Modeling Latent Dimensions of Human Beliefs

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Cognitive Science

Background:

• Human perception and beliefs shape societal interactions.
• Traditional topic modeling relies on word lists, lacking contextual nuance.
• Exploring large-scale belief systems often requires extensive manual effort.

Purpose of the Study:

• Introduce LaBel (Latent Beliefs Model) as an alternative to topic modeling.
• Automate the exploration of latent semantic dimensions and human beliefs.
• Represent discovered dimensions as generated phrases for enhanced interpretability.

Main Methods:

• Utilize transformer-based embeddings to capture semantic information.
• Employ a novel modification of autoregressive transformers for text generation.
• Condition text generation on vector input formats.

Main Results:

• LaBel generates contextual phrases (e.g., "the world is truly in your favor") instead of word lists.
• Achieved higher accuracies in belief classification tasks compared to traditional topic modeling.
• Demonstrated effectiveness in an intrusion task for evaluating dimension quality.

Conclusions:

• LaBel offers a more natural and context-aware method for uncovering latent beliefs.
• The model automates belief exploration, reducing reliance on human experts for large datasets.
• LaBel enhances semantic representation and classification accuracy in belief analysis.