Detection of news propagation barriers, being economical, cultural, political, time zonal, or geographical, is still an open research issue. We present an approach to barrier detection in news spreading by utilizing Wikipedia-concepts and metadata associated with each barrier. Solving this problem can not only convey the information about the coverage of an event but it can also show whether an event has been able to cross a speci c barrier or not. Experimental results on IPoNews dataset (dataset for information spreading over the news) reveals that simple classi cation models are able to detect barriers with high accuracy. We believe that our approach can serve to provide useful insights which pave the way for the future development of a system for predicting information spreading barriers over the news.
The phenomenon of event-centric news spreading due to globalization has been
exposed internationally [
Political Barrier. If news about a speci c event is disseminated from publishers having di erent political alignment, we can say that the news related to that event crossed the political barrier.
Geographical Barrier. We say that some news related to a speci c event overpasses geographical barriers if that event gets attention by publishers of countries located in di erent geographical regions.
Time Zone Barrier. We can claim that event-centric news has crossed the time zone barrier if it has been published by publishers located in di erent time zones.
Economic Barrier. It can be asserted that a piece of event-centric news has crossed economic barriers if it is published in countries having di erent economic conditions.
In this paper, we propose a methodology for detection of di erent barriers
during information propagation in form of news that utilize data (IPoNews) [
Following are the main scienti c contributions of this paper: { A novel methodology for barrier detection in news spreading. { Experimental comparison of several simple classi cation models that can serve as a baseline. 1.2
Observing the spreading of news on a particular event over time, we want to predict whether a barrier (cultural, political, geographical, time zone, economical) is likely to hamper information while information propagates over the news (binary classi cation). 2
Multiple barriers come across event-centric news speci cally when the news is
concerned about international or national events. According to news ow
theories, multiple determinants impact international news spreading. The economic
power of a country is one of the factors that in uence news spreading. Moreover,
economic variations has di erent in uence for di erent events (e.g. protests,
conicts, disasters) [
News agencies tend to follow the national context in which journalists
operate. One of the related examples is the SARS epidemic study which found
that cross-national contextual values such as political and economic situations
impact the news selection [
News classi cation for di erent kinds of problems is a well-known topic since
the past and features used to classify varies depending upon the problem. [
We utilized dataset "A dataset for information spreading over the news (IPoNews)"
that consists of pairs of news articles that were labeled based on the level of their
similarity, as described in [
The dataset contains a list of pairs of news articles annotated with one of the labels such as "information-Propagated", "Unsure", or "Information-NotPropagated" (see Table 2). The information is considered to be propagated if the cosine similarity score of the two articles in the pair is above a prede ned threshold ( 0.7 for Information-Propagated, < 0.4 for Information-not-Propagated, otherwise Unsure). We restructured the original dataset to include only examples labeled as spreading information. In this way, we have pair of news articles where we observe information spreading from one to the other. Furthermore, for each example, instead of having a pair of articles, we kept only the article that was published earlier. In this way, each example contains an article that spreads information. 3 http://wiki er.org/info.html, https://github.com/abdulsittar/IPoNews
The original dataset describes in Section 3 contains pairs of articles along with the information on whether there was the propagation of information related to a speci c event or not. We used only examples labeled as propagating information 4. Based on the available metadata for articles, we ignored articles that do not have metadata information in our database (see Section 3.4). Table 3 shows the statistics for each barrier after ltering the original dataset. As our dataset already mention (see Section 3) if information in news is spreading from an article to another based on Wikipedia-concepts, we utilized the most frequent (top 300) Wikipedia-concepts as features. Figure 1 portrays these Wikipedia-concepts for all three events in form of word clouds. Barriers knowledge refers to a database that contains metadata about each barrier. Figure 3 shows schema of database and Table 4 presents barriers along with their characteristics. Each barrier depends on one main information that is the country name of the headquarter of the news publishers. Since the utilized data 4 https://doi.org/10.5281/zenodo.3950064 set already contains headquarter of publishers therefore we fetched the country associated with headquarters. For economical barrier, we fetched economical pro le for each country using \"The Legatum Prosperity Index"" 5. Cultural di erences among di erent regions were collected using Hofstede's national culture dimensions (HNCD). For time zone and geographical barrier, we stored general UTC-o set, latitude, and longitude. For political barrier we are using the political alignment of the newspaper/magazine that we determined based on Wikipedia infobox at their Wikipedia page. For instance, for Austrian newspaper "Der Standard" we nd social liberalism as political alignment (See Figure 2), for British newspaper "Daily Mail" we nd right-wing as political alignment, for German "Stern" magazine there is no information in its Wikipedia infobox on the political alignment thus we label political alignment as unknown.
We represented each barrier with a speci c pro le containing a list of features. Table 4 depicts the list of features for each barrier. Economic and cultural barriers consist of a vector of length 11 and 6 features whereas geographical, time zone, and political only contain 1 or 2 features such as latitude-longitude, UTCo set, and political alignment. We queried the metadata information for each article and generated a CSV le for each barrier. We annotated each article based on that meta information to be used for model training and classi cation. For economic and cultural barriers, we calculated cosine similarity between vectors of economical values and vectors of cultural values. Score greater than the threshold value of 0.9 labeled as FALSE otherwise TRUE. We set the lowest value as a threshold based on the fact that if two countries have a little gap concerning culture or economical values then there exists a barrier. For geographical barriers, we compared the latitude and longitude of the country of each publisher. If a country name or lat/lat appeared to be the same then we annotated it with FALSE otherwise TRUE. Lastly, for Barrier Features
Rank, Safety-Security, Economic PEenrtseornparils-Fe-rCeeodnodmit,ioGnso,veMrnaarknecte-,InSforcaisatlr-uCcatpuirtea,l,EIcnovneosmtmice-nQt-uEanlivtyir,onment, Living-Conditions, Health, Education, Natural-Environment
Power-Distance, Cultural Uncertainty-Avoidance-By-Individuals, Individualistic-Cultures,
Masculinity-Femininity, Long-Term-Orientation, Indulgence-Restraint Geographical Latitude, Longitude Time Zone UTC-o set Political Political-Alignment time-zone and political barriers, we followed the same process that was for the geographical barrier. if political alignment or UTC-o set appeared to be the same for a pair then it is annotated with FALSE otherwise TRUE. Figure 4 depicts the class distribution for each barrier. We can notice unbalanced class distribution with majority of the examples being False. This is especially true for Cultural and Political barrier with 91 percent of example being False. Thus in our evaluation we rely more on F1 measure than classi cation accuracy. 4.1
For each barrier, we have a list of news articles where each article is associated with 300 Wikipedia-concepts and features related to that barrier. The task is to predict the status S of each barrier B.
S = f (C; B) f is the learning function for barrier detection, C is donating here Wikipediaconcepts related to an article and B is the list of features related to a speci c barrier (see Table 4). 4.2
We utilized dataset IPoNews [
We used the following methods as baselines for all our models. { Uniform: Generates predictions uniformly at random. { Strati ed: Generates predictions by respecting the training set's class distribution. { Most Frequent: Always predicts the most frequent label in the training set. 5.2
We trained popular classi cation models for each barrier such as SVM, kNN, Decision Tree, Random Forest, and Naive Bayes using Scikit-Learn. We applied a strati ed 10-fold cross-validator to split the dataset for training and testing. For Random Forest, kNN, and Decision Tree, we varied the size of n-estimator, value of k, and max-leafs and chosen the one with the best score on test data respectively. Implementation of this methodology to barrier detection can be found on GitHub 6. 5.3
Due to imbalance in the class distribution for all barriers, we used micro averaged precision and recall to evaluate our models. 7 { Micro-Precision: The precision of average contributions from each class is calculated in micro-precision whereas the following question is answered by precision: What proportion of positive predictions was correct? It is de ned as:
{ Micro-Recall: Recall of average contributions from each class is calculated in micro-recall whereas the following question is answered by recall: What proportion of actual positives was predicted correctly? It is de ned as:
5.4
Table 5 shows the results of all the classi ers for each barrier along with baselines. Analysis of the experimental results show that overall all the machine learning models outperform the three baselines. For all the barriers, we can notice MicroRecall is equal to Micro-Precision. The best performing baseline is the "Mostfrequent" with Micro-F1 for economic, cultural, geographical, time zone, and political barrier equal to 0.70, 0.90, 0.58, 0.70, and 0.90 respectively. The best performing models on all the barriers are Decision Tree, Random Forest, and kNN. Looking at Micro-F1, we can see that on the Economic and Cultural barrier kNN achieved the best performance of 0.75 and 0.95 respectively. On Geographical barriers, kNN and Decision Tree performed the best achieving 0.81. On Time-Zone, the best performing classi er is Random Forest with Micro-F1 6 https://github.com/cleopatra-itn/BarrierDetection-Classi cation 7 https://peltarion.com/knowledge-center/documentation/evaluationview/classi cation-loss-metrics/micro-recall 0.83. On Political barriers, SVM, kNN, and Random Forest achieve the best Micro-F1 score of 0.97.
In terms of classi cation accuracy, we can see that Random Forest outperforms the baselines as well as the other four classi ers for the rst four barriers. Notice that Random forest performs better than decision tree but takes more time. Naive-Bayes achieves a little bit lower classi cation accuracy than the Decision Tree for the rst four barriers. On the political barrier Naive-Bayes achieves the best classi cation accuracy (0.98) but lower Micro-F1 (0.66). 6
It is highly important to detect the barriers while information propagates
specifically through the news. For journalists, marketers, and social scientists, the
phenomenon of knowing which barrier appeared most frequently for what type of
events, is signi cantly helpful to solve business and marketing problems. In this
regard, we proposed a simple methodology. Though its results are good enough
for three types of events, we would like to enhance features as well as events. We
used only Wikipedia-concepts and meta information to detect barriers. In the
future, we would like to use DMoz categories provided by Event Registry [
The research described in this paper was supported by the Slovenian research agency under the project J2-1736 Causalify and co- nanced by the Republic of Slovenia and the European Union's Horizon 2020 research and innovation program under the Marie Sklodowska-Curie grant agreement No 812997.