2026 Shared Task Description

Basic Task Description

Background

Sentiment classification is one of the most popular NLP tasks and a textbook example of sequence classification over text. Sarcasm classification is a closely related task: sarcasm often inverts surface sentiment, making the two tasks jointly informative and jointly challenging. However, the performance of large language models on both tasks varies considerably across language varieties, including national varieties such as Australian and British English. This shared task focuses on sentiment and sarcasm classification for text in Australian (en-AU) and British (en-UK) English, using the en-AU and en-UK subsets of BESSTIE.

Goal

The goal is to build systems that classify both the sentiment and the sarcasm of a given text, while remaining robust across the two English varieties. Participants are provided with labelled examples containing:

• Data source (source): Google Places reviews (google) and Reddit posts/comments (reddit).
• Variety (variety): en-AU or en-UK.
• Text (text)
• Sentiment Label (sentiment): negative (0), positive (1).
• Sarcasm Label (sarcasm): not sarcastic (0), sarcastic (1).

Data Files and Submission

We will use CodaBench for this year's competition (ALTA Shared Task 2026). The details about data formats, submission, and evaluation metrics will be provided in the competition website.

Important Dates