Investigating Moral Foundations from Web Trending Topics
aa r X i v : . [ c s . S I] F e b I NVESTIGATING M OR AL F OUNDATIONS FROM W EB T R ENDING T OPIC S
Jean Marie Tshimula
Department of Computer ScienceUniversité de SherbrookeQC J1K 2R1, Canada [email protected]
Belkacem Chikhaoui
LICEF Research CenterUniversité TÉLUQQC H2S 3L5, Canada [email protected]
Shengrui Wang
Department of Computer ScienceUniversité de SherbrookeQC J1K 2R1, Canada [email protected] A BSTRACT
Moral foundations theory helps understand differences in morality across cultures. In this paper,we propose a model to predict moral foundations (MF) from social media trending topics. We alsoinvestigate whether differences in MF influence emotional traits. Our results are promising and leaveroom for future research avenues.
Nowadays, media, governments, and organizations keep a constant eye on social media for uncovering the very latesttrends. Web trending topics include various opinions on the matters covered in the community. The exploitation oftrending topics depends strongly on the goals and interests of each entity. For instance, healthcare organizations maystudy social media language associated with trending topics to track mental health, symptom mentions and changesin people’s well-being throughout the pandemic. A government may continuously monitor web trends for identifyingrapidly growing divisive and controversial topics that can produce political tensions and an endlessly chaotic situation,and tarnish the country’s reputation and image.While some trends may involve strong emotional and passionate debate, we need to be clear about the fact that somepeople may be influenced by a wide variety of social and emotional forces and even indulged in emotional manip-ulation. Research shows that psycholinguistic features and sentiment analysis can provide substantial insights intothe issues that affect emotional stability, intensity, and reactions [5]. In this paper, we are interested in examiningmoral foundations theory to discern moral differences in a broad spectrum of opinion on social media. It is of greatimportance to understand moral differences at cultural and individual levels because morality guides human social in-teractions and can potentially conduct to divergence of opinion, polarity, violence, and hostility when there are moralshocks within a community. Understanding moral foundations can yield promising results in terms of perceiving theintended meaning of the text data because the concept of morality provides additional information on the unobservablecharacteristics of information processing and non-conscious cognitive processes [3], [7]. Researchers believe that fivepsychological dimensions are functioning as foundations for moralities around the world [4].The five moral foundations are care/harm, fairness/cheating, loyalty/betrayal, authority/subversion, and pu-rity/degradation. Each dimension possesses virtues and vices:• Care/harm is associated with the protection of self and others from harm’s way.• Fairness/cheating is related to the evolutionary process of reciprocal altruism and generates ideas of justice,rights, and autonomy.• Loyalty/betrayal is based on the expressions of self-sacrifice for both ends of the virtue-vice spectrum, suchas patriotism-betrayal, faithfulness-unfaithfulness.• Authority/subversion underlies virtues of leadership and followership, including deference to legitimate au-thority and respect for traditions. Purity/degradation is associated with sanctity in the virtue dimension and degradation and pollution in thevice dimension.To the best of our knowledge, our paper is the first to address the problem of moral foundations and emotional traitsin web trending topics. We utilized Moral Foundations Dictionary (MFD) and MoralStrength [1] in which thefive aforementioned psychological dimensions are considered. We combined MFD and MoralStrength and removedduplicate words for each dimension. We considered the words of each dimension as labels and then resorted to zero-shot text classification (ZSC) [11] to assign a probability to each label in text data.With the probabilities of labels obtained, we conducted 10-fold cross-validation to split our training and testing setsand trained a support vector machine (SVM) classifier to predict each dimension. Additionally, we used MFD-MoralStrength together with the Linguistic Inquiry and Word Count (LIWC) to investigate whether differences inmoral foundations influence some emotional traits. LIWC dictionary is a widely used psychometrically validated sys-tem for psychology-related analysis of language and word classification [8]. LIWC includes word categories that havepre-labeled meanings created by psychologists. LIWC categories have been also been independently evaluated fortheir correlation with psychological concepts. For each input sequence, we computed the number of observed words,using LIWC and focusing on five LIWC subcategories of psychological processes: positive emotion, negative emotion,anger, anxiety, and sadness. Specifically, we performed the Pearson correlation ( r ) between the MFD-MoralStrengthword scores and LIWC features extracted from the text data.We conducted extensive experiments to investigate two trending topics: coronavirus lockdown measures and the West-ern Exit (WEXIT) separatist movement in Canada. We collected 857,294 coronavirus lockdown-related tweets postedbetween 12 March 2020 and 25 May 2020. Specifically, we extracted tweets bearing the words or hashtags: covid,coronavirus, Tables 1 and 2 show the results of the Pearson correlation ( r ) between moral word scores and psycholinguistic featuresextracted from the lockdown and WEXIT dataset, respectively. For coronavirus lockdown, we observe that all correla-tions between positive emotion and virtue moral foundations ( Care , Fairness , Loyalty , Authority , Purity )as well as all correlations between negative emotion and vice moral foundations (
Harm , Cheating , Betrayal , Degradation ) are statistically significant at p < . ; except for one vice moral foundation, Subversion ( p > . ). We report that Degradation is associated with a relatively high correlation with negative emotion , anger , anxiety , and sadness ( p < . ) during the lockdown period. This reflects the feelings that touch directlyon mental health.Table 1: Pearson correlation ( r ) between moral word scores and LIWC features during coronavirus lockdown inCanada. Care Harm Fairness Cheating Loyalty Betrayal Authority Subversion Purity DegradationPos. emotion 0.151 0.014 0.183 0.161 0.246 -0.021 0.053 0.027 0.186 -0.084Neg. emotion -0.093 0.131 0.054 -0.022 -0.042 -0.030 0.017 0.235 -0.094 0.191Anger -0.041 0.128 0.105 -0.019 -0.093 0.191 -0.052 0.176 0.058 0.173Anxiety 0.038 -0.189 0.033 0.101 0.097 -0.013 -0.006 0.238 0.003 0.219Sadness 0.195 0.081 0.132 0.004 -0.093 0.155 0.032 0.107 -0.031 0.112Research reveals that many Canadians have seen their stress levels double since the onset of the pandemic and arestruggling with fear and uncertainty about their own and their loved ones’ health [2]. Survey research conducted byMental Health Research Canada found that feelings of depression are rising constantly [6]. Before the pandemic, 7%of Canadians reported high levels of depression. This rate has risen to 16% during the lockdown period and 22%predict high levels of depression if social isolation continues for two more months. Our results are alarming andindicate potential signals relevant to mental health that can aid mental health services in assessing the impact of thepandemic on the population and implementing healthier coping strategies to build resilience. https://moralfoundations.org https://github.com/oaraque/moral-foundations r ) between moral word scores and LIWC features during a specific period of WEXIT.Care Harm Fairness Cheating Loyalty Betrayal Authority Subversion Purity DegradationPos. emotion 0.252 -0.044 0.209 -0.093 0.102 -0.097 0.160 0.122 0.103 0.017Neg. emotion 0.057 0.203 0.073 0.154 -0.046 0.024 0.188 0.001 0.020 0.113Anger -0.026 0.145 0.003 -0.059 0.004 0.180 0.028 0.163 0.064 0.154Anxiety -0.025 0.152 -0.069 0.271 0.171 0.163 0.166 0.132 0.017 0.150Sadness 0.283 -0.089 0.152 -0.032 0.168 0.021 -0.051 -0.059 0.258 0.008Table 3: Classification results for moral foundations using ZSC+SVM. F-1 is used to evaluate the model performance.Care Harm Fairness Cheating Loyalty Betrayal Authority Subversion Purity DegradationLockdown 0.671 0.673 0.704 0.647 0.688 0.708 0.706 0.710 0.640 0.692Wexit 0.697 0.630 0.708 0.669 0.697 0.658 0.671 0.693 0.688 0.645As for the WEXIT, it refers to a movement that advocates for separation from Canada. We report that the correla-tions between all vice moral foundations and negative emotion and anger as well as all virtue moral foundationsand anxiety and sadness are statistically significant at p < . . We also note that there are some significantcorrelations between vice moral foundations and negative emotion , and virtue moral foundations and positiveemotion ( p < . ). Our results show relatively low-levels of sadness for Harm , Subversion , and
Degradation ( p > . ). We observe that WEXIT conversations include dominant emotional traits for vice moral foundations.This could be considered as strong evidence to argue that this trending topic may have sparked stormy debates andemotional statements. Identifying WEXIT proponents and opponents normally requires further analysis such as stancedetection [9], measures of divergent opinions, and community detection to track persistent members of ever-growingcommunities and their linguistic idiosyncrasies.Table 3 presents the performance results for virtue and vice moral dimensions classification during coronavirus lock-down measures and the WEXIT movement in Canada. We observe that the F-1 scores are higher and show the abilityto predict moral foundations, with F-1 scores of over 0.6 for the ten classes. Our model leverages large-scale socialmedia text data stemming from trending topics. The absence of adequate annotated data on moral foundations maybe challenging. To overcome this problem, we used psychologically validated and annotated dictionaries that indicatemoral foundations’ dimensions. We applied these resources to track moral foundations using ZSC, a model that doesnot require data to be annotated beforehand to predict text data; it learns a classifier on one set of labels and thenevaluates on a different set of labels that the classifier has never seen before. Note that the lack of annotated data doesnot affect the generalizability of the findings and model performance. We presented the first experiments towards investigating moral foundations from large-scale social media text datafrom trending topics. Our results provide strong evidence that we can predict moral foundations with an accuracy ex-ceeding 0.6 and we can jointly investigate emotional traits and moral foundations. Though the results are encouraging,this problem leaves room for future research. In the future, we aim to study natural language inference and behavioraldeterioration in moral foundations [10].
References [1] Oscar Araque and Lorenzo Gatti and Kyriaki Kalimeri. MoralStrength: Exploiting a moral lexicon and embeddingsimilarity for moral foundations prediction.
Knowledge-Based Systems , 191(105184), 2020.[2] The Centre for Addiction and Mental Health (CAMH). Mental health in Canada: Covid-19 and beyond".
CAMHPolicy Advice (July 2020 ), 2020.[3] Jesse Graham. Morality beyond the lab.
Science , 345(6202):1242–1242, 2014.[4] Jonathan Haidt and Jesse Graham. When morality opposes justice: Conservatives have moral intuitions thatliberals may not recognize.
Social Justice Research , 20(1):98–115, 2007.35] Jeffrey H. Kahn and Renée M. Tobin and Audra E. Massey and Jennifer A. Anderson. Measuring emotionalexpression with the linguistic inquiry and word count.
The American Journal of Psychology , 120(2):263–286,2007.[6] Mental Health Research Canada (MHRC). Mental health in crisis: How covid-19 is impacting Canadians". https://bit.ly/2UUvyKU (2020, accessed 07 Jan. 2021).[7] Ece C. Mutlu and Toktam A. Oghaz and Ege Tütüncüler and I. Garibay. Do bots have moral judgement? Thedifference between bots and humans in moral rhetoric.
Proceeding of 2020 IEEE/ACM International Conferenceon Advances in Social Networks Analysis and Mining , 2020.[8] James W. Pennebaker and Ryan L. Boyd and Kayla N Jordan and Kate Blackburn. The development and psycho-metric properties of LIWC2015.
Austin, TX: University of Texas at Austin , 2015.[9] Jean Marie Tshimula and Belkacem Chikhaoui and Shengrui Wang. A pre-training approach for stance classifica-tion in online forums.
Proceeding of 2020 IEEE/ACM International Conference on Advances in Social NetworksAnalysis and Mining , 2020.[10] Jean Marie Tshimula and Belkacem Chikhaoui and Shengrui Wang. On predicting behavioral deteriorationin online discussion forums.
Proceeding of 2020 IEEE/ACM International Conference on Advances in SocialNetworks Analysis and Mining , 2020.[11] Wenpeng Yin and Jamaal Hay and Dan Roth. Benchmarking zero-shot text classification: Datasets, evalua-tion and entailment approach.