https://orcid.org/0000-0002-8842-163X
Figures
Abstract
Language is both a cause and a consequence of the social processes that lead to conflict or peace. “Hate speech” can mobilize violence and destruction. What are the characteristics of “peace speech” that reflect and support the social processes that maintain peace? This study used existing peace indices, machine learning, and on-line, news media sources to identify the words most associated with lower-peace versus higher-peace countries. As each peace index measures different social properties, they can have different values for the same country. There is however greater consensus with these indices for the countries that are at the extremes of lower-peace and higher-peace. Therefore, a data driven approach was used to find the words most important in distinguishing lower-peace and higher-peace countries. Rather than assuming a theoretical framework that predicts which words are more likely in lower-peace and higher-peace countries, and then searching for those words in news media, in this study, natural language processing and machine learning were used to identify the words that most accurately classified a country as lower-peace or higher-peace. Once the machine learning model was trained on the word frequencies from the extreme lower-peace and higher-peace countries, that model was also used to compute a quantitative peace index for these and other intermediate-peace countries. The model successfully yielded a quantitative peace index for intermediate-peace countries that was in between that of the lower-peace and higher-peace, even though they were not in the training set. This study demonstrates how natural language processing and machine learning can help to generate new quantitative measures of social systems, which in this study, were linguistic differences resulting in a quantitative index of peace for countries at different levels of peacefulness.
Citation: Liebovitch LS, Powers W, Shi L, Chen-Carrel A, Loustaunau P, Coleman PT (2023) Word differences in news media of lower and higher peace countries revealed by natural language processing and machine learning. PLoS ONE 18(11): e0292604. https://doi.org/10.1371/journal.pone.0292604
Editor: Mihajlo Jakovljevic, Hosei University: Hosei Daigaku, JAPAN
Received: May 30, 2023; Accepted: September 24, 2023; Published: November 1, 2023
Copyright: © 2023 Liebovitch et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The programs used to analyze that data in the current study are available at: https://github.com/mbmackenzie/power-of-peace-speech https://github.com/wpqc21/ArticleClassifier/tree/main/ArticleClassifierHSSC https://github.com/smilelinnn/Article-Classification The data analyzed in the current study are available at: Liebovitch, Larry et al. (Forthcoming 2023). Words in news media in low and high peace countries [Dataset]. Dryad. https://doi.org/10.5061/dryad.2v6wwpzv6.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Importance of language
Communication through language has been highlighted as the single most important process in constructing our reality [1, 2]. Language also plays a critical role in conflicts. The extreme power of “hate speech” to mobilize destruction and violence is evident around the globe. In Kenya, hate speech over social media and in blogs played a central role in inciting ethnic divides and conflict [3]. In Nigeria, hate speech in the news was identified as a major driver of election violence [4]. Studies in Poland have shown that exposure to hate speech leads to lower evaluations of victims, greater distancing, and more outgroup prejudice [5]. Peacekeepers working in conflict zones are currently using data science and natural language processing methods to track hate speech–monitoring hostile news accounts, blogs, and broadcast and social media posts in order to provide early warning predictions of increases in ethnic tensions or violence in local communities [6]. These studies have focused on the prevention of destructive conflicts, approaching peace as the absence of harmful conflict.
However, highly peaceful societies have been found to evidence other conditions and processes in addition to an absence of violence that distinguish them from low peace nations, including the prevalence of non-warring norms, values and rituals [7]. Highly peaceful societies are also significantly more stable and have the lowest probability of lapsing into violence [8]. Yet, scarce research has been devoted to unpacking the conditions promoting higher levels of sustainable peace [9, 10]. To build a foundation for sustainably peaceful societies, it is imperative to understand the drivers of peace. This has led to an increasing number of studies of “positive peace” [11–16] to understand the active social forces that work together to generate and maintain peace in a society.
Linguistic features of peace and conflict can be found in all aspects of language including in phonology, grammar, semantics, pragmatics, and discourse [17]. This leads us to ask, what are the properties of “peace speech” that are the other face of the coin from “hate speech”? Peace speech is a basic linguistic structure that may help to build and sustain peacefulness between people and between groups [18–21]. There is only limited empirical evidence identifying the specific features and effects of peace speech [2]. As noted by peace linguist Patricia Friedrich [19], “Just how much a change in vocabulary can shape the outcome of interactions should be a matter to be empirically verified by peace linguistics as soon as possible, so we can all move from the realm of possibility to the realm of empirical evidence and corroboration” (p. 120). Our aim here, is to use machine learning to identify some of the linguistic features of peace speech, namely the most frequently used words in lower-peace and higher-peace countries.
Measuring peace
There are several approaches to measuring peace. In some measures, peace is viewed as an objective state that can be defined, quantified, and measured according to a standardized set of parameters. This “technocratic” [22] approach assumes that peace consists of criteria that do not vary from case to case, and seeks to compare and rank cases, often in order to drive policy and funding. Each of these indices consider a wide array of indicators that capture discrete elements of peacefulness and rank countries on their performance or attainment of these elements [23]. We used these peace studies to train our machine learning models:
- The Global Peace Index (GPI) which measures peacefulness and its economic value [24]. It uses 23 indicators of the absence or fear of violence, each with a different weight factor from 2.0 to 5.0, that cover ongoing domestic and international conflict, societal safety and security, and militarization. Examples of these measures include: number of deaths from internal organized conflict, political instability, and number of armed services personnel per 100,000 people. It is published by the Institute for Economics and Peace (IEP), a non-profit company founded by IT entrepreneur and philanthropist Steve Killelea.
- The Positive Peace Index (PPI), which measures the conditions for peace in a society to flourish [25, 26]. It uses 24 indicators of the absence or fear of violence over 8 categories that include: acceptance of the rights of others, free flow of information, low levels of corruption, and well functioning government. Examples of these measures include: internet use over the last three months, perceptions of how often public sector employees steal, embezzle, or misappropriate public funds, and perceptions of the quality of public services. It is also published by IEP.
- The Human Development Index (HDI), which measures a long and healthy life and a decent standard of living [27]. It measures a country’s achievements towards "goalposts" in health, education, and standard of living set by the United Nations. Examples of these measures include: life expectancy at birth, mean years of schooling, and the Gross National Income (GNI) per capita. It is published by the United Nations Development Program.
- The World Happiness Index (WHI), which measures happiness as perceived by people themselves and their community [28]. It measures how perceptions of life satisfaction measured by polling correlate with Gross Domestic Product (GDP), life expectancy, generosity, social support, freedom, and corruption. It is published by the Sustainable Development Solutions Network, powered by the Gallup World Poll data.
- The Fragile States Index (FSI), which measures fragility, risk and vulnerability [29]. It measures qualitative and quantitative data in 4 areas: cohesion, economic, political, and social. Examples of the indicators measured are group grievance, economic decline, human rights and rule of law, and refugees and internally displaced persons. It is published by the Fund for Peace.
We chose these indices as they are the most often used measures of peace and are highly respected in the peace studies community. They represent a broad set of measures on the rights, stability, and quality of life characteristically found in peaceful societies and lacking in non-peaceful societies, as measured by government, non-profit, and private sources. The limitations of these indices include that they: a) are often based on incomplete data sets, b) are averaged across very different areas within nations to achieve national averages, c) based on vastly different assumptions and conceptualizations of what constitutes peacefulness, and d) are based on linear assumptions of cause and effect.
Alternative approaches hold that peace is highly context specific [22]. This local-centric approach centers on those who live within the context being measured, proposing that it is the members who live within a given society who should define what constitutes peace. This approach is more participatory and seeks to include people from a given context to identify, define, and weight indicators of peace. Examples of this approach include the Everyday Peace Indicators [30] and Generations for Peace [31]. These methods help to address some of the limitations of top-down approaches to measuring peacefulness.
Goals of this study
Our primary goal here was to identify the words and their frequency of use in media articles that are most important in differentiating lower-peace and higher-peace countries. Certainly, words alone do not capture all the linguistic subtleties of language, but they can serve as a good starting point to explore the linguistic differences between lower-peace and higher-peace cultures. These words are the conduits of the social processes that underlie conflict and peace and may therefore provide insights into identifying those social processes and so have useful applications in conflict prevention and peace building.
Having developed our machine learning model to analyze media articles that accurately classified countries as lower-peace and higher-peace in this study, we then used that same machine learning model to provide a quantitative peace index, not only for the lower-peace and higher-peace countries in the training set, but also for other intermediate-peace countries that were not in the training set.
Methods
Overview
Over the previous centuries science has proceeded by using observations, experiments, data, and intuitions, to form theoretical frameworks that could then be supported or falsified by further experimental data. That is a top-down approach, from thoughts (theory) to data. In these studies, we went in the reverse direction, from the bottom-up, from data to thoughts (results) [32–34].
We used a data-driven, machine learning framework. This is different from a traditional theory-driven framework. We believe that both theory-driven and data-driven frameworks can give useful insights in peace studies. In this data-driven approach we create a machine learning "model", a set of equations. That model has inputs and outputs. We have no "theory" about how the outputs depend on the inputs. Instead, we use known examples from our data to "train" the model. Using those examples, we adjust the model, so that it gives good outputs from those inputs. Then, we can use that model: 1) to tell us which inputs were most important in determining the outputs and 2) to provide good output predictions from new inputs. In this paper, the inputs were the word frequencies from the news media in each country and the outputs were the level of peace in that country. The machine learning model: 1) reports which words were most important in differentiating lower-peace from higher-peace countries and 2) provides a quantitative machine learning peace index output computed from the input of the frequency of words from news media sources.
Using the data-driven, machine learning theoretical framework, we made no a priori assumptions or hypotheses about which topics or which words would be the most important in finding the differences in news media between lower-peace and higher-peace countries. We used all the words in the NOW (News on the Web) corpus because it has a large amount of news media data on a large range of different topics.
An overview of the strategy used is shown in Fig 1. First, to focus on the differences between lower and higher peace countries, words likely to be common in both were removed by natural language processing. Also removed are names of people, places, and companies that would be confounding variables to predict the level of peace not related to language itself. Then the machine learning method is “trained” on countries of different levels of peace, that is, it is given some word data from those countries and then the parameters of the machine learning model are adjusted so that input into the model yields the correct output classification. Another set of “hyperparameters” on the how the machine learning algorithm works can also be adjusted, but those were held constant in the work presented here. This model is then “tested” by determining the statistical accuracy of its predictions given new word data from countries of different levels of peace. Next machine learning importance methods are used to find the words that are most important in the machine learning model in making its classification. This step determines the words that are most significant in differentiating lower and higher peace countries. This strategy to find the most important features in predicting the correct classification is typical of many applications of machine learning in data science and has been reliably used in other natural language applications, classifications based on features with numerical values, and image analysis [35, 36].
One machine learning model was trained on only lower-peace and higher-peace countries. It was then used to generate a quantitative machine learning peace index of any country, which could be in either the lower, intermediate, or higher-peace regimes. Therefore, it was applied to countries that are both similar and different to the countries that it was trained on.
Data collection and pre-processing
We used the 723,574 media articles with 57,819,434 words published between January 2010 and September 2020, from the News on the Web corpus [37]. The NOW corpus consists of news media from the 20 countries that have substantial local sources in English. It is one of the largest corpus of words in English. We used only local sources in English as we did not have confidence in automated systems to correctly translate and give correct local contextural meanings from other languages in a larger set of countries.
We used all the data from all the many different media sources and all the many different types of articles from all the countries in the NOW corpus. This data consists of on-line newspaper and magazine articles about accidents, business, crime, education, the arts, government, healthcare, law, literature, medicine, politics, real estate, religion, sports, war, as well as book, music, and movie reviews, and could include any article downloaded from its media sources. A sample of these sources include: AlterNet, Austin American-Statesman, Business Insider, Business Wire (press release), Chicago Tribune, FOX43.com, Jerusalem Post, Israel News, KCCI Des Moines, Kentwired, KOKI FOX 23, POWER magazine, Press of Atlantic City, The Jewish Press, USA TODAY, and Vulture.
This data is broadly representative of the language used in news media in that it includes a wide number and variety of media sources. It is biased in that all the sources are in English, which means it is best representative of countries where English is the native language, but less representative of countries where English is not the native language. Fairness is more difficult to ascertain, it is as fair, or as not fair, and the sources of its data. Since it has data from a wide variety of sources, it does include a wide variety of viewpoints.
We used the following natural language pre-processing steps. All the stop words (common words), named entities (proper nouns such as names of people, places, and companies) and phrases unrelated to the article’s content (such as ads) were removed by the methods of Jung et al. [38] and manually as necessary. Removing all such names risks losing important information as those names can be used as markers of the ideologies associated with them. On the other hand, because machine learning models can use such low-frequency identifiers to make classifications, as is standard in natural language processing applications, we removed those names. Two countries from this corpus (Pakistan and South Africa) did not have a sufficient number of articles in English for further processing and were omitted from the machine learning model. We then recorded the 300 most frequent words and their frequency of occurrence among all the articles combined for each country, which resulted in 767 unique words across all the 18 countries shown in Table 1.
We chose to use the 300 most frequent words, as the word frequency has declined significantly by that rank. Typically, known as Zipf’s Law, word frequencies are inversely proportional to word rank, so that the frequency of the 300th word would be approximately 1/300 of the most common word, or 0.0033. We did not explore other values for this parameter, but the word frequencies are already so low at the 300th word, that we would not expect that adding additional words would significantly alter the result.
Machine learning models
Our strategy was to identify the words that are most important in a machine learning model in making the classification of the level of peace. To train a machine learning model required a training set of data with a known measure of peace in each country. However, as each existing peace index measures different social properties, there is no detailed agreement in the numerical values of these indices for each country. However, we can successfully use these indices, as described below, to group countries into three overall classes of lower-peace (class 0), higher-peace countries (class 1), and intermediate-peace (class 2), which we then used in our first 3-class machine learning model. As described in the results section below, that 3-class model was not very good at predicting the level of peace in a country. As there is more consensus amongst these indices for the countries that are at the extremes of lower-peace and higher-peace, we also developed a second, independent 2-class machine learning model using only the lower-peace (class 0) and higher-peace (class 1) countries. Using such extreme cases can help to clarify the differences between them. For example, Voukelatou et al. [39] compared three peaceful countries (Portugal, Iceland, and New Zealand) with three of the most war-torn countries (DR Congo, Pakistan, and Yemen). This is called the Extreme Groups Approach (EGA) in psychology, where it must be used cautiously as to not artificially inflate statistical accuracy [40]. It is however appropriate and useful in standard machine learning to predict group membership, here to predict whether a country is lower-peace or higher-peace.
To determine the lower-peace and higher-peace countries, we first found the average values, over the years 2010–2019, for the GPI, PPI, WHI, FSI, and HDI indices for each country, as shown in Table 2. These indices were chosen as they are among the more prominent measures of levels of peace, conflict, and well-being at the country level.
GPI = Global Peace Index, PPI = Positive Peace Index, WHI = World Happiness Index, FSI = Fragile States Index, and HDI = Human Development Index.
Each index also uses a different range from lower-peace to higher-peace to measure overall peace, respectively from 5–1, 5–1, 0–10, 120–0, and 0–1. To more easily compare them, as shown in Table 3, we linearly scaled the average of each index over the range 0–100, where 0 is lowest-peace and 100 is highest peace for the countries we analyzed.
Color coded independently for each index, lower-peace group = red, higher-peace group = green, and intermediate-peace group = yellow.
Each of the 5 peace indices (GPI, PPI, WHI, FSI, HDI) we used has its own unique theoretical framework [24–29]. We sought to use the best consensus of these indices to determine the lower-peace and higher-peace countries. For each index, we then ordered each country by its average value and divided that list into thirds. The lower-peace countries were then defined as those with 3 or more of the 5 indices in the lowest group in that index, the higher-peace countries as those with 3 or more indices in the highest group in that index, and the intermediate-peace countries as those not in either group. We also compared this result to the average of the 5 indices. We used the unweighted average as there is no clear criterion on how to weight each index. Both methods gave the same results for the choices of countries in the lower-peace and higher-peace classes. The lower-peace countries chosen by the first method all had the lowest average of the 5 indices and the higher-peace countries chosen by the first method all had the highest average of the 5 indices.
The NOW corpus consists of data from 20 countries chosen by the criteria of having substantial local on-line news in English. We initially made this assignment for all those 20 countries and kept that same assignment for the entire analysis, when later we removed two countries because they had a much smaller number of articles.
Table 3 also shows how each of the five peace indices rates each country, the lower-peace group countries (in red): Bangladesh, Kenya, Nigeria, and Tanzania; the intermediate-peace group countries (in black): Ghana, Hong Kong, India, Jamaica, Malaysia, Philippines, Sri Lanka, and the United States; and the higher-peace group countries (in green): Australia, Canada, Ireland, New Zealand, Singapore, and the United Kingdom.
Table 4 shows the number of articles and words in the data from the countries used for the 3-class model of the lower-peace, intermediate-peace, and higher-peace countries.
Table 5 shows that data for the 2-class model of the lower-peace and higher-peace countries.
We used the random forest and logistic regression classifiers [41, 42] to train and test the 3-class model of lower-peace (class 0), higher-peace (class 1), and intermediate-peace (class 2) and independently the 2-class model of lower-peace (class 0) and higher-peace (class 1). In all cases the features were the normalized frequency of the 767 most frequently used words across all the countries in the data. There are different ways to both train and test such models [36]. As typically done, we first trained each model by using 80% of the data and then tested it on the remaining 20% of the data. We also used a leave-one-out cross-validation method [43] where the model is trained on all but one country, tested on the excluded country, and this is repeated for each different country being excluded. This makes more efficient use of the information in the data but requires additional computational time for the repeated trainings.
Machine learning peace index
As shown in Fig 1, the data-driven, machine learning approach leads from the data through natural language processing, to the training and testing of the machine learning model. We then used the machine learning model in two different ways. First, as already described, we used its importance methods to identify the word differences that best classify (that is, predict) which countries are lower or higher peace. Second, we also used it to provide a quantitative peace index from the media data.
From the 2-class model, trained on only the word frequencies in the lower-peace and higher-peace countries, these equations determined the probability, p, that a country is in the higher-peace class (class 1). In the binary classification task, the model classifies countries as lower peace if p < 0.5 and higher peace if p ≥ 0.5. We used this value of p, which is a quantitative measure of the probability of being in the higher-peace class (class 1), as a measure of the level of peace in any country in either the lower, intermediate, or higher-peace regimes. Therefore, it was applied to countries that are both similar and different to the countries that it was trained on. To be consistent with our scaling of the other peace indices in Table 3, we defined a machine learning peace index as 100 x p, so that 0 is lowest peace and 100 is highest peace.
Results
Performance measures
Table 6 shows the mean ± sem of the performance measures for random guessing and the random forest and logistic regression classifiers on the 3-class and 2-class models. These are the Accuracy = (TP+TN)/(FP+FN+TP+TN), the Precision = TP/(TP+FP), the Recall = TP/(FN+TP), and F1 = 2(Precision x Recall)/(Precision + Recall), where TP = true positive, TN = true negative, FP = false positive, and FN = false negative.
3-class models are: lower-peace, intermediate-peace, and higher-peace. 2-class models are: lower-peace and higher-peace.
Unlike a plausibility probe (in social sciences) where hypotheses are first tested for potential validity, in the data-driven, machine learning approach used here, the performance measures (shown in Table 6) quantify the ability of the machine learning model to successfully represent the data.
First we consider the results from the 3-class model using the lower-peace, intermediate-peace, and higher-peace countries. Random guessing of the three classes, averaged over 20 guesses, yielded an accuracy of 0.356, within the error expected around 0.333. The 80/20 train/test split using all 18 countries, was only weakly more successful, with accuracies of 0.525 for the random forest and 0.388 for logistic regression. Nonetheless, for example, for the random forest, this accuracy is still statistically significantly greater than random guessing (Z = (Δmeans)/sem = 4.14, p < 2.0 x 10−5, one-tailed).
We had expected that the more efficient leave-one-out cross-validation using all the data from 17 countries to predict the class of the one country not used in each training, would significantly improve the accuracy. This was not the case. For random forest the accuracy only improved slightly from 0.525 to 0.567 and for logistic regression the accuracy increased a little more from 0.388 to 0.611.
Second, we now consider the results from the 2-class model using only the lower-peace and higher-peace countries. Now the leave-one-out cross-validation model dramatically increased the accuracy to 0.960 for the random forest model and 1.000 for the logistic regression model. (The 80/20 split of the 2-class model is not shown in Table 6, as there are only 2 or 3 predicted values for each run, which is insufficient to compute the performance measures. However, the 10 predictions of the leave-one-out cross-validation models are statistically significant with an average over 20 runs of 9.6 our of 10 correct classifications for random forest and 10 out of 10 correct classifications for logistic regression, and the probability that this is due to chance is p = 0.00977, Binomial(n = 10,k = 1,p = 0.5) for the random forest model and p = 0.00098 Binomial(n = 10,k = 0,p = 0.5) for the logistic regression model.) This dramatic improvement in the accuracy of the 2-class model over the 3-class model makes sense in the following way. Since the peace indices of the humans disagree with each other how can our machine learning model figure it out? These 5 peace indices have a large range of values for each country. Averaged over all the countries, the range of the normalized (0–100) peace indices for each country is 30.68, with a minimum range of 4.76, a maximum range of 64.72, and a standard deviation of 19.06. This range of values for each country is illustrated in Fig 2. But, the peace indices of the humans are more aligned with each other for the countries that are extremely low or high peace, so they are pretty sure which are the most lower-peace and higher-peace countries. This is quantified by the standard deviations of the peace indices in each class: 8.14 for the higher-peace countries, 17.57 for the lower-peace countries, and 24.12 for the intermediate-peace countries.
Overlapping values are color coded by the center and surround of the markers and the black lines are the range of those values for each country.
Thus, training on the 2-class model using only the lower-peace and higher-peace countries makes it possible for the machine learning model to properly associate the level of peace with the word frequencies.
The number of articles and words from each country is shown in Table 5. The excellent values for the performance measures in Table 6 for the 2-class model demonstrate that the differences in numbers of articles or words between the countries had no significant negative impact on these results. For example, the 20 random forest runs for the 2-class model there were no mis-classifications for 188 classifications, only 5 times were higher-peace countries mis-classified as lower-peace, and only 7 times were lower-peace countries mis-classified as higher-peace. The 20 logistic regression runs all converged to the same values with no mis-classifications.
NOTE: Because the performance measures of the 2-class model were so much better than that of the 3-class model, all the following results are based on the 2-class model.
Most frequent and important words in lower-peace and higher-peace countries
The machine learning model was used to find the words most important in differentiating lower-peace and higher-peace countries. The machine learning model is not making its classification of lower-peace and higher-peace based on a few occurrences of a few specific words from a few media sources. It is making its classification based on the frequencies of all the most common 767 words, across all the media sources within that country, and across all types of articles. As shown in shown in Fig 3, the frequency of the most common 100 words, range from approximately 28,000 to 410,000 occurrences in the data set. Fig 3 shows the words that were most frequently used in the articles in the lower-peace and higher-peace countries. We also used the feature_importances method from the random forest classifier to determine which of these words were most important in correctly predicting whether a country is lower-peace or higher-peace. The highest frequency words were more likely the words of the highest feature importance, but interestingly, many words of lower frequency were also important in predicting whether a country was lower-peace or higher-peace. Fig 3 shows the 100 most frequently used words in higher-peace and lower-peace countries with the words of highest feature importance highlighted in yellow.
Yellow indicates the words of highest feature importance in making the higher-peace and lower-peace classification by the random forest feature importance method. The word frequencies, of words reported within the countries, also shown in the figure, range from approximately 28,000 to 410,000 occurrences.
A decision tree graph is a good method to classify data, but it can learn to fit the training data so tightly that it does not generalize to properly classify new data. The random forest classifier was developed to avoid such overfitting by creating many different classification trees from random subsets of the data, literally a random forest, to make the classification. Hence, each time it is run, the model will identify a slightly different set of words that best classifies a country as lower-peace and higher-peace. We found that the words of high frequency and high importance were very similar in each run, but that there was more variation in the words identified of lower-frequency and lower-importance.
Figs 4 and 5 show word clouds of the words of highest feature importance, with the size of those words scaled to their frequency of occurrence, in green for higher-peace countries and in red for lower-peace countries. The word cloud in Fig 4 for the higher-peace countries has the most words of high feature importance and high frequency of occurrence (36/100) associated with daily activities: “time” “like”, “game”, “play”, “good”, “team”. On the other hand, Fig 5 for the lower-peace countries, has the most words of high feature importance and high frequency of occurrence (41/100) about social structures, such as: “state”, “government”, “country”, “court”, “general”, “law”. A preliminary and speculative analysis of these results suggests that lower-peace countries are characterized by words of government control and fear. The direction of the arrow of causality is not clear. Do the social realities lead to these words, or do these words lead to the social realities? Can “peace speech” in news and social media enhance the prospects for peace or only reflect it?
Machine learning peace index
We used logistic regression to determine a quantitative peace index, not only for the lower-peace and higher-peace countries in the training set, but also for other intermediate-peace countries that were not in the training set. Logistic regression, trained on the set of independent variables and their classes, can also be used to predict the probability of a class, given new values of the independent variables. The logistic regression model used the word frequencies from each country to compute the probability p that country was higher-peace (class 1). This model was retrained on all the 10 countries and was not an average of all the 10 models using the leave-one-out cross-validation. The machine learning peace index was then equal to 100 times p. Table 7 shows the 18 countries in rank order of this machine learning peace index compared to the GPI, PPI, WHI, FSI, and HDI indices. There are three important findings from this computation.
Training set of countries: lower-peace = red and higher-peace = green.
First, as can be seen in Table 7, although the model was trained only on the lower-peace and higher-peace countries, and it has never been given any data whatsoever about the intermediate-peace countries, it correctly ranks those intermediate-peace countries in between the lowest lower-peace and highest higher-peace countries. This important result confirms that the machine learning model has learned something real and substantive from the word frequencies in the lower-peace and higher-peace countries, that correctly generalizes to the intermediate-peace countries.
Second, unlike the positivist approaches to measuring peace, which a priori choose social indicators from their conceptualization of peace, the machine learning peace index is data driven and free of any assumptions about which words or their frequencies are most representative of peace. The choices of which words, and their frequencies, are important in measuring peace, arise solely from training the machine model, with samples of media articles from countries identified as lower-peace and higher-peace. This is a new and valuable data driven, bottom-up approach. It is the reverse of a classical top-down approach where a conceptual framework is used to hypothesize which words best measure peace and then test that hypothesis with data. As previous measures of peace had used a top-down approach based on a priori assumptions, here we explored what could additional be learned by using a bottom-up, data driven, data science approach. We chose to explore that bottom-up approach because it could, and in fact did here, provide us new insights into the differences in language between lower and higher peace countries that had never before been formulated into hypotheses to be tested. Every flower of a different color adds beauty to the garden.
Third, Table 7 also shows that the machine learning peace index for each country is similar to the average of the 5 peace indices of that country, linear regression r2 = 0.8349. This is true even though the machine learning peace index is based on the frequency of words in news media and the other peace indices are each based on different theoretical frameworks and measurements. Our machine learning peace index correlates slightly higher with the PPI (r2 = 0.8628), FSI (r2 = 0.8581), WHI (r2 = 0.8378), and HDI (r2 = 0.8007) than the GPI (r2 = 0.3308). It appears that our machine learning peace index is capturing essential aspects of peace from the frequency of words alone in news media that aligns with the entirely different and more complex measures based on national data, economic statistics, and polling data, used by these traditional peace indices.
Discussion
The language that we use to communicate across our differences both reflects our internal view of the world and influences our external world. “Hate speech” can mobilize violence and destruction. Much less is known about “peace speech” that characterizes peaceful cultures and that may also help to generate or sustain peace. Our long-range aim is to identify the linguistic features of speech that characterizes lower-peace and higher-peace societies. In this study, we identified the words in media articles most associated with lower-peace and higher-peace countries. Certainly, words alone do not capture all the linguistic subtleties of language, but they can serve as a good starting point to explore the linguistic differences between lower-peace and higher-peace cultures.
We used a novel data science approach to identify those words. These data science methods, developed in computer science, which are widely used in commerce, are now being increasingly applied to gain new understanding of systems in the physical, biological, medical, and social sciences. A classical approach would be to use theoretical concepts to generate sets of words expected to be more frequently found in lower-peace and higher-peace countries and then test whether that is indeed the case. Instead, we used a modern data science approach to identify those words that are the most important in predicting whether a country is lower-peace or higher-peace.
We found that the words that are most important in differentiating lower-peace and higher-peace countries are those shown in Figs 4 and 5. These words suggest that lower-peace countries are characterized by words of government, order, control and fear (e.g. government, state, court), while higher-peace countries are characterized by words of optimism for the future and fun (e.g. time, like, game). Words are both a cause and a consequence of the social processes that lead to lower or higher levels of peace. The link between these words and their associated social processes needs to be developed further. Having identified those words, at least, provides a starting point for that exploration.
Having trained the machine learning model to use words to recognize the differences between lower-peace and higher-peace countries gave us the opportunity to rank countries on their level of peace, as shown in Table 7. Current peace indices use conceptual frameworks to choose data believed to be indicators of peace. Our machine learning peace index is agnostic to such theoretical assumptions or frameworks. The parameters of the machine learning model arise only from its ability to use word frequencies to correctly classify countries as lower-peace or higher-peace. How does our quantitative machine learning peace index compare to other measures of peace? As can be seen from Table 7, the overall ranking of countries by our machine learning peace index is similar (linear regression r2 = 0.8349) to the overall rankings by five other peace indices based on their theoretical conceptual frameworks. We note that since our machine learning peace index correlates to the average of these 5 peace indices for each country, perhaps we could have used those averages to train our machine learning model. The issue in that approach is that those averages depend on the relative weights assigned to each index and we had no clear criteria on how to assign those weights. This means that part of the correlation between our machine learning index and the averages of the 5 peace indices may be due to that particular choice of equal weighting. Nonetheless, with that understanding, this still provides a useful relative comparison between the machine learning and the traditional peace indices. Recently, other machine learning methods have been used to show that events from the GDELT (Global Data on Events, Locations, and Tone) digital news database [44] successfully correlates with, and can even predict, the values of the GPI over time [39, 45]. Those studies used pre-assigned event categories, while our work here used machine learning to identify the words that differentiate lower and higher levels of peace without prior assumptions on what those words would be. Our approach, for example, has led to the unanticipated finding that news stories about “games”, “teams”, and “play” are representative of higher levels of peace.
In order to avoid difficulties in translation, we restricted our analysis to sources in English. This means that the data we analyzed may reflect a Western bias in the countries chosen because those countries have the most extensive news media in English, and as many of the higher peace countries are in the Global North, while the lower peace countries are in the Global South. That may influence the words determined from the lower-peace and higher-peace countries and the quantitative values of the machine learning model peace index.
Future directions for these studies include identifying the social processes reflected in the different sets of words in the lower-peace and higher-peace countries. One promising approach is to use the word frequencies to identify societies with “‘tight’ cultural groups that have strong norms and little tolerance for deviance while other ‘loose’ groups that have weaker norms and more tolerance for dissent.” [46, 47]. Some social challenges may be more successfully addressed by a tighter society and others by a looser society. More advanced natural language processing, such as Google’s BERT, Bidirectional Encoder Representations from Transformers [48] that captures more meaning-level information because it analyzes whole sentences at a time could also be used. A preliminary study by Liu et al. [49] using BERT, showed that the prediction accuracy only decreased 4% when the words in the articles were scrambled into a random order. This suggests that the word vocabulary alone, rather than more sophisticated linguistic features, plays a significant role in differentiating lower-peace from higher-peace countries. Our results here can also be tested by analyzing larger data sets that include more countries. The challenge here is to properly balance the increased data with the increased bias from the fewer news sources in countries where English is not the primary language. To expand our analysis beyond English, we also want to explore using other languages versions of BERT, such as multilingual BERT [48] which covers up to 104 languages as well as non-English monolingual BERTs in French [50], Spanish [51], Dutch [52], Chinese [53], Finnish [54], and Russian [55]. We are also considering analyzing more local geographic regions than countries which will allow us to study correlations with other cultural factors.
The peace indices shown in Table 2, and their values normalized from 0–100 shown in Table 3, have different values for the same country. What sense can we make of those differences? Each index uses its own assumptions as to what are the indicators of peace and how to weight their relative importance. Could it be that they are all correct? Understandings of peace likely vary in different contexts. As noted by Roger Mac Ginty [22], “Different communities are likely to define peace in different ways” (p. 59). We speculate that there is no one, single measure of peace. There could be different ways, ethnically, culturally, politically, socially, historically, economically, that countries can be peaceful and sustain their peaceful character. John M. Gottman and his collaborators [56, 57] have identified the 4 most negative emotions that “describe communication styles that, according to our research, can predict the end of a relationship.” Those communication styles are: criticism, contempt, defensiveness, and stonewalling, which are all characterized by an underlying lack of emphatic connection. Do all lower-peace countries share this same underlying lack of empathic connections? In the opening sentence of Ann Karenina, Leo Tolstoy writes, “All happy families are alike; every unhappy family is unhappy in its own way.” Perhaps, peace may be just the opposite of Tolstoy’s families. Perhaps, there are many ways countries can be peaceful, but only one way that they are not peaceful.
References
- 1. Luhmann N (1987) Soziale Systeme: Grundriß einer allgemeinen Theorie. Suhrkamp, Frankfurt
- 2.
Karlberg M (2011) Discourse Theory and Peace. In: Christie DJ (ed) The Encyclopedia of Peace Psychology. Blackwell Publishing Ltd, p. 87
- 3. Kimotho SG, Nyaga RN (2016) Digitized ethnic hate speech: Understanding effects of digital media hate speech on citizen journalism in Kenya. Adv Lan Lit Stu 7(3): 189–200
- 4. Ezeibe C (2021) Hate Speech and Election Violence in Nigeria. J Asi Afr Stu, 56(4): 919–935. https://doi.org/10.1177/0021909620951208
- 5. Soral W, Bilewicz M, Winiewski M (2018) Exposure to hate speech increases prejudice through desensitization. Agg Beh, 44(2): 136–146 pmid:29094365
- 6. PeaceTech Lab (2020) Combating Hate Speech. https://www.peacetechlab.org/hate-speech. Accessed 1 Aug 2022
- 7. Fry DP, Souillac G, Liebovitch LS, Coleman PT, Agan K, Nicholson-Cox E, et al. (2021). Societies within peace systems avoid war and build positive intergroup relationships. Humanities and Behavioral Sciences Communications 8, 17. https://doi.org/10.1057/s41599-020-00692-8
- 8. Diehl PF, Goertz G, Gallegos Y. (2019) Peace data: Concept, measurement, patterns, and research agenda. Con Man Pea Sci. https://doi.org/10.1177/0738894219870288
- 9.
Coleman PT, Deutsch M (eds) (2012) The Psychological Components of Sustainable Peace. Springer, New York
- 10.
Coleman P.T., Fisher J., Fry D.P., Liebovitch L. Chen-Carrel A., Souillac G. (2020). How to Live in Peace? Mapping the Science of Sustaining Peace: A Progress Report. American Psychologist
- 11.
Fry DP (2006) The human potential for peace: An anthropological challenge to assumptions about war and violence. Oxford University Press, USA
- 12.
Deutsch M, Coleman PT (2016) The psychological components of a sustainable peace: An introduction. In Brauch HG, Spring UO, Grin J, Scheffran J (eds) Handbook on sustainability Transition and Sustainable Peace, Springer, New York, p. 139
- 13. Diehl PF (2016) Exploring peace: Looking beyond war and negative peace. Int St Qua, 60(1):1–10
- 14.
Goertz G, Diehl PF, Balas A (2016) The puzzle of peace: The evolution of peace in the international system. Oxford University Press, USA
- 15.
Mahmoud Y, Makoond A (2017) Sustaining peace: What does it mean in practice? International Peace Institute
- 16. Advanced Consortium of Cooperation, Conflict, and Complexity (2018) Sustaining Peace Project. http://sustainingpeaceproject.com. Accessed 2 Dec 2021
- 17.
Bolívar A (2011) Language, Violent and Peaceful Uses of. In: Christie DJ (ed): The Encyclopedia of Peace Psychology. Blackwell Publishing Ltd, https://doi.org/10.1002/9780470672532.wbepp146
- 18. Friedrich P (2007) English for peace: Toward a framework of Peace Sociolinguistics. Wor Eng 26(1): 72–83. https://doi.org/10.1111/j.1467-971X.2007.00489.x
- 19.
Friedrich P (2019) Applied Linguistics in the Real World. Routledge, London
- 20. Gomes de Matos F (2000) Harmonizing and humanizing political discourse: The contribution of peace linguists. Peace and Conflict: J Peace Psych 6:339–344
- 21. Ngabonziza AJD (2013) The Importance of Language Studies in Conflict Resolution. J Afr Con Pea Stu 2(1): 33–37. http://dx.doi.org/10.5038/2325-484X.2.1.4
- 22. Mac Ginty R (2013) Indicators+: A proposal for everyday peace indicators. Eval Pro Pl 36(1): 56–63. pmid:22868180
- 23.
Caplan R (2019) Measuring Peace: Principles, Practices, and Politics. Oxford University Press, USA
- 24. Institute for Economics & Peace (2019) Global Peace Index 2019: Measuring Peace in a Complex World. http://visionofhumanity.org/reports. Accessed 1 Aug 2022
- 25. Institute for Economics & Peace (2019) Positive Peace Report 2019: Analysing the Factors that Sustain Peace. http://visionofhumanity.org/reports. Accessed 1 Aug 2022
- 26. Institute for Economics & Peace (2021) Positive Peace Report 2021: Analysing the Factors that Sustain Peace. http://visionofhumanity.org/reports. Accessed 1 Aug 2022
- 27. United Nations Development Programme (2021) Human Development Index. https://hdr.undp.org/data-center/human-development-index#/indicies/HDI. Accessed 1 Aug 2022
- 28. Helliwell JF, Layard R, Sachs JD (2019) World Happiness Report 2019. https://worldhappiness.report/ed/2019/. Accessed 1 Aug 2022
- 29.
Fund for Peace (2019) Fragile States Index Annual Report 2019. The Fund for Peace, Washington DC. https://fundforpeace.org/2019/04/10/fragile-states-index-2019/. Accessed 1 Aug 2022
- 30. Firchow P, Ginty RM (2017) Measuring peace: Comparability, commensurability, and complementarity using bottom-up indicators. Intl St Rev 19(1): 6–27. https://doi.org/10.1093/isr/vix001
- 31.
Yusuf S, Voss SJ (2018) The generations for peace institute compendium of participatory indicators of peace. Generations for Peace Institute
- 32. A743 (2019) Which is Better? Systemic (Holistic) or Symptomatic (Reductionistic) Approach to Data Science, Oct 7, 2019. https://medium.com/@A743241/which-is-better-systemic-bottom-up-or-symptomatic-top-down-approach-to-data-science-9bae2afca518. Accessed 23 Jan 2023
- 33. Reutter A (2020) Top-Down vs. Bottom-Up Approaches to Data Science, June 9, 2020. https://blog.dataiku.com/top-down-vs.-bottom-up-approaches-to-data-science. Accessed 23 Jan 2023
- 34. Investopedia Team (2022) Top-Down vs. Bottom-Up: What’s the Difference?, Updated September 06, 2022. https://www.investopedia.com/articles/investing/030116/topdown-vs-bottomup.asp Accessed 23 Jan 2023
- 35.
Lane H, Howard C, Hapke HM (2019) Natural Language Processing in Action. Manning, Shelter Island, NY
- 36.
Raschka S, Mirjalili V (2019) Python Machine Learning 3rd Ed. Packt>, Birmingham, UK
- 37. NOW (2021) News on the web corpus. https://www.english-corpora.org/now/
- 38. Jung J, Lee H, Kwon HJ, Mackenzie M, Lim TY (2021) power-of-peace-speech. https://github.com/mbmackenzie/ power-of-peace-speech/. Accessed 2 Dec 2021
- 39. Voukelatou V, Miliou I, Giannotti F, Pappalardo L (2022) Understanding peace through the world news. EPJ Data Science 11:2 pmid:35079561
- 40. Preacher KJ, Rucker DD, MacCallum RC, Nicewander WA (2005) Use of the Extreme Groups Approach: A Critical Reexamination and New Recommendations. Psy Met APA, 10(2): 178–192 pmid:15998176
- 41. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. (2011). Scikit-learn: Machine Learning in Python sklearn.ensemble.RandomForestClassifier -. J Mach Learn Res, 12:2825–2830, 2011b
- 42. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. (2011) Scikit-learn: Machine learning in Python sklearn.linear model.logisticregression J Mach Learn Res 12: 2825–2830
- 43. Cross-validation (statistics). Wikipedia. https://en.wikipedia.org/wiki/Cross Accessed 4 April 2023
- 44. Leetaru K (2013) The GDELT project. https://www.gdeltproject.org/
- 45. Voukelatou V, Pappalardo L, Miliou I, Gabrielli L, Giannotti F (2020) Estimating countries’ peace index through the lens of the world news as monitored by GDELT. Paper presented at IEEE 7th International Conference on Data Science and Advanced Analytics (DSAA) Sydney, 6–9 October 2020, p 216–225
- 46. Jackson JC, Gelfand MJ, De S, Fox A (2019) The loosening of american culture over 200 years is associated with a creativity–order trade-off. Nat Hum Beh 3(3):244–250
- 47. Gelfand MJ, Jackson JC, Pan X, Nau D, Pieper D, Denison E, et al. (2021) The relationship between cultural tightness–looseness and covid-19 cases and deaths: a global analysis. Lan Pla He 5(3):e135–e144 pmid:33524310
- 48. Devlin J, Chang MW, Lee K, Toutanova K (2019) BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. http://arxiv.org/abs/1810.04805, https://huggingface.co/bert-base-multilingual-cased
- 49. Liu H, Qi H, Wu X, Zhou Y, Zhu W. Power of Peace Speech https://github.com/wz2536/power-of-peace-speech_CapstoneFall2021 Accessed 2 Dec 2021
- 50. Le H, Vial L, Frej J, Segonne V, Coavoux M, Lecouteux B, et al. (2020) FlauBERT: Unsupervised Language Model Pre-training for French. https://arxiv.org/abs/1912.05372
- 51. Cant J., Chaperon G, Fuentes R, Ho J-H, Kang H, Perez J. (2023) Spanish Pre-trained BERT Model and Evaluation Data. https://arxiv.org/abs/2308.02976
- 52. de Vries W., can Cranenburgh A, Bisazza A, Caselli T, van Noord G, Nissim M. (2019) BERTje: A Dutch BERT Model. https://arxiv.org/abs/1912.09582
- 53. Cui Y, Che W, Liu T, Qin B, Yang Z. (2019) Pre-Training with Whole Word Masking for Chinese BERT. https://arxiv.org/abs/1906.08101
- 54. Virtanen A, Kanerva J, Ilo R, Luoma J, Luotolahti J, Salakoski T, et al. (2019) Multilingual is not enough: BERT for Finnish. https://arxiv.org/abs/1912.07076
- 55. Kuratov Y, Arkhipov M. (2019) Adaptation of Deep Bidirectional Multilingual Transformers for Russian Language. https://arxiv.org/abs/1905.07213
- 56.
Coan JA, Gottman JM (2007) The Specific Affect Coding System (SPAFF). In: Coan JA, Gottman JM (eds) Handbook of Emotion Elicitation and Assessment. Oxford University Press, New York, p 267–285
- 57. Lisitsa E (2022) The Four Horsemen: Criticism, Contempt, Defensiveness, and Stonewalling. https://www.gottman.com/blog/the-four-horsemen-recognizing-criticism-contempt-defensiveness-and-stonewalling/. Accessed 29 Jul 2022