Detecting and Interpreting Sarcasm for Autistic Teens: Design and Evaluation of a Multi-Model BERT-Based Assistant by Ananya Surana

Detecting and Interpreting Sarcasm for Autistic Teens: Design and Evaluation of a Multi-Model BERT-Based Assistant

Hackathon Project: SubText

https://m3qhyucffql4psn7h5wezx.streamlit.app/

By Ananya Surana (Student at The Cathedral and John Connon School), 1st August 2026.

Reviewed by Orin Stephney (Principal, Lancers International School).

1. Introduction

In recent years, digital communication has become the predominant form of communication between adolescents, wherein much of the meaning in text messages, tweets and DMs is implicit rather than explicit. Tone, sarcasm, flirtation, irony, humour and intent thus must be inferred from contextual clues rather than the dictionary meaning of words. This poses a challenge for neurodivergent teenagers (teens), like those with autism spectrum disorder and other social communication differences, who may have trouble interpreting these nuanced signals, which contributes to misunderstandings, social anxiety, and reduced confidence in online interactions.

Despite recent breakthroughs in sentiment analysis and emotion detection, most existing systems are geared toward identifying the emotion contained in a message rather than explaining what that emotion implies in a real-world interaction. A communication can be labelled as negative, sardonic, or frustrated, but the user is still left to determine the speaker's true intent. This constraint is especially important for neurodivergent people, who may struggle to interpret meaning from indirect language, sarcasm, implied requests, or subtle tone shifts. These issues can be exacerbated in text-based communication, which lacks facial expressions, body language, and verbal clues.

This study introduces SubText, a communication-support tool that helps users grasp the meaning of texts rather than merely labelling their sentiment. The technology recognizes implicit intent and emotional tone in conversations using a refined BERT model and converts these indications into understandable explanations. Additionally, it offers response options based on the interaction's context. The research investigates how natural language processing can be used to both analyze and facilitate talks by fusing contextual language understanding with useful communication support. It is anticipated that this method will outperform traditional sentiment-analysis systems in identifying subtle social signs and offering more significant assistance.

Given the highly contextual nature of sarcasm, implicit emotion, and social intent in digital conversations, transformer-based architectures were selected as the primary modelling approach. Specifically, BERT was chosen due to its bidirectional contextual encoding capabilities, which enable more effective interpretation of nuanced linguistic patterns than lexicon-based, conventional machine-learning, or recurrent neural-network approaches.

To Read the full paper, please click here -