Changes in suicide-related tweets before and during the COVID-19 pandemic in France: The importance of social media monitoring in public health prediction.

INTRODUCTION: The COVID-19 pandemic impacted mental health, as demonstrated by numerous studies. In recent years, especially during the pandemic, the use of social networks, including Twitter, increased. This suggests that this media could help with mental health monitoring, as attested by previous studies. METHOD: We conducted a multidisciplinary study on French tweets that were posted between January 1, 2019, and December 31, 2021. We selected the tweets via the Twitter API (Application Programming Interface) using five keywords relating to suicide: want to die, suicidal ideation, commit suicide, suicidal, and suicide attempt. A word frequency analysis was performed, and the data were compared with the number of emergency visits for suicidal ideation before and during the COVID-19 pandemic as recorded by the French national suicide observatory. RESULTS: We observed that 189,005 tweets were related to suicide in 2019, 261,993 in 2020 (+38.62% of that observed in 2019), and 301,177 in 2021 (+59.35% of that observed in 2019). We also observed an increase in the number of tweets containing control words in 2020 (+30.07% of that observed in 2019), but in 2021, the number almost fell back to the level of that in 2019 (+5.96% of that observed in 2019). Furthermore, the difference between both ratios (of suicide-related tweets and of tweets containing control words) was most significant during the third lockdown. The change in the number of suicide-related tweets followed a curve that overlapped with the change in the number of emergency visits following suicidal ideations, as reported by the French national suicide observatory. In conclusion, Twitter can be an adequate and reliable tool for screening for suicidal ideation in the general population.

Psychiatry on Twitter: Content Analysis of the Use of Psychiatric Terms in French.

BACKGROUND: With the advent of digital technology and specifically user-generated contents in social media, new ways emerged for studying possible stigma of people in relation with mental health. Several pieces of work studied the discourse conveyed about psychiatric pathologies on Twitter considering mostly tweets in English and a limited number of psychiatric disorders terms. This paper proposes the first study to analyze the use of a wide range of psychiatric terms in tweets in French. OBJECTIVE: Our aim is to study how generic, nosographic, and therapeutic psychiatric terms are used on Twitter in French. More specifically, our study has 3 complementary goals: (1) to analyze the types of psychiatric word use (medical, misuse, or irrelevant), (2) to analyze the polarity conveyed in the tweets that use these terms (positive, negative, or neural), and (3) to compare the frequency of these terms to those observed in related work (mainly in English). METHODS: Our study was conducted on a corpus of tweets in French posted from January 1, 2016, to December 31, 2018, and collected using dedicated keywords. The corpus was manually annotated by clinical psychiatrists following a multilayer annotation scheme that includes the type of word use and the opinion orientation of the tweet. A qualitative analysis was performed to measure the reliability of the produced manual annotation, and then a quantitative analysis was performed considering mainly term frequency in each layer and exploring the interactions between them. RESULTS: One of the first results is a resource as an annotated dataset. The initial dataset is composed of 22,579 tweets in French containing at least one of the selected psychiatric terms. From this set, experts in psychiatry randomly annotated 3040 tweets that corresponded to the resource resulting from our work. The second result is the analysis of the annotations showing that terms are misused in 45.33% (1378/3040) of the tweets and that their associated polarity is negative in 86.21% (1188/1378) of the cases. When considering the 3 types of term use, 52.14% (1585/3040) of the tweets are associated with a negative polarity. Misused terms related to psychotic disorders (721/1300, 55.46%) were more frequent to those related to depression (15/280, 5.4%). CONCLUSIONS: Some psychiatric terms are misused in the corpora we studied, which is consistent with the results reported in related work in other languages. Thanks to the great diversity of studied terms, this work highlighted a disparity in the representations and ways of using psychiatric terms. Moreover, our study is important to help psychiatrists to be aware of the term use in new communication media such as social networks that are widely used. This study has the huge advantage to be reproducible thanks to the framework and guidelines we produced so that the study could be renewed in order to analyze the evolution of term usage. While the newly build dataset is a valuable resource for other analytical studies, it could also serve to train machine learning algorithms to automatically identify stigma in social media.

Emotionally Informed Hate Speech Detection: A Multi-target Perspective.

Hate Speech and harassment are widespread in online communication, due to users' freedom and anonymity and the lack of regulation provided by social media platforms. Hate speech is topically focused (misogyny, sexism, racism, xenophobia, homophobia, etc.), and each specific manifestation of hate speech targets different vulnerable groups based on characteristics such as gender (misogyny, sexism), ethnicity, race, religion (xenophobia, racism, Islamophobia), sexual orientation (homophobia), and so on. Most automatic hate speech detection approaches cast the problem into a binary classification task without addressing either the topical focus or the target-oriented nature of hate speech. In this paper, we propose to tackle, for the first time, hate speech detection from a multi-target perspective. We leverage manually annotated datasets, to investigate the problem of transferring knowledge from different datasets with different topical focuses and targets. Our contribution is threefold: (1) we explore the ability of hate speech detection models to capture common properties from topic-generic datasets and transfer this knowledge to recognize specific manifestations of hate speech; (2) we experiment with the development of models to detect both topics (racism, xenophobia, sexism, misogyny) and hate speech targets, going beyond standard binary classification, to investigate how to detect hate speech at a finer level of granularity and how to transfer knowledge across different topics and targets; and (3) we study the impact of affective knowledge encoded in sentic computing resources (SenticNet, EmoSenticNet) and in semantically structured hate lexicons (HurtLex) in determining specific manifestations of hate speech. We experimented with different neural models including multitask approaches. Our study shows that: (1) training a model on a combination of several (training sets from several) topic-specific datasets is more effective than training a model on a topic-generic dataset; (2) the multi-task approach outperforms a single-task model when detecting both the hatefulness of a tweet and its topical focus in the context of a multi-label classification approach; and (3) the models incorporating EmoSenticNet emotions, the first level emotions of SenticNet, a blend of SenticNet and EmoSenticNet emotions or affective features based on Hurtlex, obtained the best results. Our results demonstrate that multi-target hate speech detection from existing datasets is feasible, which is a first step towards hate speech detection for a specific topic/target when dedicated annotated data are missing. Moreover, we prove that domain-independent affective knowledge, injected into our models, helps finer-grained hate speech detection.

Combining glass box and black box evaluations in the identification of heart disease risk factors and their temporal relations from clinical records.

BACKGROUND: The determination of risk factors and their temporal relations in natural language patient records is a complex task which has been addressed in the i2b2/UTHealth 2014 shared task. In this context, in most systems it was broadly decomposed into two sub-tasks implemented by two components: entity detection, and temporal relation determination. Task-level ("black box") evaluation is relevant for the final clinical application, whereas component-level evaluation ("glass box") is important for system development and progress monitoring. Unfortunately, because of the interaction between entity representation and temporal relation representation, glass box and black box evaluation cannot be managed straightforwardly at the same time in the setting of the i2b2/UTHealth 2014 task, making it difficult to assess reliably the relative performance and contribution of the individual components to the overall task. OBJECTIVE: To identify obstacles and propose methods to cope with this difficulty, and illustrate them through experiments on the i2b2/UTHealth 2014 dataset. METHODS: We outline several solutions to this problem and examine their requirements in terms of adequacy for component-level and task-level evaluation and of changes to the task framework. We select the solution which requires the least modifications to the i2b2 evaluation framework and illustrate it with our system. This system identifies risk factor mentions with a CRF system complemented by hand-designed patterns, identifies and normalizes temporal expressions through a tailored version of the Heideltime tool, and determines temporal relations of each risk factor with a One Rule classifier. RESULTS: Giving a fixed value to the temporal attribute in risk factor identification proved to be the simplest way to evaluate the risk factor detection component independently. This evaluation method enabled us to identify the risk factor detection component as most contributing to the false negatives and false positives of the global system. This led us to redirect further effort to this component, focusing on medication detection, with gains of 7 to 20 recall points and of 3 to 6 F-measure points depending on the corpus and evaluation. CONCLUSION: We proposed a method to achieve a clearer glass box evaluation of risk factor detection and temporal relation detection in clinical texts, which can provide an example to help system development in similar tasks. This glass box evaluation was instrumental in refocusing our efforts and obtaining substantial improvements in risk factor detection.