TextEssence: A Tool for Interactive Analysis of Semantic Shifts Between Corpora

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Data Visualization

Background:

• Word and concept embeddings capture language regularities but are underutilized for corpus analysis.
• Comparative analysis of corpora using embeddings is challenging.

Purpose of the Study:

• Introduce TextEssence, an interactive system for comparative corpus analysis using embeddings.
• Propose a novel measure for embedding confidence to identify high-quality embeddings.

Main Methods:

• Developed a web-based interface for TextEssence.
• Implemented visual, neighbor-based, and similarity-based embedding analysis modes.
• Introduced a new embedding confidence measure based on nearest neighborhood overlap.

Main Results:

• TextEssence provides an interactive platform for comparative corpus analysis.
• The proposed confidence measure aids in selecting reliable embeddings.
• A case study on COVID-19 literature demonstrates the system's utility.

Conclusions:

• TextEssence facilitates the study of corpus characteristics and relationships through embeddings.
• The system enhances the application of embeddings in corpus linguistics and NLP.
• TextEssence is accessible online for broader use and research.