Learning Supervised Topic Models for Classification and Regression from Crowds.

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Science

Background:

• Large-scale document analysis necessitates advanced topic modeling techniques.
• Supervised topic models are valuable but challenged by noisy, multi-annotator data common in real-world applications.
• The single-annotator assumption is often impractical due to annotation ambiguity and volume.

Purpose of the Study:

• To propose novel supervised topic models for classification and regression tasks.
• To address the challenges of annotator heterogeneity and bias in crowdsourced data.
• To develop scalable algorithms for analyzing large document datasets.

Main Methods:

• Development of two supervised topic models: one for classification and one for regression.
• Implementation of an efficient stochastic variational inference algorithm.
• Empirical evaluation against state-of-the-art supervised topic models.

Main Results:

• The proposed models effectively account for annotator heterogeneity and biases.
• The stochastic variational inference algorithm demonstrates scalability to large datasets.
• Empirical results show superior performance compared to existing approaches.

Conclusions:

• The developed supervised topic models offer a robust solution for analyzing real-world, crowdsourced document data.
• The models provide a practical approach to handling noisy annotations and diverse annotator characteristics.
• This work advances supervised topic modeling for large-scale classification and regression problems.