approach for combining social media and authoritative data towards identifying useful information for disaster management.

International Journal of Geographical Information Science 2015; 29(4): 667–689.

• (46)

Ashktorab Z, Brown C, Nandi M et al.

Tweedr: Mining twitter to inform disaster response.

In Proceeding of the 11th International Conference on Information Systems for Crisis Response and Management (ISCRAM). ACM, pp. 77–80.

• (47)

Dashti S, Palen L, Heris MP et al.

Supporting disaster reconnaissance with social media data: a design-oriented case study of the 2013 colorado floods.

In Proceedings of the 11th International ISCRAM Conference. pp. 18–21.

• (48)

Herfort B, Schelhorn SJ, Albuquerque JPd et al.

Does the spatiotemporal distribution of tweets match the spatiotemporal distribution of flood phenomena? a study about the river elbe flood in june 2013.

In International Conference on Information Systems for Crisis Response and Management, 11. The Pennsylvania State University.

• (49)

Murthy D and Longwell SA.

Twitter and disasters: The uses of twitter during the 2010 pakistan floods.

Information, Communication & Society 2013; 16(6): 837–855.

• (50)

Shaw F, Burgess J, Crawford K et al.

Sharing news, making sense, saying thanks.

Australian Journal of Communication 2013; 40(1): 23.

• (51)

Kongthon A, Haruechaiyasak C, Pailai J et al.

The role of twitter during a natural disaster: Case study of 2011 thai flood.

In 2012 Proceedings of PICMET’12: Technology Management for Emerging Technologies. IEEE, pp. 2227–2232.

• (52)

Twitter.

https://dev.twitter.com/overview/documentation, 2017.

Accessed: 2017-28-02.

• (53)

GNIP.

https://gnip.com/, 2017.

Accessed: 2017-28-02.

• (54)

Taboada M, Brooke J, Tofiloski M et al.

Lexicon-based methods for sentiment analysis.

Computational linguistics 2011; 37(2): 267–307.

• (55)

Khan AZ, Atique M and Thakare V.

Combining lexicon-based and learning-based methods for twitter sentiment analysis.

International Journal of Electronics, Communication and Soft Computing Science & Engineering (IJECSCSE) 2015; : 89.

• (56)

Medhat W, Hassan A and Korashy H.

Sentiment analysis algorithms and applications: A survey.

Ain Shams Engineering Journal 2014; 5(4): 1093–1113.

• (57)

Linguistic Inquiry and Word Count (LIWC).

http://liwc.wpengine.com/, 2017.

Accessed: 2017-28-02.

• (58)

Karami A and Zhou L.

Improving static sms spam detection by using new content-based features.

Proceedings of the 20th Americas Conference on Information Systems (AMCIS) Savannah, GA 2014; .

• (59)

Karami A and Zhou L.

Exploiting latent content based features for the detection of static sms spams.

Proceedings of the 77th annual meeting of the association for information science and technology (ASIST) Seattle, WA 2014; .

• (60)

Karami A and Zhou B.

Online review spam detection by new linguistic features.

Proceedings of the iConference Irvine, CA 2015; .

• (61)

Franklin E.

Some theoretical considerations in off-the-shelf text analysis software.

In Proceedings of the Student Research Workshop associated with RANLP. pp. 8–15.

• (62)

Golder SA and Macy MW.

Diurnal and seasonal mood vary with work, sleep, and daylength across diverse cultures.

Science 2011; 333(6051): 1878–1881.

• (63)

Gonçalves P, Araújo M, Benevenuto F et al.

Comparing and combining sentiment analysis methods.

In Proceedings of the first ACM conference on Online social networks. ACM, pp. 27–38.

• (64)

Nolen-Hoeksema S and Morrow J.

A prospective study of depression and posttraumatic stress symptoms after a natural disaster: the 1989 loma prieta earthquake.

Journal of personality and social psychology 1991; 61(1): 115.

• (65)

Yin D, Bond SD and Zhang H.

Anxious or angry? effects of discrete emotions on the perceived helpfulness of online reviews.

MIS Quarterly 2014; 38(2): 539–560.

• (66)

Zhang D, Luo T and Wang D.

Learning from lda using deep neural networks.

In Natural Language Understanding and Intelligent Applications. Springer, 2016.

pp. 657–664.

• (67)

Mohammadi M, Al-Fuqaha A, Sorour S et al.

Deep learning for iot big data and streaming analytics: A survey.

IEEE Communications Surveys & Tutorials 2018; .

• (68)

Zhao W.

Research on the deep learning of the small sample data based on transfer learning.

In AIP Conference Proceedings, volume 1864. AIP Publishing, p. 020018.

• (69)

Chilimbi TM, Suzue Y, Apacible J et al.

Project adam: Building an efficient and scalable deep learning training system.

In OSDI, volume 14. pp. 571–582.

• (70)

Erhan D, Bengio Y, Courville A et al.

Why does unsupervised pre-training help deep learning?

Journal of Machine Learning Research 2010; 11(Feb): 625–660.

• (71)

Severyn A and Moschitti A.

Twitter sentiment analysis with deep convolutional neural networks.

In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, pp. 959–962.

• (72)

Mahmud M, Kaiser MS, Hussain A et al.

Applications of deep learning and reinforcement learning to biological data.

IEEE transactions on neural networks and learning systems 2018; 29(6): 2063–2079.

• (73)

Blei DM, Ng AY and Jordan MI.

Latent dirichlet allocation.

Journal of machine Learning research 2003; 3: 993–1022.

• (74)

Mcauliffe JD and Blei DM.

Supervised topic models.

In Proceedings of the Advances in neural information processing systems. pp. 121–128.

• (75)

Hong L and Davison BD.

Empirical study of topic modeling in twitter.

In Proceedings of the first workshop on social media analytics. ACM, pp. 80–88.

• (76)

Karami A and Gangopadhyay A.

Fftm: A fuzzy feature transformation method for medical documents.

Proceedings of BioNLP Baltimore, MD 2014; .

• (77)

Karami A, Gangopadhyay A, Zhou B et al.

A fuzzy approach model for uncovering hidden latent semantic structure in medical text collections.

Proceedings of the iConference Irvine, CA 2015; .

• (78)

Lu Y, Mei Q and Zhai C.

Investigating task performance of probabilistic topic models: an empirical study of plsa and lda.

Information Retrieval 2011; 14(2): 178–203.

• (79)

Karami A and Pendergraft NM.

Computational analysis of insurance complaints: Geico case study.

International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation Washington DC 2018; .

• (80)

Karami A.

Fuzzy Topic Modeling for Medical Corpora.

PhD Thesis, University of Maryland, Baltimore County, 2015.

• (81)

Karami A, Gangopadhyay A, Zhou B et al.

Fuzzy approach topic discovery in health and medical corpora.

International Journal of Fuzzy Systems 2018; 20(4): 1334–1345.

• (82)

Karami A, Gangopadhyay A, Zhou B et al.

Flatm: A fuzzy logic approach topic model for medical documents.

In 2015 Annual Meeting of the North American Fuzzy Information Processing Society (NAFIPS). IEEE.

• (83)

Mallet.

http://mallet.cs.umass.edu/topics.php, 2017.

Accessed: 2017-28-02.

• (84)

Steyvers M and Griffiths T.

Probabilistic topic models.

Handbook of latent semantic analysis ; 427(7): 424–440.

• (85)

JournalStar.

http://www.pjstar.com/article/20151008/NEWS/151009377, 2015.

Accessed: 2017-28-02.

• (86)

Senna A and Collins J.

South Carolina flood: Thousands without running water.

http://www.thestate.com/news/nation-world/national/article37801893.html, 2015.

Accessed: 2017-28-02.

• (87)

Wallace A.

Curfew, boil water advisory issued as flooding continues.

http://www.coladaily.com/2015/10/04/officials-issue-warnings-floods-sweep-columbia-area/, 2015.

Accessed: 2017-28-02.

• (88)

PRNewswire.

Massive flooding in south carolina; many consumers lack flood insurance, 2015.

http://www.prnewswire.com/news-releases/massive-flooding-in-south-carolina-many-consumers-lack-flood-insurance-300154495.html, Accessed: 2017-28-02.

• (89)

Ohnesorge S.

Homeless man loses “home” in SC flooding.

http://www.wbtv.com/story/30203344/homeless-man-loses-home-in-sc-flooding, 2015.

Accessed: 2017-28-02.

• (90)

2015 DOT Annual Report.

http://dc.statelibrary.sc.gov/bitstream/handle/10827/21838/DOT_Annual_Report_2015-2016.pdf?sequence=1&isAllowed=y, 2016.

Accessed: 2017-28-02.

• (91)

US Department of Commerce.

The historic south carolina floods of october 1–5, 2015, 2015.

http://www.weather.gov/media/publications/assessments/SCFlooding_072216_Signed_Final.pdf, Accessed: 2017-28-02.

• (92)

GreenvileOnline.

Storm system brings property damage, flooding, record rains, 2015.

http://www.greenvilleonline.com/story/news/local/2015/06/03/storm-system-brings-property-damage-flooding-record-rains/28403225/, Accessed: 2017-28-02.

• (93)

John McDermott.

S.C. insurers brace for surge of flood claims.

http://www.postandcourier.com/business/s-c-insurers-brace-for-surge-of-flood-claims/article_d67b6317-2145-5063-ba1b-61850ded7c0e.html, 2015.

Accessed: 2017-28-02.

• (94)

Ferris R.

South Carolina’s rain and floods, by the numbers.

http://www.cnbc.com/2015/10/08/south-carolinas-rain-and-floods-by-the-numbers.html, 2015.

Accessed: 2017-28-02.

• (95)

Lu F.

Dogs left homeless after S.C. flooding arrive in Mass. in search of new owners.

http://www.masslive.com/news/index.ssf/2015/10/dogs_left_homeless_after_sc_fl.html, 2015.

Accessed: 2017-28-02.

• (96)

Urcan I.

Flood hazards perception: The result of an opinion survey made in the little towns from lower aries corridor.

Journal of Catastrophe 2012; 11: 202–210.

• (97)

Pennebaker JW, Boyd RL, Jordan K et al.

The development and psychometric properties of liwc2015.

Technical report, 2015.

• (98)

Najafabadi MM.

A research agenda for distributed hashtag spoiling: Tails of a survived trending hashtag.

In Proceedings of the 18th Annual International Conference on Digital Government Research. ACM, pp. 21–29.

• (99)

Ferrara E, Varol O, Davis C et al.

The rise of social bots.

Communications of the ACM 2016; 59(7): 96–104.