Visualizing Temporal Topic Embeddings with a Compass.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Dynamic topic modeling (DTM) reveals how latent topics change over time.
• Current DTM methods separate document and word representations, hindering temporal analysis.
• This separation prevents a unified embedding space for analyzing word and document evolution.

Purpose of the Study:

• To propose an enhanced dynamic topic modeling approach by integrating temporal word embeddings.
• To enable direct comparison of word and document embeddings across time within dynamic topics.
• To facilitate visualizations combining temporal word embeddings and document contexts for topic analysis.

Main Methods:

• Expanding the compass-aligned temporal Word2Vec methodology for dynamic topic modeling.
• Developing a unified embedding space for words and documents across time.
• Creating novel visualizations for temporal word embeddings within topic structures.

Main Results:

• The proposed method achieves competitive performance in topic relevancy and diversity.
• Demonstrates effectiveness across temporal datasets of varying sizes.
• Provides insightful visualizations of temporal word embeddings and global topic evolution.

Conclusions:

• The new method offers a unified framework for analyzing temporal topic dynamics.
• Enables deeper understanding of how word usage and topics evolve concurrently.
• Advances the field of dynamic topic modeling with integrated temporal embeddings and visualizations.