An Exploratory Analysis of the Relation between Offensive Language and Mental Health

Ana-Maria Bucur, Marcos Zampieri, Liviu P. DinuACL 2021

TL;DR

This paper explores the connection between offensive language and mental health, specifically depression. By applying state-of-the-art offensive language detection models to social media posts from individuals with and without depression, the study finds that depressed individuals are more likely to use offensive language, especially those showing signs of depression. This work opens new avenues for understanding the linguistic markers of mental health conditions.

What Problem Does This Paper Address?

The pervasive nature of offensive language on social media has been a significant area of research, with numerous computational models developed to identify and categorize it. However, most of this research focuses on general population datasets, leaving a gap in understanding how offensive language is used by specific groups, particularly those with mental health conditions. While some studies have begun to explore this, such as identifying increased profanity use in users with mood disorders, there's a need for more empirical evidence and a deeper analysis of the nature of offensive language used by individuals with depression. This paper aims to fill this gap by investigating the relationship between offensive language and depression using publicly available mental health datasets.

Methods

The researchers employed a two-pronged approach using transformer-based models, specifically BERT.

  1. Offensive Language Detection and Categorization: They fine-tuned a BERT model on the Offensive Language Identification Dataset (OLID) to classify posts into "offensive" (OFF) vs. "non-offensive" (NOT) categories (Level A). A second BERT model was fine-tuned on OLID for categorizing offensive language into "targeted insult or threats" (TIN) vs. "untargeted profanity" (UNT) (Level B). The offensive score for each post was derived from the softmax output of the BERT model.
  2. Signs of Depression Detection: To identify posts exhibiting signs of depression within the mental health datasets, they utilized a heuristic approach proposed by Ríssola et al. (2020). This heuristic combines sentiment polarity (using TextBlob), emotion detection (from EmoLex), and a depression score (from NRC Affect Intensity Lexicon). Posts were flagged as showing signs of depression if they had negative polarity, contained sadness or disgust emotions, and had a depression score above 0.1.

Datasets

The study utilized three publicly available English datasets:

  • Offensive Language Identification Dataset (OLID): This dataset, containing 14,100 manually annotated Twitter posts, was used for training and evaluating the offensive language detection and categorization models. It features a hierarchical annotation taxonomy.
  • Reddit Self-reported Depression Diagnosis (RSDD): This dataset comprises posts from users with self-reported depression and control users. The authors used the training split, which includes over 5 million posts from depressed users and over 30 million from control users. Posts containing explicit depression terms were removed to anonymize users.
  • eRisk 2018: This dataset also contains posts from users labeled with depression and control users. The study used both train and test splits, totaling approximately 90,000 submissions from depressed users and 985,000 from control users. In this dataset, only posts with the exact mention of diagnosis were removed.

Key Insights and Ideas

  • Depression and Offensive Language Correlation: The most striking finding is the statistically significant difference in offensive language use between individuals with depression and control groups. Depressed individuals, particularly those showing signs of depression, are more likely to use offensive language.
  • Emotion Regulation as a Potential Driver: The use of offensive language might serve as an emotion regulation strategy for depressed individuals to alleviate distress, drawing parallels with studies on pain and swearing.
  • Nature of Offensive Language Differs: While depressed individuals use more offensive language overall, those showing signs of depression are less likely to use targeted offensive language. Instead, their offensive language tends to be untargeted profanity.
  • Negative Sentiment Dominance: Posts from individuals showing signs of depression are characterized by a higher frequency of negative words (e.g., "bad," "hate," "sick," "death"), aligning with existing literature on cognitive biases in depression. Conversely, users without overt signs of depression in their posts might use more sexual or profane terms.
  • Self-Deprecation and Negative Self-Evaluation: The tendency for depressed individuals to use more self-deprecating language and less deprecation of others is consistent with cognitive theories suggesting depressed individuals often have a negative evaluation of themselves.

Key Results

  • RQ1: Users with self-reported depression diagnosis are more likely to use offensive language than control groups. This trend is even more pronounced in individuals showing signs of depression.
  • RQ2:
    • Posts from individuals showing signs of depression are less likely to contain targeted offensive language (insults) compared to untargeted profanity.
  • A higher percentage of offensive posts from individuals showing signs of depression contained negative polarity words (e.g., "bad," "hate," "sick," "suffer").
  • In contrast, posts from users without signs of depression tended to include more sexual-related content.
  • The offensive language detection model achieved 0.85 Precision, 0.74 Recall, and 0.77 F1 score on the OLID test data (Level A).
  • The offensive language categorization model achieved 0.78 Precision, 0.84 Recall, and 0.80 F1 score (Level B).

Limitations and Future Work

Limitations: * The study acknowledges potential biases in model predictions and analysis arising from the annotation and sampling techniques of the used datasets. * The authors did not use any demographic information in their models or analysis.

Future Work: * Investigate the interplay between offensive language and sentiment polarity in mental health datasets. * Analyze the targets of offensive posts using the OLID Level C annotation. * Conduct a more detailed analysis of the function of profanity and vulgarity in these datasets. * Extend the analysis to other languages, leveraging existing datasets and cross-lingual embedding models.

Summary updated

BibTeX
@inproceedings{bucur2021exploratory,
  author = {Ana-Maria Bucur and Marcos Zampieri and Liviu P. Dinu},
  title = {{An Exploratory Analysis of the Relation between Offensive Language and Mental Health}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure