Sensing the public's reaction to crime news using the 'Links Correspondence Method'.

Public media such as TV or newspapers, paired with crime statistics from the authority, raise awareness of crimes within society. However, in today's digital society, other sources rapidly gain importance as well. The Internet and social networks act heavily as information distribution platforms. Therefore, this paper aims at exploring the influence of the social Web service Twitter as an information distribution platform for crime news. In order to detect messages with crime-related contents, the Links Correspondence Method (LCM) is introduced, which gathers and investigates Twitter messages related to crime articles via associated Web links. Detected crime tweets are analysed in regard to the distance between the location of an incident and the location of associated tweets, as well as regards demographic aspects of the corresponding crime news. The results show that there exists a spatial dependency regarding the activity space of a user (and the crime-related tweets of this user) and the actual location of the crime incident. Furthermore, the demographic analysis indicates that the type of a crime as well as the gender of the victim has great influence on whether the crime incident is spread via Twitter or not.

Introduction

Media and the public's concern about crime

Media, and in particular TV and newspapers, affect people's feelings, impressions, judgments and viewpoints about crime. Many scholars have supported with their findings the positive association between media and crime perception; yet this is not a new research field. Interests on the Media's effects on crime perception have started since the beginning of the last century. Fenton (1910) argues that human habits are unconsciously and consciously acquired as we grow up and are influenced, among many other factors, by the newspapers as well. Based on this logical assumption, Fenton was interested to investigate whether newspaper publications had an agenda setting effect about anti-social activity or influenced and unwittingly promoted this type of behaviour.

Current research still acknowledges the positive association, but shifted its interest to the effects on crime perception in general. For instance, fear of crime is an essential and timeless area of research in criminology. Interestingly enough, the findings of previous studies disagree considerably in this scientific domain. Significant relationships between crime news and fear of crime (FOC) or fear of victimisation have been proved in many studies (Chiricos, Eschholz, & Gertz, 1997; Jaehnig, Weaver, & Fico, 1981; Smolej & Kivivuori, 2006; Williams & Dickinson, 1993).

Furthermore, it seems that crime type also plays important role in people's fear of crime and according to Jaehnig et al. (1981) this is more closely associated when newspapers emphasise on violent crimes. However, among the scholars there is considerable disagreement about the extent to which the media reflect or form opinions (Duffy & Rowden, 2005). Though, according to Williams and Dickinson (1993), Schlesinger, Tumber, and Murdock (1991), and Duffy and Rowden (2005), our study area (UK) seems to be affected.

Geographic information of crimes

Crime incidents feature geographic aspects. These aspects appear in form of a geographic location related to the crime scene. But not only the crime scene as such is georeferenced, the patterns of activities, routes and behaviour of victims and criminals feature geographical aspects as well. This observation manifested itself within the literature as the crime pattern theory (Brantingham & Brantingham, 1984, 1993). It is reasoned that crime incidents happen at predictable locations.

These locations are based on the geographical intersection of the criminals so-called awareness space and crime chances (Wortley & Mazerolle, 2013). As the development of technology advanced, Information and Communication Technology (ICT)-based systems found their ways to various application areas, in case of police work and crime analysis in form of geographic information systems (GIS). These systems support the police by proving the possibility to not only display temporal and spatial crime data, but also to develop models to identify crime hot spots and in further consequence to predict potential crime areas as well (Sherman, Gartin, & Buerger, 1989).

Not only the crime incidents as such generate geographic information, the public reactions to them do as well. Media and in particular the newspapers present a major public information source about crimes. In a survey related to the fear of crime produced by different cartographic practices, 58% of the respondents reported that newspapers are the primary source for crime information (Groff et al., 2005). Likewise, in the UK more than 60% of people are regular newspaper readers (Duffy & Rowden, 2005). As this information is spread among the community, people react to it via expressing their opinions and reactions to different topics, news, or events in social networks. As these reactions are usually in textual, verbatim description, as well as the profile information of the associated users, natural language processing (NLP)-based techniques and methods are crucial for retrieving and analysing these data. The use of NLP in GIS is an emerging research field. Current research projects cover heterogeneous topics from location-based services (Haav, Kaljuvee, Luts, & Vajakas, 2009), geoparsing and geocoding (Kounadi, Lampoltshammer, Leitner, & Heistracher, 2013) up to concepts in the field of human computer interaction (Thomas, Sripada, & Noordzij, 2012). A comprehensive overview of on-going research in this area can be found in Lampoltshammer (2012). This paper aims at exploring the influence of the social Web service Twitter as an information distribution platform for crime news.

Methodology

In this section the Links Correspondence Method (LCM) is introduced. LCM utilises online influential sources as well as sources of opinion expression in order to estimate the degree or variations of public's interest in particular topics. In this study, the investigated topic is crime articles, the selected influential sources are online versions of newspapers and the source of opinion expression is Twitter.

Since our topic is dealing with crime information, which generates negative emotions, we employ the LCM to estimate patterns, variations and the degree of concern related to specific crime articles.

The step-by-step outline of the Links Correspondence Method is shown in Fig. 1. It consists of two main phases: i) data extraction, and ii) the analysis of associations between the crime incidents and the crime-related tweets. In the first phase, we extract texts from sources of opinion expressions (in our case Twitter messages) that contain Web links pointing to influential sources. The texts provide information about the location and also the links' popularity which is expressed by the degree of occurrences of these links. In addition, the location of the topic that is discussed (in our case locations of crime incidents) and other characteristics of that topic are extracted from the links. The first phase results in four derivatives: a) the links' locations, b) the texts' locations, c) the topic analysis of the links, and d) the frequency of the links within the texts.

Fig. 1

Links Correspondence Method: Influential sources and sources of opinion expression are utilised to assess the interest in particular topics.

In the second phase, a spatial point pattern analysis is performed over all locations. The aim is to understand whether there exists any spatial dependence associated with people's interest in the examined topic. Depending on the information that can be extracted from the sources of opinion expression, the analysis can be performed at a local level or at a global level. The second step of this phase is to analyse the specific characteristics of the topic in order to infer the degree of interest in some characteristics than others. This is achieved by regression analysis employing the characteristics as explanatory values for link frequencies.

Data for the analysis

Selection of crime news from online versions of newspapers

The study area that was selected for the application of the LCM is London, United Kingdom (UK). The crime news is selected based on three conditions: i) an incident occurred during the year 2012, ii) an incident was either of type acquisitive or violent crime, and iii) the crime article included the location information of at least the street name and the municipality. The two distinct crime types were selected to detect variations of concerns associated to different types of crimes. Finally, the location resolution was set to the street level for global as well as local spatial analysis. Nevertheless, for two crime incidents we allowed the resolution of the location information to be the municipality level in order to obtain a meaningful sample size.

Based on each street we calculated the coordinates of their centre points and from the municipalities we calculated the coordinates of the spatial mean centre. In regard to the online versions of newspapers from which we obtained the crime articles' links, our strategy consisted of two steps. First, we selected a total of 30 crime news, consisting of 15 acquisitive crimes and 15 violent crimes. The crime news selection was based on several of the most popular online versions of newspapers in the UK. The popularity is assessed by readability rates, which seem to vary across different Web sources.

Since we wanted to capture as many worldwide Twitter messages as possible, we selected UK newspapers based on national and global ratings (Comscoredatamine.com, 05.04.2014; eBizMBA.com, 05.04.2014; Journalism.co.uk, 05.04.2014; PressGazette.co.uk, 05.04.2014; ThePaperboy.com, 05.04.2014). The online versions of newspapers employed in this study are: The Telegraph, Mail Online, The Independent, The Times, and The Guardian. From each newspaper, we selected five crime articles (different in each newspaper). For some cases, we were not able to find five articles that would satisfy our selection criteria; therefore we additionally utilised the local newspapers Metro and London Evening Standard. Eventually, we obtained thirty primary links associated to the thirty crime incidents. Subsequently, we employed each article's key words on a Google-based Internet search and extracted up to four additional online versions of newspapers' links about the same crime incident. In total we tested 113 links that referred to the 30 crime incidents.

The Twitter database

Digital micro blogging and social media platforms have recently emerged as a new public forum, well suited as source of opinion expression (Pak & Paroubek, 2010). The social media community is huge and still growing. According to the market trend company eMarketer (05.04.2014), almost one of four people globally is currently utilising some kind of social platform.

Forecasts state that this trend will increase up to one out of four people by 2017. In our study, we utilise messages generated by the users of the Twitter networking service. We retrieved Twitter data in two independent ways, creating separate and to some extent different datasets. The first one was collected through one of the commercial social media aggregators (Topsy, 05.04.2014). This data set was filtered for all worldwide tweets that contained one of the previously identified links related to crime news. Messages within this data set could only be georeferenced based on location information declared within a user's profile. This level was sufficient to differentiate the response rate from different countries. However, for the local scale analysis, this level of granularity was too large. Therefore, we utilised the complementary dataset of all georeferenced messages that were posted within UK in 2012, collected directly from the Twitter Streaming API (Twitter, 05.04.2014). This snapshot from the entire Twitter database was cleaned from errors and artificial tweeting noise similarly as described by Hawelka et al. (2014). Both datasets were joined, based on unique identifiers of Twitter users (user names). This procedure enabled to amend crime-related messages from the first dataset by the home location of their authors estimated based on the locations recorded in the second dataset (detailed procedure is described in the further section of the paper).

Results and discussion

Global dependence

In the Twitter social network there exist, apart from the individuals' accounts, accounts by online versions of newspapers or other organisations related to news and media as well. Hence, the list of individual messages was processed to exclude messages from such accounts. In total, we obtained 3116 worldwide messages. For the global spatial analysis, we employed the location provided by the users within their profiles. Subsequently, this information was processed to remove empty or irrelevant information. Afterwards, the country level for each message was inferred. In the end, we retrieved 1821 messages. The results are presented in Fig. 2 by a cartogram that alters the size of the countries based on the crime tweets' density.

Fig. 2

Global distribution of London crime news posted on Twitter. The areal size of the countries represents the proportion of tweets of the London crime tweets dataset sent from the particular country.

The messages were sent from 80 countries in total; however, 15 countries accounted for more than 90% of the total messages. The messages' cumulative frequency for these countries is shown in Fig. 2 and, as it was expected, UK users sent the majority of those (60.8%).

There seem to be two factors that influence the frequency of the messages; namely the spoken language and the proximity to the crime location. Among the countries with the highest frequency are neighbouring countries located in Europe, indicating that the information is spatially dependent. This means that the closer the country is to the crime incident the more frequent are the messages. We can observe this e.g. for France and Germany. Also, countries with a high frequency of crime tweets have either English as one of their official languages such as India or Nigeria, if not the first one such as Canada and USA.

Another important factor that influences the spatial distribution of retrieved messages is the usability of the Twitter among particular nations. Some of the countries with top amounts of crime tweets, such as Brazil, are among the top countries in terms of Twitter accounts (Semiocast, 05.04.2014). Kulshrestha, Kooti, Nikravesh, and Gummadi (2012) analysed in their work the distribution of Twitter users world-wide and compared the adoption rate to the socio-economic status in terms of a countries population. The results revealed a highly unequal distribution paired with a high correlation of the adoption rate and the population status. However, easy and free access to Twitter or other social media services is not guaranteed in all countries. Abrol and Khan (2010) indicated in their work that the number of people using Twitter generally increases with the number of inhabitants of a country. However, free access is not for granted in every country and therefore Twitter user counts may go down on a global perspective. In China for example, the government tries to ‘protect’ citizens by forbidding access to social services such as Twitter (Wang, 2012). Another example these days can be found in Turkey, where the government, despite a clear court decision, refuses to provide free access to social media to its people (Fraser, 03.04.2014).

Local dependence

This part of the analysis aims at the exploration of local spatial relationships between locations of users who posted the crime messages and the locations of the crime incidents as such. Hence, a better estimation of the users' location is required. As the initial hypothesis it was assumed that the closer the crime incident is to a user's activity space, chances are higher that the user is concerned about the crime incident and posts about it. Consequently, we applied a method that estimates a user's home location based on the complete database of georeferenced tweets posted in the UK within the year of 2012 (Fig. 3).

Fig. 3

Home estimation method: Extraction of messages from Twitter's georeferenced database based on user name and application of spatial clustering to estimate the home locations of the users with crime tweets.

‘Activity space’ as a term has been used in several geographical studies in a variety of applications such as exploring the accessibility to healthcare facilities (Sherman, Spencer, Preisser, Gesler, & Arcury, 2005) or to describe the spatial patterns of offenders (Brantingham & Brantingham, 2008, p. 78). In this context it is used to describe the areas of the night daily activities of the users.

Space-time prism (Miller, 1991) is a very popular concept in geography nevertheless it includes the time aspect as well. The altitude of Fig. 4 shows the density of the locations and is not related with time. Hence, it was the ‘routine activity’ theories that inspired us to use the term ‘activity space’ and not the space-time prism literature. Also, space-time prism has been used to assess accessibility whereas we are more interested in the generic travelling behavioural patterns of the Twitter users. The idea of exploring the users activity spaces based on the topic that they discuss (i.e. type or gender of crimes) is an interesting idea, but goes beyond the scope and analytical effort, of this study. The activity space figure is used for illustrative purposes that can further support the results of the distance decay model.

Fig. 4

Examples of weekday night activity spaces of users and locations of crime incidents. The activity spaces are shown as the density of activity locations over the street network.

First, we extracted all georeferenced messages of users who posted links from the crime articles. Messages originating outside of the city of London were removed. As concerning the users' home locations it was assumed that Twitter messages outside working hours and messages not being sent on weekends represent a reasonable subset. The remaining locations were considered as potential home locations and were separated by user (in total 55 users in London). Then for each user we applied two spatial analytical steps. First, a spatial clustering was performed (nearest neighbour hierarchical clustering, (Everitt, Landau, & Leese, 1993)) and the locations within the clusters were selected. These were defined as weekday-night activity spaces. Afterwards, the highest density cluster was selected and the central point (the point that had the smallest accumulated distance to all other points in the cluster) was calculated, which represents the estimated home location of a user. If the points within the highest density cluster formed more than one sub-clusters, the two steps were re-applied for this cluster only. In of an insufficient amount of points to conduct spatial clustering, we calculated directly the central point. It is worth mentioning that the highest density cluster consisted of more than 70% of all points for the vast majority of users. This implies that after the initial weekend and daytime hours' exclusion the users' posting location patterns are heavily concentrated around a few or one location only. Finally, the users' estimated home locations were joined with the crime messages resulting in 61 georeferenced messages that were analysed against the locations of the respective 30 crime incidents.

A primary visualisation was created for the weekday-night activity spaces of the users in order to observe spatial relationships between activity spaces and locations of crime incidents. We observed three types of relationships, which are shown in Fig. 4. Some of the crime incidents were very close to the activity spaces (green density walls (in the web version), Fig. 4b), whereas others were inside the activity space (blue density walls (in the web version), Fig. 4c). Nevertheless, a few of the crime incidents were located far away of any activity space (yellow density walls (in the web version), Fig. 4a).

To understand precisely if any spatial dependence existed within our dataset we calculated the Euclidean distances from the estimated home locations of the authors of Twitter messages to the locations of the crime incidents. The locations were grouped into distance intervals of 2 km each and then the cumulative percentage was calculated as shown in Fig. 5a. The results show that there is a higher concentration of crime tweets that are being sent by users with an estimated home location in a close proximity of the incident comparing to the number of tweets sent by users living further away. More specifically, 75% of all crime tweets are sent by users living within a 16 km buffer zone from the crime incident though the rest are sent from a buffer zone of more than 16 km up to 38 km. Also, in order to detect if there is any decreasing trend in the data the frequency of the crime tweets compared to the distance from the crime incident is further analysed (Fig. 5b). The two functions with a decreasing trend that fitted significantly the data points are the linear and the quadratic function. However, there is an increase in the R squared value from the linear to the quadratic model which indicates that the latter model fits the data somewhat better than the former (linear: R2 = 0.644, quadratic: R2 = 0.662). That is due to the curve effect of the quadratic function that describes better the initial increasing slope of the points. Nevertheless, the data do not show a strong spatial dependency that would have been observed if a lognormal, exponential or a power model could have been fitted to our data. This could be the result of the small volume of the dataset that was available after the home estimation method since only 61 messages were used for this analysis.

Fig. 5

Distance from users' estimated home locations to the crime incidents' locations. Sub figure (a) at the top shows the cumulative percentage of the crime tweets and the respective distance from the crime incidents whereas sub figure (b) at the bottom shows the distance decay fitted models between the frequency of the tweets and their distance from the crime incidents.

Topic dependence

In the last part of the analysis the crimes' characteristics were grouped as independent variables, which aimed to explain the depended variable of the count of the messages. Here we used the initial 3116 worldwide messages since the restriction of the location attribute was not substantial for this analysis. The characteristics were identified from the crime articles.

We selected only those characteristics from which we found information in the majority of the articles. These are: i) the offender's gender, ii) the offender's age, iii) the victim's gender, iv) the victim's age, and v) the crime type (acquisitive or violent). The technique that was employed was the ordinal logistic regression analysis. The explanatory variables were measured at ordinal (age) and nominal (type and gender) scales and the dependent variable was transformed into an ordinal scale of five classes with increasing order.

First, we employed the Pearson product–moment correlation coefficient for all variables and we found a positive correlation between the two explanatory variables and the messages' frequency. These are the victim's gender (r = 0.663, n = 30, p < 0.01) and the crime type (r = 0.445, n = 30, p < 0.05). The cumulative percentages of these two variables are shown in Fig. 6a and Fig. 6b. Fig.6a shows that there are no violent crimes within our sample that have not been posted on Twitter. On the contrary, one third of the acquisitive crimes have not been posted at all. Similarly, Fig. 6b shows that there are no crimes concerning men being victims that have not been referred on Twitter, whereas 20% of the crimes where the victim was a woman have not been mentioned on Twitter. In general, if a line appears lower than another it indicates that the respective category (men being a victim and type being a violent crime) is associated with higher x values (more tweets) than the other categories (women being a victim and type being an acquisitive crime).

Fig. 6

Examples of weekday night activity spaces of users and locations of crime incidents. The activity spaces are shown as the density of activity locations over the street network.

Then, we utilised these two variables in the ordinal regression model. The results of the model revealed that if the victim is a woman or the crime type is acquisitive, there are less chances for the crime incident to be posted on Twitter than if the victim is a man or the type of the crime incident is violent. In particular, for an additional victim being women, controlling for crime type, the odds of that incident being in a higher messages' class are lower by 98.17% (Wald2 (1) = 8.831, p = .003). Similarly, for an additional acquisitive crime, controlling for victim gender, the odds of that incident being in a higher messages' class are lower by 97.98% (Wald2 (1) = 7.596, p = .006).

Future recommendations

The details in this section are describing the initial methodology we developed for this paper. As this approach turned out not to be successful, the approach was reworked to its current state. However, we still see some valuable information in our experiences that we would like to share with the community.

At the very beginning of this research work, we focussed on the available georeferenced Twitter data from London, UK in the year 2012. Unfortunately, the percentage of the georeferenced tweets is only about 1–2% of the total amount of tweets (Morstatter, Pfeffer, Liu, & Carley, 2013). Therefore, it turned out to be difficult to achieve valuable results with this type of dataset only. Nevertheless, we would like to describe and discuss our experiences to provide additional information that could turn out to be useful for upcoming research projects within the community.

To start with, we filtered the entire Twitter database from 2012 to obtain a subset containing only tweets from London, UK (about 11 Mio. entries). Afterwards, a crime vocabulary was applied to distinguish crimes of the types acquisition and violent. In order to create an appropriate crime vocabulary for our research different information sources were used. Firstly, the crime words had to be classified and for doing that we examined the typology used in the following UK-based websites: UKCrimeStats (05.04.2014), UK National Statistics (05.04.2014), Office for National Statistics (05.04.2014), and PoliceUK (05.04.2014) and finally we concluded in the classification provided by the Office for National Statistics. Then, each crime category was enriched by respective formal vocabulary that was extracted from the above-mentioned Web sites along with other sources (Douglas, Burgess, Burgess, & Ressler, 2013; Dressler, 2002; Encyclopedia, 05.04.2014; Wikipedia, 05.04.2014). Furthermore, a second essential part here was to include slang words and expressions, as there is a high likelihood that some users in the informal Twitter social network prefer this writing style for their messages.

For the slang vocabulary, the following Web sources of information were used: Miskatonic (05.04.2014), SlangSearch (05.04.2014), Ask Metafilter (05.04.2014), Classic Crime Fiction (05.04.2014), Mental Floss (05.04.2014), and English Club (05.04.2014). Subsequently, each word or expression had to be examined visually so as to filter ambiguities or oracular words. That was especially the case for the slang words that tend to have formal meanings but are also used in the slang language to define an alternative meaning. Striking examples are ‘the ice’ that is used to describe a stolen diamond or ‘the rabbit’ that defines a prisoner who has a history of escape attempts. After the filtering, the vocabulary consisted of verbs, nouns, terminology and expressions that show an explicit relationship with the crime type they have been assign to.

The final vocabulary was then utilised to filter the London tweets to generate a subset for each of these vocabularies. The acquisition crimes subset featured about 1400 tweets and the violent crimes subset featured about 16,000 tweets.

In order to check that the filtered results were usable, we selected random samples out of both subsets (n = 1000 for each set) and performed a visual interpretation of the tweets.

The results showed that 33% of the acquisition crimes tweets and 15% of the violent crimes tweets were actually related to crimes. We then employed an alternative approach by examining the headlines of newspaper crimes and the main body of the news items to identify the most important keywords for each news item and utilised them as search criteria within the acquisition and violent crime tweets. Based on these keywords, we conducted a similarity analysis by applying the cosine similarity measurement (Xie & Liu, 2008). The highest similarity we could find was at 15%. None of the hits were related to crimes though. We also tried to combine the most prominent keywords as search criteria within the categories. They received 0 results. By this experience we learned that human language may occur very complicated to mimic with automatic processing, especially with short texts such as Twitter messages. However, the above-described approach is promising for other social communication platforms that feature bigger junks of textual information per user.

Discussion and conclusions

This paper investigated the relevance of the social Web service Twitter as an information distribution platform for crime news. To enable the exploration of Twitter data, the Links Correspondence Method (LCM) was introduced to gather and investigate messages related to crime articles via Twitter messages in the area of London, UK. The results show that there exists a spatial dependency regarding the activity space of a user (and the crime tweets of this user) and the actual location of the crime incident. Furthermore, the topic analysis indicates that the type of crime as well as the gender of the victim has great influence on whether the crime incident is spread via Twitter or not. It is important to mention that the current amount of georeferenced tweets is not enough to provide sufficient information about the interrelation of the spread of crime news via Twitter. The LCM closes this gap by fusing non-georeferenced tweets and location information of crime incidents. However, it may not be neglected that the actual number of available tweets strongly influences the outcome and accuracy of the analysis. Still, the LCM provided promising and valuable first results.

When working with Twitter data, a certain bias towards the results has to be acknowledged. The social media service Twitter has a specific profile of users, e.g., young adults and minority groups tend to be overrepresented in tweets as well as urban areas have more tweets associated than rural areas (Smith & Brenner, 2012). Similar findings were also published by Mitchell and Page (2013). Still, Twitter data and social media have a high potential for research-driven work. White and Roth (2010) suggested that social media and in particular Twitter can be used for criminal investigative analysis. From this perspective, the centrality of tweets to the crime incidents seems to be a valuable finding for future research. Our results show that 75% of the geo-located tweets are being sent within a 16 km buffer zone from the crime incidents. Future research in crime investigation may utilise this information and focus on examining messages from social media that are sent in the near areas from the crime locations. To this extend, natural language processing algorithms for clustering unstructured text such as the ‘Latent Dirichlet Allocation topic modelling’ (Blei, Ng, & Jordan, 2003) could reveal to the criminologists and spatial crime analysts valuable information from observers in the neighbourhood where the crime occurred. Considering the amount of information included in worldwide datasets such as Twitter, the centrality of the tweets could facilitate and give a clear direction during the big data mining pre-processing steps.

According to the Crime Survey for England and Wales (Office for National Statistics, 11.04.2014), males are more frequently violent crime victims than women (3.2% male victimisation rate and 1.9% victimisation rate). Furthermore, regarding homicide offences male victims are more than double than female victims (380 male victims and 171 female victims in England and Wales, 2012/13). These findings are in line with the results of our topic dependence analysis. This analysis showed that news about violent crimes with males as victims are being discussed more intensively than those with female victims. The report for crime in England and Wales in 2010/2011 (Chaplin, Flatley, & Smith, 2011) states that people are perceiving themselves less in danger of being a victim of violent crime than the actual numbers of incidents tell. This means that there exists an awareness deficit. Coming back to our results described before, this implies that within the unawareness, violent crimes towards women are even more underrepresented in the discussion and the overall picture. Thus, we suggest as a prevention strategy that the results of the Twitter crime analysis may be used to identify areas with a particular low discussion of violent crimes, in particular crimes towards women, and use this information to dedicate awareness campaigns of various kinds to those particular areas.

Upcoming research work will be dedicated to a larger data set associated to a more specific type of crime to compare the outcomes. Furthermore, additional themes could be fear of crime through altitude analysis, temporal analysis through the dispersion of tweets over time after the incident occurs and their implications, or socio-economic analysis with local built-in environment. The LCM as such is flexible and generic enough to be transferred to any other topic of interest, paired with various kinds of social media.