VADIS: A Visual Analytics Pipeline for Dynamic Document Representation and Information-Seeking

Rui Qiu, Yamei Tu, Po-Yin Yen, Han-Wei Shen

Introduction

In the rapidly expanding field of biomedical research, clinicians and researchers often find themselves overwhelmed by the sheer volume of available literature. Navigating thousands of papers to find specific clinical insights is a significant challenge. Traditional document visualization tools, while helpful, typically rely on "static" embeddings—mathematical representations of text that remain fixed regardless of the user's specific information needs. This rigidity often results in visualizations that group documents by broad structural similarities rather than specific relevance to a query.

In their paper, "VADIS: A Visual Analytics Pipeline for Dynamic Document Representation and Information-Seeking," Rui Qiu and colleagues present a novel system designed to overcome these limitations. VADIS (Visual Analytics pipeline for Dynamic Document Representation and Information-Seeking) introduces a "steerable" visualization framework. By utilizing natural language prompts, users can dynamically reorganize the document map, prioritizing relevance and semantic similarity in a way that aligns with their immediate research goals.

The Limitations of Static Visualization

Existing document exploration tools generally suffer from three core issues:

  1. Static Representations: Most tools generate a document map once using pre-trained models. If a researcher is looking for "treatments for ADHD," the map might still cluster papers by "study population" (e.g., children vs. adults) because the underlying model was not tuned to the specific nuances of the query.
  2. Relevance vs. Similarity: Standard scatterplots are effective at showing how similar documents are to one another, but they rarely illustrate how relevant those documents are to the user's specific question.
  3. The "Black Box" Problem: It is often difficult for users to understand why a document has been highlighted or positioned in a certain way, leading to a lack of trust in the search results.

The VADIS Pipeline

The VADIS system addresses these hurdles through a three-component architecture designed for interactivity and interpretability.

1. Prompt-based Attention Model (PAM)

At the heart of VADIS is the Prompt-based Attention Model (PAM). Unlike standard models that use a generic [CLS] token to represent an entire document, PAM employs a dual-encoder architecture based on BioBERT and a custom attention layer. When a user enters a prompt, the model calculates attention scores between the prompt and every token in the corpus.

  • Dynamic Embeddings: The document’s vector is calculated as a weighted sum of its tokens based on these attention scores, meaning the representation of the document literally changes based on the query.
  • Relevance Scores: The sum of these attention scores provides a direct, quantitative measure of how well a document matches the user's prompt.

2. Relevance-Preserving Document Map

VADIS replaces the traditional, often cluttered scatterplot with a circular grid layout. This layout uses a "target" or "bullseye" metaphor to organize information:

  • The Center: The center of the circle represents the highest relevance. The closer a document is to the center, the higher its relevance score to the current prompt.
  • Neighborhoods: Spatial proximity on the grid reflects semantic similarity. Documents that are similar to each other are placed in the same "neighborhood."
  • The Algorithm: The authors adapted the Self-Organizing Map (SOM) algorithm to handle the dual constraints of relevance (radial distance) and similarity (angular/spatial distance) on a discrete grid, preventing the overlapping issues common in traditional dimensionality reduction techniques.

3. Corpus-level Attention Visualization

To solve the "black box" problem, VADIS employs Non-negative Matrix Factorization (NMF) on the attention scores across the entire corpus. This process extracts "topics of interest," providing the user with a summary of which keywords the model is prioritizing. This allows researchers to see the specific terms driving the model's relevance rankings.

Evaluation and Key Results

The authors evaluated VADIS using several major Question-Answering (QA) benchmarks, including SQuAD, TriviaQA, and NaturalQA, alongside the specialized emrQA dataset (derived from clinical notes) and a real-world ADHD corpus of ~2,100 abstracts.

Retrieval and Clustering Performance

  • Accuracy: PAM demonstrated superior retrieval accuracy compared to traditional baselines. On the emrQA dataset, it achieved a 70.1% Top-10 accuracy, significantly outperforming the 54.2% achieved by BM25.
  • Clustering Quality: Using the Adjusted Rand Index (ARI), the authors found that PAM’s dynamic embeddings clustered documents much more effectively according to user interests (ARI of 0.71) compared to static BioBERT embeddings (ARI of 0.26).

Qualitative Case Study

In a practical case study, a clinical expert used VADIS to explore the ADHD corpus. Through iterative refinement—searching for "treatments," then weighting "side effects" more heavily—the expert was able to narrow down thousands of papers to specific discussions on drug treatments (like Methylphenidate) and their associated side effects (like headaches) within a single session.

Discussion and Future Work

While VADIS represents a significant advancement, the authors note several areas for improvement. There is an inherent trade-off between preserving relevance and similarity; if these two constraints conflict, the grid may develop irregular empty spaces. Additionally, the current system requires users to manually adjust weights for similarity versus relevance, a process the authors hope to automate in future iterations. Finally, like many attention-based models, PAM can be sensitive to "noisy" prompts that are entirely unrelated to the corpus.

Conclusion

VADIS moves beyond the paradigm of static document maps, offering a dynamic, "human-in-the-loop" approach to information seeking. By treating attention as the bridge between the user's intent and the visual layout, VADIS transforms the search process into an interpretable, exploratory conversation with data.

Summary updated

BibTeX
@article{qiu2025vadis,
  author = {Rui Qiu and Yamei Tu and Po-Yin Yen and Han-Wei Shen},
  title = {{VADIS: A Visual Analytics Pipeline for Dynamic Document Representation and Information-Seeking}},
  journal = {IEEE Trans. Vis. Comput. Graph.},
  year = {2025},
  doi = {10.1109/tvcg.2024.3456339},
}
Figure