Visual Exploration of Neural Document Embedding in Information Retrieval: Semantics and Feature Selection

Area of Science:

• Computational Linguistics
• Information Science
• Data Visualization

Background:

• Neural document embeddings convert variable-length texts into semantic vector representations, benefiting applications like information retrieval (IR).
• The 'black-box' nature of these embeddings hinders understanding of semantic encoding and utilization.
• Biomedical IR in healthcare decision-making highlights the need for interpretable neural document embeddings.

Purpose of the Study:

• To develop a visual analytics system for exploring neural document embeddings.
• To gain insights into the underlying embedding space and its semantic encoding.
• To promote the effective utilization of neural document embeddings in downstream applications, particularly IR.

Main Methods:

• Developed a visual analytics system to visualize neural document embeddings as a configurable document map.
• Enabled exploration of the embedding space to identify salient neural dimensions (semantic features).
• Supported feature selection with instant visual feedback for improved IR performance.

Main Results:

• Demonstrated the usefulness and effectiveness of the visual analytics system.
• Identified inspiring findings through use cases, showcasing improved IR performance.
• Facilitated guidance, reasoning, and semantic analysis of neural document embeddings.

Conclusions:

• The visual analytics system enhances understanding and confidence in neural document embeddings.
• It empowers designers and developers to achieve more favorable performance in application domains.
• This approach promotes the effective application of neural document embeddings in IR and beyond.