A machine learning approach to detect potentially harmful and protective suicide-related content in broadcast media.

Suicide-related media content has preventive or harmful effects depending on the specific content. Proactive media screening for suicide prevention is hampered by the scarcity of machine learning approaches to detect specific characteristics in news reports. This study applied machine learning to label large quantities of broadcast (TV and radio) media data according to media recommendations reporting suicide. We manually labeled 2519 English transcripts from 44 broadcast sources in Oregon and Washington, USA, published between April 2019 and March 2020. We conducted a content analysis of media reports regarding content characteristics. We trained a benchmark of machine learning models including a majority classifier, approaches based on word frequency (TF-IDF with a linear SVM) and a deep learning model (BERT). We applied these models to a selection of more simple (e.g., focus on a suicide death), and subsequently to putatively more complex tasks (e.g., determining the main focus of a text from 14 categories). Tf-idf with SVM and BERT were clearly better than the naive majority classifier for all characteristics. In a test dataset not used during model training, F1-scores (i.e., the harmonic mean of precision and recall) ranged from 0.90 for celebrity suicide down to 0.58 for the identification of the main focus of the media item. Model performance depended strongly on the number of training samples available, and much less on assumed difficulty of the classification task. This study demonstrates that machine learning models can achieve very satisfactory results for classifying suicide-related broadcast media content, including multi-class characteristics, as long as enough training samples are available. The developed models enable future large-scale screening and investigations of broadcast media.

"Breaking the Silence" Suicide Prevention Media Campaign in Oregon, April 7-14, 2019.

Background: Between April 7 and 14, 2019, the "Breaking the Silence" media engagement campaign was launched in Oregon. Aims: We aimed to assess the consistency of media content related to the campaign with media guidelines and the quantitative footprint on Twitter (now X) over time. Method: Media items related to the campaign were analyzed regarding focus and consistency with media guidelines for suicide reporting and compared with other suicide-related reports published in the same time frame, as well as with reporting in Washington, the control region. Tweets related to the campaign were retrieved to assess the social media footprint. Results: There were n = 104 media items in the campaign month, mainly in the campaign week. Items typically used a narrative featuring suicide advocacy or policy/prevention programs. As compared to other items with a similar focus, they scored better on several protective characteristics listed in media recommendations. Stories of coping with adversity, however, were scarce. The social media footprint on Twitter was small. Limitations: Inability to make causal claims about campaign impact. Conclusion: Media items from the Breaking the Silence campaign appeared mainly consistent with media guidelines, but some aspects, such as stories of recovery, were under-represented.

LEIA: Linguistic Embeddings for the Identification of Affect.

The wealth of text data generated by social media has enabled new kinds of analysis of emotions with language models. These models are often trained on small and costly datasets of text annotations produced by readers who guess the emotions expressed by others in social media posts. This affects the quality of emotion identification methods due to training data size limitations and noise in the production of labels used in model development. We present LEIA, a model for emotion identification in text that has been trained on a dataset of more than 6 million posts with self-annotated emotion labels for happiness, affection, sadness, anger, and fear. LEIA is based on a word masking method that enhances the learning of emotion words during model pre-training. LEIA achieves macro-F1 values of approximately 73 on three in-domain test datasets, outperforming other supervised and unsupervised methods in a strong benchmark that shows that LEIA generalizes across posts, users, and time periods. We further perform an out-of-domain evaluation on five different datasets of social media and other sources, showing LEIA's robust performance across media, data collection methods, and annotation schemes. Our results show that LEIA generalizes its classification of anger, happiness, and sadness beyond the domain it was trained on. LEIA can be applied in future research to provide better identification of emotions in text from the perspective of the writer.

Power pose effects on approach and avoidance decisions in response to social threat.

Individuals' opportunities for action in threatening social contexts largely depend on their social power. While powerful individuals can afford to confront aggressors and dangers, powerless individuals need others' support and better avoid direct challenges. Here, we investigated if adopting expansive or contracted poses, which signal dominance and submission, impacts individuals' approach and avoidance decisions in response to social threat signals using a within-subject design. Overall, participants more often chose to avoid rather than to approach angry individuals, but showed no clear approach or avoidance preference for fearful individuals. Crucially, contracted poses considerably increased the tendency to avoid angry individuals, whereas expansive poses induced no substantial changes. This suggests that adopting power-related poses may impact action decisions in response to social threat signals. The present results emphasize the social function of power poses, but should be replicated before drawing strong conclusions.

Social Drivers and Algorithmic Mechanisms on Digital Media.

On digital media, algorithms that process data and recommend content have become ubiquitous. Their fast and barely regulated adoption has raised concerns about their role in well-being both at the individual and collective levels. Algorithmic mechanisms on digital media are powered by social drivers, creating a feedback loop that complicates research to disentangle the role of algorithms and already existing social phenomena. Our brief overview of the current evidence on how algorithms affect well-being, misinformation, and polarization suggests that the role of algorithms in these phenomena is far from straightforward and that substantial further empirical research is needed. Existing evidence suggests that algorithms mostly reinforce existing social drivers, a finding that stresses the importance of reflecting on algorithms in the larger societal context that encompasses individualism, populist politics, and climate change. We present concrete ideas and research questions to improve algorithms on digital platforms and to investigate their role in current problems and potential solutions. Finally, we discuss how the current shift from social media to more algorithmically curated media brings both risks and opportunities if algorithms are designed for individual and societal flourishing rather than short-term profit.

Collective emotions during the COVID-19 outbreak.

The COVID-19 pandemic has exposed the world's population to unprecedented health threats and changes to social life. High uncertainty about the novel disease and its social and economic consequences, together with increasingly stringent governmental measures against the spread of the virus, likely elicited strong emotional responses. We analyzed the digital traces of emotional expressions in tweets during 5 weeks after the start of outbreaks in 18 countries and six different languages. We observed an early strong upsurge of anxiety-related terms in all countries, which was related to the growth in cases and increases in the stringency of governmental measures. Anxiety expression gradually relaxed once stringent measures were in place, possibly indicating that people were reassured. Sadness terms rose and anger terms decreased with or after an increase in the stringency of measures and remained stable as long as measures were in place. Positive emotion words only decreased slightly and briefly in a few countries. Our results reveal some of the most enduring changes in emotional expression observed in long periods of social media data. Such sustained emotional expression could indicate that interactions between users led to the emergence of collective emotions. Words that frequently occurred in tweets suggest a shift in topics of conversation across all emotions, from political ones in 2019, to pandemic related issues during the outbreak, including everyday life changes, other people, and health. This kind of time-sensitive analyses of large-scale samples of emotional expression have the potential to inform risk communication. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

Association of 7 million+ tweets featuring suicide-related content with daily calls to the Suicide Prevention Lifeline and with suicides, United States, 2016-2018.

OBJECTIVE: The aim of this study was to assess associations of various content areas of Twitter posts with help-seeking from the US National Suicide Prevention Lifeline (Lifeline) and with suicides. METHODS: We retrieved 7,150,610 suicide-related tweets geolocated to the United States and posted between 1 January 2016 and 31 December 2018. Using a specially devised machine-learning approach, we categorized posts into content about prevention, suicide awareness, personal suicidal ideation without coping, personal coping and recovery, suicide cases and other. We then applied seasonal autoregressive integrated moving average analyses to assess associations of tweet categories with daily calls to the US National Suicide Prevention Lifeline (Lifeline) and suicides on the same day. We hypothesized that coping-related and prevention-related tweets are associated with greater help-seeking and potentially fewer suicides. RESULTS: The percentage of posts per category was 15.4% (standard deviation: 7.6%) for awareness, 13.8% (standard deviation: 9.4%) for prevention, 12.3% (standard deviation: 9.1%) for suicide cases, 2.4% (standard deviation: 2.1%) for suicidal ideation without coping and 0.8% (standard deviation: 1.7%) for coping posts. Tweets about prevention were positively associated with Lifeline calls (B = 1.94, SE = 0.73, p = 0.008) and negatively associated with suicides (B = -0.11, standard error = 0.05, p = 0.038). Total number of tweets were negatively associated with calls (B = -0.01, standard error = 0.0003, p = 0.007) and positively associated with suicide, (B = 6.4 × 10-5, standard error = 2.6 × 10-5, p = 0.015). CONCLUSION: This is the first large-scale study to suggest that daily volume of specific suicide-prevention-related social media content on Twitter corresponds to higher daily levels of help-seeking behaviour and lower daily number of suicide deaths. PREREGISTRATION: As Predicted, #66922, 26 May 2021.

Detecting Potentially Harmful and Protective Suicide-Related Content on Twitter: Machine Learning Approach.

BACKGROUND: Research has repeatedly shown that exposure to suicide-related news media content is associated with suicide rates, with some content characteristics likely having harmful and others potentially protective effects. Although good evidence exists for a few selected characteristics, systematic and large-scale investigations are lacking. Moreover, the growing importance of social media, particularly among young adults, calls for studies on the effects of the content posted on these platforms. OBJECTIVE: This study applies natural language processing and machine learning methods to classify large quantities of social media data according to characteristics identified as potentially harmful or beneficial in media effects research on suicide and prevention. METHODS: We manually labeled 3202 English tweets using a novel annotation scheme that classifies suicide-related tweets into 12 categories. Based on these categories, we trained a benchmark of machine learning models for a multiclass and a binary classification task. As models, we included a majority classifier, an approach based on word frequency (term frequency-inverse document frequency with a linear support vector machine) and 2 state-of-the-art deep learning models (Bidirectional Encoder Representations from Transformers [BERT] and XLNet). The first task classified posts into 6 main content categories, which are particularly relevant for suicide prevention based on previous evidence. These included personal stories of either suicidal ideation and attempts or coping and recovery, calls for action intending to spread either problem awareness or prevention-related information, reporting of suicide cases, and other tweets irrelevant to these 5 categories. The second classification task was binary and separated posts in the 11 categories referring to actual suicide from posts in the off-topic category, which use suicide-related terms in another meaning or context. RESULTS: In both tasks, the performance of the 2 deep learning models was very similar and better than that of the majority or the word frequency classifier. BERT and XLNet reached accuracy scores above 73% on average across the 6 main categories in the test set and F1-scores between 0.69 and 0.85 for all but the suicidal ideation and attempts category (F1=0.55). In the binary classification task, they correctly labeled around 88% of the tweets as about suicide versus off-topic, with BERT achieving F1-scores of 0.93 and 0.74, respectively. These classification performances were similar to human performance in most cases and were comparable with state-of-the-art models on similar tasks. CONCLUSIONS: The achieved performance scores highlight machine learning as a useful tool for media effects research on suicide. The clear advantage of BERT and XLNet suggests that there is crucial information about meaning in the context of words beyond mere word frequencies in tweets about suicide. By making data labeling more efficient, this work has enabled large-scale investigations on harmful and protective associations of social media content with suicide rates and help-seeking behavior.

Validating daily social media macroscopes of emotions.

Measuring sentiment in social media text has become an important practice in studying emotions at the macroscopic level. However, this approach can suffer from methodological issues like sampling biases and measurement errors. To date, it has not been validated if social media sentiment can actually measure the temporal dynamics of mood and emotions aggregated at the level of communities. We ran a large-scale survey at an online newspaper to gather daily mood self-reports from its users, and compare these with aggregated results of sentiment analysis of user discussions. We find strong correlations between text analysis results and levels of self-reported mood, as well as between inter-day changes of both measurements. We replicate these results using sentiment data from Twitter. We show that a combination of supervised text analysis methods based on novel deep learning architectures and unsupervised dictionary-based methods have high agreement with the time series of aggregated mood measured with self-reports. Our findings indicate that macro level dynamics of mood expressed on an online platform can be tracked with social media text, especially in situations of high mood variability.

Dashboard of Sentiment in Austrian Social Media During COVID-19.

To track online emotional expressions on social media platforms close to real-time during the COVID-19 pandemic, we built a self-updating monitor of emotion dynamics using digital traces from three different data sources in Austria. This allows decision makers and the interested public to assess dynamics of sentiment online during the pandemic. We used web scraping and API access to retrieve data from the news platform derstandard.at, Twitter, and a chat platform for students. We documented the technical details of our workflow to provide materials for other researchers interested in building a similar tool for different contexts. Automated text analysis allowed us to highlight changes of language use during COVID-19 in comparison to a neutral baseline. We used special word clouds to visualize that overall difference. Longitudinally, our time series showed spikes in anxiety that can be linked to several events and media reporting. Additionally, we found a marked decrease in anger. The changes lasted for remarkably long periods of time (up to 12 weeks). We have also discussed these and more patterns and connect them to the emergence of collective emotions. The interactive dashboard showcasing our data is available online at http://www.mpellert.at/covid19_monitor_austria/. Our work is part of a web archive of resources on COVID-19 collected by the Austrian National Library.

Repeatedly adopting power postures does not affect hormonal correlates of dominance and affiliative behavior.

BACKGROUND: Adopting expansive vs. constrictive postures related to high vs. low levels of social power has been suggested to induce changes in testosterone and cortisol levels, and thereby to mimic hormonal correlates of dominance behavior. However, these findings have been challenged by several non-replications recently. Despite this growing body of evidence that does not support posture effects on hormone levels, the question remains as to whether repeatedly holding postures over time and/or assessing hormonal responses at different time points would yield different outcomes. The current study assesses these methodological characteristics as possible reasons for previous null-findings. Additionally, it investigates for the first time whether expansive and constrictive postures impact progesterone levels, a suggested correlate of affiliative motives and behavior. By testing the effects of repeated but short posture manipulations in between the blocks of a social task while using a cover story, it further fulfills the conditions previously raised as potentially necessary for the effects to occur. METHODS: A total of 82 male participants repeatedly adopted an expansive or constrictive posture for 2 min in between blocks of a task that consisted in categorizing faces based on first impressions. Saliva samples were taken at two different time points in a time window in which hormonal responses to stress, competition and other manipulations are known to be strongest. RESULTS: Neither testosterone and cortisol levels linked to dominance behaviors, nor progesterone levels related to affiliative tendencies, responded differently to adopting expansive as opposed to constrictive postures. The present results suggest that even repeated power posing in a context where social stimuli are task-relevant does not elicit changes in hormone levels.

Impact of self-esteem and sex on stress reactions.

Positive self-evaluation is a major psychological resource modulating stress coping behavior. Sex differences have been reported in self-esteem as well as stress reactions, but so far their interactions have not been investigated. Therefore, we investigated sex-specific associations of self-esteem and stress reaction on behavioral, hormonal and neural levels. We applied a commonly used fMRI-stress task in 80 healthy participants. Men compared to women showed higher activation during stress in hippocampus, precuneus, superior temporal gyrus (STG) and insula. Furthermore, men outperformed women in the stress task and had higher cortisol and testosterone levels than women after stress. Self-esteem had an impact on precuneus, insula and STG activation during stress across the whole group. During stress, men recruit regions associated with emotion and stress regulation, self-referential processing and cognitive control more strongly than women. Self-esteem affects stress processing, however in a sex-independent fashion: participants with lower self-esteem show higher activation of regions involved in emotion and stress regulation, self-referential processing and cognitive control. Taken together, our data suggest that men are more engaged during the applied stress task. Across women and men, lower self-esteem increases the effort in emotion and stress processing and cognitive control, possibly leading to self-related thoughts in stressful situations.