Sarcasm Detection in Autism Forum Data (Oct 2023)

The overall aim of this project is to learn more about the ways autistic people understand, interpret, and use sarcasm in online communities. Many autistic people find it hard to detect sarcasm in speech, which may lead to misinterpretation, stress, and difficulties with social interactions. A playwright-based web scraper that gathers high volumes of authentic forum interaction allows researchers to portray how sarcasm naturally occurs in real autistic communication, rather than in artificially simulated or lab-generated samples. With such data the research aims to understand not only where confusion occurs, but also where autistic people do successfully use sarcasm, humor, or irony.

Upon scraping data, we fed this to BART and other transformer-based models to identify parts of the text likely containing sarcasm so that manual reviewing can be reduced, resulting in more efficient large-scale analysis. Understanding which models get it right and which do not helps us understand the nature of machine comprehension as different from human comprehension when it comes to detecting sarcasm. These contrasts show where AI can fail when it comes to tone recognition and indicate dangers in automated moderation or dialogue systems.

More generally at a societal level, the aim of this approach is to make online environments more comfortable, safe, and intuitive for autistic users. By looking at communication patterns on a large scale, we can help inform tools for making moderation systems more efficient, reduce the amount of miscommunication that happens online, and help individuals with autism feel included instead of misunderstood. This project represents a step forward in technology that listens attentively, adapts more thoughtfully, and recognizes the diverse ways that people communicate.