Is gender-based violence a confluence of culture? Empirical evidence from social media

GBV and culture

GBV is a social ill evident across all the countries irrespective of their economy, language, and demography. However, with country, the type of GBV, its intensity, and the reaction of people vary (Fakunmoju, Bammeke et al., 2017). For example, in the USA, dating violence is more common than in Africa where there are comparatively lesser instances of dating violence (Johnson et al., 2015). On the other hand, in Africa, intimate partner violence is more prominent as compared to North America (https://apps.who.int/iris/bitstream/handle/10665/85239/9789241564625_eng.pdf). This implies that the same GBV is represented differently in a different country. This implies that GBV is a global evil but the context of GBV changes with the country. There have been many research to understand the causes and factors leading to GBV (Jewkes, Levin & Penn-Kekana, 2002; Jewkes, Jama-Shai & Sikweyiya, 2017; Marine & Lewis, 2020). These works have claimed that a country’s culture can characterize the persisting GBV in the country. Every culture has norms, prejudices, and societal rules that design the behaviour of people towards GBV. For example, in Malawi, the concept of polygamy and dowry is evident in the culture, and these perpetuate GBV in Malawi (Bisika, 2008). Similar research in many other countries like UK (Aldridge, 2021), Ethiopia (Le Mat et al., 2019), Cambodia  (Palmer & Williams, 2017), and many other countries (Djamba & Kimuna, 2015; Raj & Silverman, 2002) have highlighted cultural norms which lead to one or other form of GBV. Not only in research but global organisations like WHO (https://www.who.int/violence_injury_prevention/violence/norms.pdf), World Bank (https://thedocs.worldbank.org/en/doc/656271571686555789-0090022019/original/ShiftingCulturalNormstoAddressGBV.pdf) have also highlighted the cultural norms of many countries that influence GBV. The socio-cultural impact is so intense that people even justify instances of GBV as a form of the social norm which cannot be questioned (Piedalue et al., 2020). However, these claims are supported by mere examples and small-scale interview-based data. Thus, the research community lags a data-driven research that justifies the claim with sufficient empirical results. In this research, we do a large-scale analysis of social network content to find evidence of confluence between the culture of a country and GBV. Next, we present the role of social media content in bridging the gap of data for various research.

Social media content

Social media has now become the new language of people, and this has generated a massive amount of data for various research. Social media data has removed the bottleneck of data requirements in numerous applications such as urban computing (Silva et al., 2019), cultural computing (Wang et al., 2017), personality computing (Samani et al., 2018) and many more. Social media content also plays a huge role in understanding people’s views and sentiments during the COVID-19 pandemic (Malagoli et al., 2021). Social media content has already substituted the tedious, time consuming, biased and under-represented survey-based data and has unlocked possibilities for research in many other directions. The utility of social media content increases with the availability of location-tagged data. The location-tagged online content has been used in numerous ethnographic research (Abdullah et al., 2015), cultural research (Cheke et al., 2020), and sociological research (Stewart et al., 2017) in recent days. Social media content is not mere data, but it captures several dimensions of human interests, psychology, and behavior. There have been works which found that the social media content reflects the real-world properties as well. García-Gavilanes, Mejova & Quercia (2014); Garcia-Gavilanes, Quercia & Jaimes (2013) have found that interaction and usage of social networks are dependent on social, economic, and cultural aspects of users. Thus, the real-world behavior of people is also mirrored in social networks. This utility of social media content motivates us to examine the relationship between GBV and a country’s culture through analysis of social media data. For this research, we use Twitter data related to GBV from different countries.

Measuring culture

Culture is an amalgamation of thoughts, beliefs, potential acts, and a lot more. A number of definitions of culture are available from previous works (Hofstede, Hofstede & Minkov, 2005). One of many definitions of culture is “a fuzzy set of assumptions and values, orientations to life, beliefs, policies, procedures and behavioural conventions that are shared by a group of people” (Spencer-Oatey & Franklin, 2012). Culture plays a vital role in many spheres of life, such as behaviour (Huang, Beshai & Yu, 2016) economy (Herrero, Jiménez & Alcalde, 2021), language (Sazzed, 2021), attitude (Shin, Chotiyaputta & Zaid, 2022). In order to ease out research based on culture, there have been several quantification of culture. Hofstede has done one such comprehensive quantification. Hofstede (Hofstede, Hofstede & Minkov, 2005) defines culture in terms of six parameters (Power Distance, Uncertainty Avoidance, Individualism, Masculinity, Long Term Orientation, Indulgence) and quantifies each one for different countries of the world. An extensive set of previous works use Hofstede Dimensions to quantify culture (García-Gavilanes, Mejova & Quercia, 2014; Garcia-Gavilanes, Quercia & Jaimes, 2013; Prakash & Majumdar, 2021). For this research, we also used Hofstede’s dimensions which have been used in a huge number of research to measure culture.

GBV through the lens of social media

Social media provides an uncensored and user-friendly medium for expressing views and opinions (Puente, Maceiras & Romero, 2021). With this, social media has become a platform for self-expression as well as for conducting online campaigns (Martínez, Pacheco & Galicia, 2021). There have been many campaigns on social media related to GBV like the #metoo, #Notokay, #StateOfWomen, #HeForShe and many more (Karuna et al., 2016). These campaigns and freedom of expression on social media have generated huge data related to GBV. The recent campaign of #Metoo observed an unprecedented response from all around the globe, thus, generating huge data related to GBV. And the event was followed by a sudden upsurge in research related to GBV using the generated data.

Thus, the data availability of social media has helped in many recent works related to GBV, which have delivered a multitude of interesting findings (Moitra, Ahmed & Chandra, 2021; Razi, 2020; Pandey et al., 2018; Khatua, Cambria & Khatua, 2018). Moreover, location-tagged social media data also assist in several cross-cultural studies related to GBV (Purohit et al., 2015; Starkey et al., 2019). In this article, we also used social media Twitter data from different countries of the world as a source of data.

Dataset and processing

We use public streams of Twitter data collected using the Twitter Streaming API. We procured 1% of public tweets provided by the API for a period of two years and five months (1st July 2016–25th Nov 2018). We remove all the duplicate tweets and retweets from the collected data as these do not add any new information (Cheke et al., 2020). From the collected tweets, we extracted GBV related tweets using a keyword matching approach as described next.

GBV Tweet extraction

UNFPA (United Nations Population Fund) domain experts have proposed three categories of GBV, namely sexual violence, physical violence, and harmful practices. They have also provided unique keywords related to each category of GBV, which have been used frequently in past literature for GBV related research (Purohit et al., 2015; ElSherief, Belding & Nguyen, 2017). Table 1 shows a total of 81 keywords constituting 29, 25, and 27 keywords from sexual violence, physical violence, and harmful practices respectively. We use the same keywords to extract relevant tweets from all three categories. The keyword set provided by UNFPA is very precise and can contain multi-words. Our methodology for extracting tweets for a particular keyword is based on the presence of the keyword in a tweet. If all the words of a multi-word keyword are present in a tweet regardless of order, we consider it a match. For example, for the category sexual violence, sexual assault is a related keyword with two words. If a tweet contains both the words sexual and assault, we consider it a match. For the cases where a tweet matches more than one category, we consider the tweet in both categories of GBV. This approach has been used in previous works in order to deliver high-precision data (Purohit et al., 2015). From the keywords related to each category of GBV, we extract tweets and create a tweet dataset from three categories, namely the sexual violence dataset, physical violence dataset, and harmful practices dataset, with a total of 0.83million, 0.53million, and 0.66million tweets, respectively. Further, we combined all three category tweets to create a GBV tweet dataset containing more than two million tweets.

Generic Tweet dataset

We created another dataset, the generic tweet dataset, to provide a better context of comparison with other categories of the dataset. This dataset is used for drawing inferences from GBV categories dataset w.r.t a generic dataset. For creating this dataset, we borrowed the methodology of ElSherief, Belding & Nguyen (2017). Our collected data is for a very long period, resulting in around 4 billion tweets. We extracted a random 1% sub-sample of total collected tweets as a generic tweet dataset. To eliminate duplicate content here as well, we removed tweets which are duplicates and retweets. Details of generic tweets data are given in Table 2.

Country-level location tagging

There are many indicators of location in a tweet, such as geotags, time zone, and profile location. Adopting the location indicators of Annamoradnejad et al. (2019) for tagging each tweet to a location, we use a three-level hierarchy of location indicative according to their accuracy levels (Kulshrestha et al., 2012). The first one is geotag, which gives the most accurate location information. If a geotag is available, then we use it for location tagging, and if it is not present, we look for the time zone data. Time zone is also an accurate way to tag country-level locations. A time zone data directly contains the user’s time zone in the form of the corresponding country name. For the cases where even time zone information is not available, then we look for the next location information in the hierarchy, i.e., location field mentioned in the user profile. Geotags and time zone contain exact country names, which can be directly mapped to a country. User profile location is an unstructured text location field that requires further processing to get country information. For this, we use the approach used by Dredze, Osborne & Kambadur (2016) where city names present in the user profile location are mapped to corresponding country names based on the Geoname (https://pypi.org/project/geonames/) world gazetteer. We borrow the list of required countries from Annamoradnejad et al. (2019) where authors have used a list of 22 countries, namely Arab countries, Argentina, Australia, Brazil, Canada, China, Colombia, France, Germany, India, Indonesia, Iran, Italy, Japan, Korea, Philippines, Russia, Spain, Thailand, Turkey, UK (United Kingdom), USA(United States of America). Arab Countries is a group of countries with a similar culture, so we merged tweets from all Arab Countries. There were very few tweets from Korea (170), so we discarded Korea from the list of considered countries and limited our research to the remaining 21 countries, each having more than 3,000 tweets. We apply the same location tagging scheme to all the GBV tweets and generic tweets. The complete data statistics are shown in Table 2 for all the categories of tweets.

We present the evaluation methodology and evaluation results of GBV tweets extraction and location tagging in ‘Results’. In this research, we want to explore the relationship between GBV online content and culture of a country. For this, we perform two analyses i. Correlation Analyses and ii. Graph Analyses. In correlation analysis, we correlate culture and its dimensions with different categories of online content in order to understand their relationship. In graph analyses, we create country graphs on the basis of parameters correlated in correlation analyses like content, and culture, which are compared using multiple graph comparison metrics in order to re-assure the observed relationships from correlation analyses. Next, we discuss the methodology used for these analyses.

Correlation analysis

In order to find a relation between the culture of a country and GBV, we calculate cultural distance and content distance between each pair of countries as detailed next.

Cultural distance

We quantify the cultural distance between two countries using cultural dimensions proposed by Geert Hofstede (Hofstede, Hofstede & Minkov, 2005). Geert Hofstede administered a huge survey among people from different countries to measure the difference in the way they behave. He has quantified six dimensions of culture (power distance¹, individualism², masculinity³, uncertainty avoidance⁴, long-term orientation⁵, indulgence⁶) for different countries in values ranging between 0–120. In order to measure cultural distance between two countries, we adopt the formulation of (Annamoradnejad et al., 2019) where authors use the euclidean distance between two countries to measure the cultural distance. The cultural distance can be formulated as shown in Eq. (1), where |D| is the total number of dimensions, $d_{c_1}^i$ and $d_{c_2}^i$ are the values of dimension dⁱ for countries c₁, c₂ respectively. (1)${\displaystyle CulturalDistance\left(C_1,C_2\right)=\sqrt{\sum _{i=1}^{\left|D\right|}{\left(d_{c_1}^i-d_{c_2}^i\right)}^2}}$

We also calculate the distance between countries on the basis of each dimension of culture proposed by Hofstede. For example, power distance is one of the dimensions of culture, and we need to calculate the distance between two countries according to power distance. For this also, we use euclidean distance, but since there is only one parameter, this becomes equivalent to |d_c1 − d_c2|. For further analyses, we calculate the cultural distance for each pair of countries on the basis of culture and six dimensions of culture.

Content distance

Online content related to a particular topic from a particular country captures country-level user comments and discussions on that topic (Cheke et al., 2020). In order to measure the difference between contents from two countries, we measure the content distance between two countries using Jaccard Similarity. First, all the tweets from each country are pre-processed to generate country-wise tweet tokens, details of which are given next.

Tweet preprocessing

We adopt the pre-processing settings of Cheke et al. (2020) to generate tweet tokens from each country. We first remove URLs, mentions, punctuation, extra spaces, stop words, and emoticons. Online acronyms and short forms are expanded using NetLingo (https://www.netlingo.com/acronyms.php). For hashtags, we removed the symbol # and kept the remaining word. Spelling and typos are corrected using Textblob (https://textblob.readthedocs.io/en/dev/). We also transliterated non-English words to English to reduce inconsistencies in language. Lastly, we tokenized each tweet using NLTK (Natural Language Toolkit). Extracted tokens from all the tweets of a country are merged to create country-wise tweet tokens. Next, for each pair of countries, we calculate content distance using the formula shown in Eq. (2), where C₁, C₂ are the set of all the tweet tokens of countries C₁ and C₂, respectively, and $\frac{|C_1\cap C_2|}{|C_1\cup C_2|}$ is the Jaccard Similarity.⁷ (2)${\displaystyle ContentDistance\left(C_1,C_2\right)=1-\frac{|C_1\cap C_2|}{|C_1\cup C_2|}.}$

We have a total of 5 categories of online content i.e., sexual violence, physical violence, harmful practices, GBV, and generic content. We apply the same methodology to extract country-wise tweet tokens from each category of online content.

Correlation

A correlation helps in understanding the relationship between two variables. Pearson correlation and Spearman correlation are two popular metrics for correlation. To establish robustness in our findings, we use both Pearson correlation, and Spearman correlation for calculating the association between content distance and cultural distance. Pearson correlation captures the linear relationship between two variables and Spearman correlation captures the monotonic relationship between two variables. Both the correlation metrics give correlation values in the range of (−1 to +1). A positive correlation value indicates that content similarity is higher for countries having higher cultural similarity and a negative correlation indicates vice versa. For calculating the correlation, we calculated content distance and corresponding cultural distance for each country pair (a total of ⁿC₂ pairs, if there are n countries). For exhaustive correlation analysis, we measured multiple correlations by keeping one correlation variable as different types of content (sexual, physical, harmful, GBV, and generic tweets) and another variable as six dimensions of culture.

P-value

For measuring the fitness of a correlation, we calculated the p-values for each correlation using the python library SciPy (https://docs.scipy.org/doc/scipy-0.14.0/reference/stats.html). A p-value represents the measure of occurrence of the correlation between two data samples by chance.

Graph analysis

Our objective is to compare GBV related content to a country’s culture. To this end, we created country graphs where edge weights are decided on the basis of different distances in terms of GBV content and culture, as mentioned in ‘Correlation Analysis’. For detailed analyses, we created multiple weighted graphs among countries with a different edge parameters. Finally, we compared created graphs using multiple graph distance metrics and graph clustering.

Country graph: A country graph created in this research is an un-directed, weighted graph G = (C, E), where C denotes the nodes of the graph, which are countries, and E denotes the set of edges between countries. For all the graphs in this research, the set of nodes C and the set of edges E are the same. The only difference is in the weights of the edges. Next, we describe the creation of edges in the required graphs.

GBV content graph: This graph captures the relationship between countries according to GBV content distance. In GBV content graph G_gbv = (C, E), the weights of the set of edges E are decided on the basis of the content distance score between two countries on the basis of GBV tweets. Here, GBV tweets are used for calculating content distance. We also create sexual violence graph, physical violence graph, and harmful practices graph where for assigning edge weights, we calculate content distance on sexual violence tweets, physical violence tweets, and harmful practices tweets, respectively.

Generic content Graph: This graph captures the relationship between countries and generic content. In the generic content graph G_rand = (C, E), the weights of the edges are assigned according to the content distance score between two countries on the basis of generic tweet data.

Cultural graph: This graph captures the cultural relationship between countries. In the cultural graph G_cult = (C, E), the weights of the edges are decided by the value of cultural distance calculated using Eq. (1).

Graph Pre-processing: For all the graphs G = (C, E), there is an edge between any pair of countries with a weight creating a complete graph. Further, all the created graphs have a different range of values of edge weight. For example, for GBV tweets graphs, edge weights will lie in the range (0,1), but for the culture graph, the values of weights can range from (0–120). To ensure consistency, we upscale edge weights in the range of (0,1) to a range of (0–120). Next, we pruned edges that are unimportant, i.e., whose weight is lesser than the median of all the edge weights. Thus, keeping only the important, i.e., higher edge weight edges in the graph. Before pre-processing, each graph is a complete graph with the same edges in all the graphs, but after pre-processing, each graph is a non-complete graph with only important edges resulting in different edges in each graph. The same is applied to all the graphs, and the final pre-processed graph is a weighted, un-directed, and non-complete graph.

We also mention that all the distances (content/culture) used to decide edge weight in all the graphs follow the axioms of distance (Kosub, 2019).

Distances:

• Quantum JSD: Quantum Jensen–Shannon Divergence (De Domenico & Biamonte, 2016) compares two weighted and undirected graphs by finding the distance between spectral entropy of density matrices.

• Degree divergence: This method (Hébert-Dufresne, Grochow & Allard, 2016) compares the degree distribution of two graphs. This methodology is applicable to weighted as well as unweighted graphs but only undirected graphs.

• Jaccard distance: Jaccard distance (Oggier & Datta, 2021) is applicable to only unweighted graphs, and its value depends on the number of common edges in the two compared graphs. For applying to our graphs, we coerced weighted graphs into unweighted ones by removing weights from all the graphs.

• Hamming distance: Hamming distance is one of the popular techniques for measuring the distance between two unweighted graphs. This is a measure of element-wise disagreement between the two adjacency matrices of the graphs. We applied Hamming distance to our graphs by coercing weighted graphs to unweighted ones by simply removing the weights.

• HammingIpsenMikhailov: This method is an enhanced version of Hamming Distance which takes into account the disagreement between adjacency matrices and associated laplacian matrices. This is applicable to weighted and undirected graphs.

• Frobenius: This is an adjacency matrix level distance that calculates L2-norms of the adjacency matrices.

• NetLSD: A metric for measuring graph distance based on spectral node signature distributions for unweighted graphs. For this, we coerced our graphs to unweighted ones by removing the weights.

GBV tweet extraction and error analysis

GBV tweet extraction is accomplished by tagging tweets using a keyword matching process. Following the keyword match verification methodology of Cheke et al. (2020), we employed three graduate annotators to manually tag the GBV category. Annotators were provided with a sample of tweets without any category information and were asked to manually tag each tweet to one or more categories of GBV (sexual violence, physical violence, harmful practices) with their own understanding and external online resources. Annotators were provided with a basic definition of GBV and its categories. For the purpose of validation, we created a balanced and shuffled sample of 6,000 tweets with 2,000 tweets from each category of GBV. For each tweet annotated by the three annotators, we select the majority category as the final category. Tweets that do not have any majority category are discarded. Considering the category tagged by annotators as the actual categories, we calculate the precision value of our keyword matching methodology for each category of GBV. The precision value for sexual violence is found to be 0.97, for physical violence 0.96, and for harmful practices to be 0.98.

Table 3 shows a few example tweets and tagged GBV categories. Examples 1–9 shows matching keywords and the tagged GBV category of the tweets from all three categories of GBV. Example 10–11 show tweets that contain keywords from more than a category of GBV. These tweets are kept in all the matching categories. In examples 12–13, keyword matching results in the wrong tagging of tweets because of contextual differences in tweets. As we can see in example tweet 12, the keywords woman and attacked belong to physical violence keywords, and hence the tweet is wrongly classified in the physical violence category. There are only a few such errors in GBV tweet category tagging arising because of changes in the context of tweets.

Evaluation of location tagging

We have a three-level hierarchy(time zone, geotags, profile location) of location tagging. Location tagging from time zone and geotags is completely accurate. For evaluating location tagging from profile location a random sample of 10, 000 tweets are given to three independent graduate annotators who were asked to manually tag a country-level location from their own understanding using online gazetteers and searches. The majority country name is selected as the final tagged country name. The profile location field with no majority among annotators is discarded. Considering the country tagged by manual annotation as the actual country, we obtained a precision score of 0.94.

Correlation analysis

Tables 4 and 5 show the results of pearson correlation and spearman correlation of different types of online content with culture and its parameters. From the tables, we can draw the following observations.

1. GBV content and all categories of GBV content show a positive correlation with culture and all its parameters by both the correlation metrics. For example, the correlation between GBV content and culture is 0.55, with a significant p-value of 0.001. Similarly, the correlation between culture and other categories of GBV i.e., sexual violence, physical violence, and harmful practices content is 0.51, 0.53, and 0.53, respectively with significant p-values.

2. The three parameters of culture uncertainty avoidance, indulgence, and individualism show comparatively higher correlation values as compared to other parameters of culture power distance, masculinity, and long-term orientation. This observation is consistent with all the categories of GBV content and with both the correlation metrics. The Pearson correlation of uncertainty avoidance, indulgence, and individualism with GBV tweet content is 0.36, 0.34, and 0.45 (Table 1). On the other hand, the pearson correlation of power distance, masculinity, and long-term orientation with GBV tweets content is 0.27, 0.16, and 0.17 respectively.

3. We also observe that the same content analyses performed for GBV content did not show a similar strong and consistent correlation with generic content. The Pearson correlation between generic content and culture is 0.33, which is much lower than the Pearson correlation between GBV content and culture, i.e., 0.55. Additionally, generic content fails to show any correlation with culture and its parameters from spearman correlation.

Observation 1 indicates that GBV content has an influence of culture and all six parameters of culture. The observation is consistent with all the categories of GBV content, i.e., sexual violence, physical violence, and harmful practices. Additionally, we also show the scatter plots of the culture and different GBV content types in Fig. 2 to reconfirm the findings. All these results imply that countries with similar culture also show higher similarity in GBV content. GBV content is composed of discussions, news, comments generated by users on the topic related to GBV. The reason for the similarity in GBV content for similar culture countries is the similarity in their discussions, news, and comments. For further understanding, we manually discerned the content of similar culture countries. According to Hofstede’s dimensions, the USA and Canada are more similar⁸ in culture as compared to USA and Iran. Similarly, Iran is more similar to Arab countries as compared to the USA. Table 6 shows the scores of cultural distance between different countries using Hofstede’s dimensions. In order to show the content differences of different culture countries, we ploted the word clouds of common frequent words of USA-Canada and Arab countries-Iran in Figs. 3A and 3B, respectively. For the countries USA and Canada, we found keywords like gfriend, whitesupremacist, objectifying as common frequent keywords. For the countries Arab countries and Iran, we found keywords like veiled, hijab, attacked, predator as common frequent keywords.

The highlights in observation 2 suggest that a few parameters of culture also play an important role in shaping the content related to GBV. Interestingly, Hofstede’s parameters uncertainty avoidance, indulgence, and individualism are found to show more impact on GBV related content than other parameters like power distance, masculinity, and long term orientation. For further reconfirming the connection between these parameters and different types of GBV content, we also show the scatter plots of these parameters and different content in Figs. 4, 5 and 6. From the scatter plots, we can evidently observe that Hofstede’s parameters uncertainty avoidance, indulgence, and individualism consistently show a close association (points are closer to the fitted line) with all types of GBV content. The same is not true for generic content. This shows a connection between these parameters of culture and different GBV categories content. The reasons for more influence of these parameters require further exploration which is outside the scope of this work. However, this observation again recommends a role of culture on GBV.

Observation 3 further strengthens the findings of Observations 1 and 2. The pattern of correlation showing a connection between culture and different categories of GBV is not the same for generic content. The lower and inconsistent correlation values of the generic content as compared to GBV content reinforce a stronger relationship between GBV content and the culture of a country. Further, the scatter plots shown in Figs. 2, 4, 5 and 6 also show that the points in all the plots of generic content are more scattered from the fitted line as compared to points in GBV and its category content plots. For example, the points in the GBV-culture plot (Fig. 2A) are closer to the fitted line, while in the generic-culture plot (Fig. 2E), the points are farther to the fitted lines showing a comparatively lower correlation. Other categories of content (sexual, physical, and harmful) also show a stronger correlation with culture as compared to generic content.

Generic content is composed of content from different topics, a few of which can be highly correlated to culture (Cheke et al., 2020), such as food, and a few can hardly show any correlation (Annamoradnejad et al., 2019), such as technology. These characteristics of generic content can be the most probable reason for showing weak correlations. Here we showed results of generic content just to provide a broader background for understanding. Next, we describe the results from graph analyses in order to validate findings from correlation analyses.

Graph analysis

We first summarize the statistical characteristics of the created graphs in our research in Table 7. All the graphs show similar basic properties because there is node correspondence in all the graphs. The edges and edge weights are the main varying parameter in the graphs. Next, we explain the details of graph comparison.

Table 8 shows the distance between different created graphs from various metrics. For each distance metric, if the distance value between a pair of graphs (G1, G2) is smaller than the distance between another pair of graphs (G3, G2), it means that the graph G2 is more similar to G1 as compared to G3. From the table, we observe that for all the metrics, the distance between the generic tweet graph and the culture graph is consistently higher than the distance between other graphs (GBV-culture, sexual violence-culture, physical violence-culture, and harmful practices-culture). For example, the metric QuantumJSD, the distance between generic graph and culture graph is 0.27 while for GBV graph and culture graph is 0.21. For the same metric, the distance between the sexual graph-culture graph, the physical graph-culture graph, and the harmful graph-culture graph is 0.21, 0.20, and 0.22, respectively. This shows that the graph created using content distance by GBV and its categories are more similar among themselves and to the graph created using cultural distance. The graph created using the generic content is consistently more distant from the cultural graph as compared to other graphs. This observation re-validates the observation from correlation analyses showing a higher degree of similarity between GBV content for similar culture countries. As two similar graphs show lesser distance by graph distance metrics, two similar graphs will also show similar clusters. Next, we present clustering in the created graphs.

Clustering the graphs

Figure 7 shows the plots of clusters in the culture graph, GBV graph, and generic content graph. From the created clusters, we observe that the clusters in the culture graph are more similar to GBV graph as compared to generic graph. The countries in the culture graph which belong to the same cluster have a larger overlap with the GBV graph rather than the generic graph. For example, the countries USA, UK, and Australia belong to the same cluster in the culture graph, and the same is also true for GBV graph. However, for the generic graph, all three countries belong to different clusters. This observation again shows a higher similarity between the culture graph and the GBV graph than the generic graph and the culture graph. Thus, we observe that the created clusters are also congruous to all other findings stating a higher level of relation between culture and GBV content, which is not the same for generic content.

Implications

In this research, we use social media data to verify connections between the culture of a country and GBV. Our findings suggest that real-world hypotheses are also evident in social media data, and their verification is no longer dependent on survey-based data. We believe that this research not only validates the hypothesis of confluence between culture and GBV but also points to the possibility of verification of other hypotheses related to GBV.

A finer analysis can also reveal culture-specific traits of GBV, which can further enhance understanding of GBV across cultures. We argue that these analyses are vital for designing culture-aware policies and strategies to curb GBV. There is a huge possibility of discovery of many more cultural norms like those pronounced by the World Bank (https://thedocs.worldbank.org/en/doc/656271571686555789-0090022019/original/ShiftingCulturalNormstoAddressGBV.pdf), which can promote GBV. Thus, this research paves a path for understanding culture-specific GBV using online social network data.

Limitations and critiques

In this section, we show a few possible limitations and how our research overcomes those. In this article, we have performed cross-cultural research using online content from Twitter. Here, we limit our research to English tweets only pertaining to two reasons. First, the GBV keywords are in English, resulting in a collection of English GBV content. Second, English has become the new lingua franca on Twitter (Choudhary et al., 2018), which delivers sufficient tweets for this research.

GBV data collection is based upon GBV keywords provided by UNFPA, which is a global organization. The provided keywords can be incomplete and non-exhaustive. There might have scope for increasing these keywords; however, GBV is a sensitive issue, and extending keywords without the intervention of social experts may introduce errors. So, we limit this research to globally available keywords only.

Online content is much inflected by a flux of ongoing news and events, which can lead to differences in data patterns in certain time periods. However, our research is based upon data from a long temporal span which diffuses such temporal inflections (Grieve, Nini & Guo, 2018).

There can be many more ways to capture the distance between countries in terms of GBV, but we have limited this to content distance using two common metrics (cosine similarity and jaccard similarity). The content distance used in this research captures the basic difference between tweet tokens of the two countries. However, the same methodology can be easily adapted to other twitter features and metrics.