It is no longer any secret that people worldwide are struggling with their mental health, in terms of diagnostic disorders as well as non-diagnostic measures like perceived stress. Barriers to receiving professional mental healthcare are significant, and even in locations where the availability of such care is increasing, our infrastructures are not equipped to find people the support they need. Meanwhile, in a highly-connected digital world, many people turn to outlets like social media to express themselves and their struggles and interact with like-minded others.
This setting---where human experts are overwhelmed and human patients are acutely in need---is one in which we believe artificial intelligence (AI) and natural language processing (NLP) systems have great potential to do good. At the same time, we must acknowledge the limitations of our models and strive to deploy them responsibly alongside human experts, such that their logic and mistakes are transparent. We argue that models that make and explain their predictions in ways guided by domain-specific research will be more understandable to humans, who can benefit from the models' statistical knowledge but use their own judgment to mitigate the models' mistakes.
In this thesis, we leverage domain expertise in the form of psychology research to develop models for two categories of emotional tasks: identifying emotional reactions in text and explaining the causes of emotional reactions. The first half of the thesis covers our work on detecting emotional reactions, where we focus on a particular, understudied type of emotional reaction: psychological distress. We present our original dataset, Dreaddit, gathered for this problem from the social media website Reddit, as well as some baseline analysis and benchmarking that shows psychological distress detection is a challenging problem. Drawing on literature that connects particular emotions to the experience of distress, we then develop several multitask models that incorporate basic emotion detection, and quantitatively change the way our distress models make their predictions to make them more readily understandable.
Then, the second half of the thesis expands our scope to consider not only the emotional reaction being experienced, but also its cause. We treat this cause identification problem first as a span extraction problem in news headlines, where we employ multitask learning (jointly with basic emotion classification) and commonsense reasoning; and then as a free-form generation task in response to a long-form Reddit post, where we leverage the capabilities of large language models (LLMs) and their distilled student models. Here, as well, multitask learning with basic emotion detection is beneficial to cause identification in both settings.
Our contributions in this thesis are fourfold. First, we produce a dataset for psychological distress detection, as well as emotion-infused models that incorporate emotion detection for this task. Second, we present multitask and commonsense-infused models for joint emotion detection and emotion cause extraction, showing increased performance on both tasks. Third, we produce a dataset for the new problem of emotion-focused explanation, as well as characterization of the abilities of distilled generation models for this problem. Finally, we take an overarching approach to these problems inspired by psychology theory that incorporates expert knowledge into our models where possible, enhancing explainability and performance.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/fxxx-wx64 |
| Date | January 2024 |
| Creators | Turcan, Elsbeth |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.002 seconds