LEIA: Linguistic Embeddings for the Identification of Affect.

Area of Science:

• Natural Language Processing
• Computational Social Science
• Affective Computing

Background:

• Social media generates vast text data, enabling emotion analysis via language models.
• Current emotion identification models face limitations due to small, costly, and noisy training datasets.
• Existing methods struggle with data size limitations and label noise in model development.

Purpose of the Study:

• To introduce LEIA, a novel language model for emotion identification in text.
• To leverage a large-scale dataset of self-annotated emotion labels for enhanced model training.
• To improve the accuracy and generalizability of emotion detection in social media content.

Main Methods:

• Developed LEIA, a language model trained on over 6 million social media posts with self-annotated emotion labels (happiness, affection, sadness, anger, fear).
• Utilized a word masking pre-training technique to improve the learning of emotion-specific words.
• Conducted in-domain and out-of-domain evaluations on multiple datasets to assess generalization capabilities.

Main Results:

• LEIA achieved a macro-F1 score of approximately 73 on in-domain tests, surpassing other supervised and unsupervised methods.
• Demonstrated robust performance across various social media platforms, user demographics, and time periods.
• Showcased generalization of anger, happiness, and sadness classification beyond the training data domain.

Conclusions:

• LEIA offers a significant advancement in emotion identification accuracy and robustness in text.
• The model's ability to generalize across diverse datasets highlights its practical applicability.
• LEIA provides a powerful tool for future research in understanding writer-centric emotions in text data.