Investigations on Text-level Psychological Stress Detection Approaches

Abstract

Psychological stress has engulfed the world and has turned out to be a major propellant for both physiological health disorders and suicides across the globe. The more concerning issue is the prevalence of psychological stress among youth and the increasing vulnerability of the young generation to stress. Hence, the detection of stress, before it becomes chronic, is of paramount importance. There exist many traditional stress detection mechanisms like interviews with psychiatrists, question naires, etc. However, the social stigma attached to these methods, makes people either avoid them or give incomplete or misleading information. Furthermore, the methods to detect stress using electronic devices and sensors are seen as invasive to daily life. Hence, this leads to searching for alternative approaches to detect stress early by capturing genuine emotions without any stigma. Social media data, with the growing popularity of social media usage, can be an ideal candidate in such a scenario. In micro-blogging sites like Twitter (now called X), people participate in a large scale to express their opinions, and daily activities in a free manner devoid of any social stigma. This makes Twitter a very lucrative resource for capturing human emotions and therefore, social media-based text-level stress detection has caught the attention of the researchers. The initial approaches using social media to detect text-level stress have focused on crowd-sourcing of the data for collection and labeling. Nevertheless, the crowd-sourcing approach is prone to a problem similar to that of questionnaires, while the manual labeling is laborious, time-consuming, and vulnerable to errors. To address this, automatic labeling based on sentence patterns of “I feel” was proved to be effective in labeling the social media data for stress detection. Using this strat egy to collect tweets, many works in the literature had proposed tweet-level stress detection methods. These techniques have also developed more detailed hand-crafted features from the text data of tweets. Furthermore, deep-learning-based methods like stacked cross autoencoders for the classification of tweet-level stress, apart from traditional machine learning techniques, were also developed. The literature of text-level stress detection problem has many gaps despite provid iii ing the initial solutions. First, there is an issue of data sparsity; most of the tweets are short in size, and hence existing approaches do not utilize text content of the data, content of previous or neighborhood tweets, and clues from the text data about sarcasm. Second, the lack of a large amount of labeled data for stress detection using social media makes it difficult to implement supervised algorithms. This gives moti vation to explore solutions based on semi-supervised learning methods which require a lesser amount of labeled data. Third, the concept of sarcasm is found to be one of the useful attributes in detecting tweet-level stress, and with sarcasm being a classifi cation problem on its own, there is a scope to study and develop multi-task learning approach to detect stress with the help of the auxiliary task of sarcasm detection. In this thesis, the issues mentioned are addressed using various machine learning-based solutions. Initially, a new approach called neighborhood-based tweet-level stress de tection(NTSD), a modified logistic regression that includes neighborhood tweets, is developed to utilize the text content of the tweets for detecting stress. In addition, a new attribute called Sarcasm_Level is proposed which captures the sarcasm present in the text content of the tweet. These solutions help to address both the data sparsity problem and the utilization of the text by capturing the related concept of sarcasm. However, it has an overhead due to the additional requirement of the neighborhood tweets. To overcome this, later, a new method called sarcasm-based tweet-level stress detection (STSD) was developed, wherein sarcasm is used to develop the modified logistic regression such that the loss of sarcastic tweets, which reflect positive and friendly moments, is penalized. Simultaneously, the loss of the non-sarcastic tweets is minimized. This makes the proposed STSD perform better without the requirement of extra data like neighborhood tweets. In the real world, most of the data is unlabeled in many scenarios, and on the contrary, most of the approaches for tweet-level stress detection are supervised models. To address this issue, a new semi-supervised approach based on logistic regression, called semi-supervised method for tweet-level stress detection (SMTSD) is developed. This utilizes semi-labeled data and the concept of sarcasm in computing the pseudo labels, thereby improving the performance of tweet-level stress detection. Though iv many approaches consider sarcasm as a useful attribute in the detection of stress at tweet-level, sarcasm detection is a classification task in itself and is related to emotion and stress detection. To this end, a multi-task approach for tweet-level stress detection (MATSD) is proposed where deep neural networks are utilized for the prediction of stress as the primary task using long short term memory (LSTM) and sarcasm as an auxiliary task using convolutional neural network (CNN), sharing a joint map layer. The simultaneous prediction of the tasks act as regularization and helps in the betterment of the efficacy of tweet-level stress prediction. To conclude, from this research, we propose different solutions for text-level stress detection by maximizing the utilization of text data.