Paper Number
1792
Paper Type
Short Paper
Abstract
Labelled data scarcity is a substantial limitation for researchers in many fields. Without labelled data, techniques such as text classification have proven to be challenging. This paper introduces an innovative approach to mitigate this issue by leveraging the emotional intelligence in Generative AI for auto-labelling. This novel methodology leverages models trained on reference datasets with similar feature representations to auto-label target datasets through n-shot learning. Focusing on emotion detection from social media text, our preliminary results demonstrate a promising trend; our model trained on GoEmotions and auto-labelling the Twitter Emotion Corpus scored an F1-Macro of 0.371, outperforming GPT-3.5-turbo. Our research applies this approach to five other social media emotion detection corpora. The research outputs are expected to contribute to easier labelling of datasets for emotion detection, with the methodology transferable to other domains lacking labelled data.
Recommended Citation
Kok-Shun, Brice Valentin; Chan, Johnny; Peko, Gabrielle; and Sundaram, David, "Generative Reference-Shot Learning for Emotions: Auto-Labelling Datasets by Leveraging Generative AI and Existing Corpora" (2024). ECIS 2024 Proceedings. 18.
https://aisel.aisnet.org/ecis2024/track03_ai/track03_ai/18
Generative Reference-Shot Learning for Emotions: Auto-Labelling Datasets by Leveraging Generative AI and Existing Corpora
Labelled data scarcity is a substantial limitation for researchers in many fields. Without labelled data, techniques such as text classification have proven to be challenging. This paper introduces an innovative approach to mitigate this issue by leveraging the emotional intelligence in Generative AI for auto-labelling. This novel methodology leverages models trained on reference datasets with similar feature representations to auto-label target datasets through n-shot learning. Focusing on emotion detection from social media text, our preliminary results demonstrate a promising trend; our model trained on GoEmotions and auto-labelling the Twitter Emotion Corpus scored an F1-Macro of 0.371, outperforming GPT-3.5-turbo. Our research applies this approach to five other social media emotion detection corpora. The research outputs are expected to contribute to easier labelling of datasets for emotion detection, with the methodology transferable to other domains lacking labelled data.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.