Using topic-noise models to generate domain-specific topics across data sources.

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Science

Background:

• Domain-specific document collections are increasingly common, sourced from diverse origins like social media and traditional texts.
• Existing topic models struggle with varying noise levels across different document types, limiting their effectiveness.
• Analyzing multi-source data requires methods to identify core topics consistently.

Purpose of the Study:

• To develop a novel topic modeling approach that handles varying document qualities and sources.
• To improve the accuracy, flexibility, and coherence of topic representations.
• To enable effective topic discovery across multiple, heterogeneous data sources.

Main Methods:

• Proposed topic-noise models that jointly model topic and noise distributions.
• Introduced Topic Noise Discriminator (TND) using word embedding spaces to approximate topic and noise.
• Developed cross-source topic blending (CSTB) by mapping topic sets to an s-partite graph to identify core topics.

Main Results:

• Demonstrated TND's ability to improve Latent Dirichlet Allocation (LDA) topic quality in noisy collections.
• Showcased TND's effectiveness in generating more coherent and diverse topic sets when ensembled.
• Empirically validated topic-noise models and CSTB on large, real-world datasets from multiple domains.

Conclusions:

• Topic-noise models offer a flexible and accurate solution for topic modeling across diverse document qualities.
• TND enhances existing topic models by explicitly addressing noise, improving topic coherence and diversity.
• CSTB provides a robust method for identifying core topics from multi-source data, facilitating cross-domain analysis.