Autoregressive Affective Language Forecasting: A Self-Supervised Task.

Area of Science:

• Computational Linguistics
• Affective Computing
• Time Series Analysis

Background:

• Human emotions evolve over time, and language use is strongly linked to emotional states.
• Modeling emotional language dynamics over time remains an underexplored area with significant potential applications.

Purpose of the Study:

• Introduce and define the task of affective language forecasting: predicting future language changes based on historical language patterns.
• Explore autoregressive characteristics of affective language forecasting, including history size, length variability, and time-step resolution.
• Develop and evaluate sequence models for predicting language-based emotions over time.

Main Methods:

• Established fundamental autoregressive properties relevant to affective language forecasting.
• Adapted popular sequence models originally designed for word sequences to model sequences of language-based emotions.
• Developed a novel dual-sequence Gated Recurrent Unit (GRU) model with decayed hidden states.

Main Results:

• A novel dual-sequence GRU model with decayed hidden states demonstrated superior performance in affective language forecasting.
• Achieved a correlation coefficient (r) of .66 in predicting language-based emotions over time.
• The study utilized a novel Twitter dataset comprising 1,900 users with weekly and daily emotion scores.

Conclusions:

• Affective language forecasting is a viable task with practical applications in mental health and consumer trend prediction.
• The proposed dual-sequence GRU model offers an effective approach for modeling temporal dynamics of language-based emotions.
• The release of the dataset and code aims to foster further research in this domain.