Semantic Fields + Gradient Flow

in Word Embedding Space

Final Presentation

GloVe 6B 300D · UMAP 3D · Unity Passthrough AR · User Study

ML Embeddings

in Physical Space

Comparative Study

Project Pivot

CFD Fluid Flow Visualization → Semantic fields and gradient flow

CFD Fluid Flow Visualization in VR

1

ParaView export + Unity integration issues

2

Streamlines inconsistent / hard to render

3

Performance + stability problems in VR

Semantic Fields and Gradient Flow

Reused vector field + streamline pipeline

More stable + easier to implement

Better evaluation design + clearer user study tasks

1

2

3

Project Overview

Continuous representation of word embeddings

Representation

• Scalar field → category density
• Vector field → gradient (semantic direction)

Visualization

• Colored density regions
• Flow lines (streamlines)
• Probe → % category composition

The Study

Controlled study comparing 2D plots vs. VR on cluster ID, similarity judgment, and spatial reasoning tasks

Key Shift from Project 1

Project 1: Discrete points → clusters

Project 2: Continuous field → transitions + boundaries

Wiki Contributions

Analysis, evaluation, and technical documentation

Semantic Field Visualization Tutorial

Step-by-step guide:

Kernel Density Estimations → scalar field → gradient → Unity rendering pipeline

Comparison: 2D vs VR Point Cloud vs VR Field Visualization

Quantitative + qualitative comparison

Cluster clarity, boundaries, flow understanding, user preference

Demo Videos

Wiki page with 2D Baseline Demo and VR/AR Demo

Scientific Visualization Connection

Applying CFD-style techniques

(streamlines, scalar fields) → word embedding space

Code & Technical Implementation

Python scripts + Unity C# pipeline

(field construction, gradients, rendering, probe interaction)

User Study + Evaluation Results

Google Form design, responses, and analysis� Includes graphs, pie charts, and qualitative insights

2D Baseline — PCA Word Embedding Plots

Contained Fields as contour lines and has vector flow streamlines

Unity 3D Semantic Fields and Gradient Lines

Survey Results: Flow Direction

Do the flow arrows tend to point towards a region or no (are there any places where many flow arrows point to)?

2D Projection

Unity 3D VR Environment

• Flow arrows consistently perceived as pointing toward cluster centers
• Stronger sense of “global directionality” toward category cores
• Some ambiguity near boundaries, but overall consistent structure

• Direction still generally toward clusters, but less uniform
• More scattered / locally inconsistent arrow perception
• Some users saw spirals, divergence, or “random” flow in regions

Takeaway

• 2D → clearer global direction
• VR → richer but more complex / less uniformly interpreted flow

Survey Results: Boundary Perception

Do the transitions/boundaries between category regions look gradual or sharp? Describe what you notice.

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

True answer: Nurse had a higher cosine similarity to doctor than engineer did based on their vectorizations.

0.5860 vs. 0.3058

2D Projection

Unity 3D VR Environment

• Mixed perception: both sharp and gradual reported
• Sharpness often tied to visible cluster separation
• Boundaries sometimes felt “artificially crisp”

• More consistent perception of gradual transitions
• Strong overlap at edges between categories
• Blending between emotional / moral / profession regions frequently noted

Takeaway

• 2D → emphasizes separation
• VR → emphasizes continuity and overlap

Survey Results: Cluster Self-Containment

Which category looks most self-contained — its contour rings are tight and don't overlap much with others?

[Insert graph here]

2D Projection

VR Projection

Survey Results: Region Mixing

Which region looks most mixed — where contour rings from multiple categories overlap?

[Insert graph here]

[Insert graph here]

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

True answer: Nurse had a higher cosine similarity to doctor than engineer did based on their vectorizations.

0.5860 vs. 0.3058

2D Projection

Unity 3D VR Environment

• Most mixed:
• Emotions + Moral Concepts (often with Professions)
• Central overlap region seen as “dense intersection”

• Same core finding, but stronger:
• Middle region heavily blended across all categories
• Overlap more spatially visible and continuous

Takeaway

• Both agree on mixed region
• VR makes overlap structure more visually explicit

Survey Results: Cluster Clarity

How clear are the clusters in the visualizations?

[Insert graph here]

2D Projection

VR Projection

Survey Results: Mystery Dots

"Dot A, B, C — which category?"

[Insert graph here]

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

[Insert graph here]

Dot A: Guilt → Emotions Dot B: Soldier → Professions Dot C: Power → Moral Concepts

VR Projection

2D Projection

Survey Results: Mystery Dot Confidence

How confident are you in your mystery dot guesses?

[Insert graph here]

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

[Insert graph here]

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

[Insert graph here]

2D Projection

VR Projection

Survey Results: Patterns

Describe any patterns or relationships you noticed.

[Insert graph here]

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

[Insert graph here]

Noisy 2D Projection

Best 2D Projection

Unity 3D VR Environment

True answer: Nurse had a higher cosine similarity to doctor than engineer did based on their vectorizations.

0.5860 vs. 0.3058

2D Projection

Unity 3D VR Environment

• Clear categorical grouping
• Emotions + moral concepts often close
• Nature consistently separated
• Professions moderately distinct

• Clusters appear closer together overall
• Boundaries less rigid than expected
• More nuanced internal structure within clusters
• Some categories interweave at edges

Takeaway

• 2D → clean separable clusters
• VR → reveals internal continuity + inter-category blending

Comparing the Visualizations

[Insert graph here]

Comparing the Visualizations

[Insert graph here]

Open Feedback (Condensed Summary)

Flow Lines

Helpful for understanding direction toward categories

But:

• Meaning not always intuitive
• Some confusion about what gradients represent
• Sparse or inconsistent visibility in VR

Confusing / Hard to Interpret

• “What exactly do flow lines represent?” (common issue)
• VR sparsity made structure harder to read at times
• Percentage / composition UI sometimes unclear or hard to interact with
• “Not sure what flow means semantically” repeated across users

Did VR Improve Understanding?

• Yes for spatial intuition and relationships
• Mixed for clarity of categories
• Helped users:
• See overlap regions better
• Understand local neighborhood structure
• But:
• Harder to interpret globally than 2D

Other Comments (Common Themes)

• VR felt more immersive but more complex
• 2D felt simpler and easier to interpret quickly
• Flow + probe interaction seen as interesting but not fully self-explanatory
• Overall: “useful but needs clearer explanation of semantics”

Takeaways & Challenges

Key Takeaways

Challenges

VR reveals semantic continuity, not just clustering

→ Users consistently saw smoother transitions and overlap regions that are hidden in 2D.

Interpreting flow semantics was difficult for users

→ “What do arrows actually mean?” was a recurring confusion.

Flow fields add directional meaning, but are not self-explanatory

→ Many users interpreted “flow” inconsistently without guidance.

Visual complexity vs interpretability tradeoff in VR

UMAP coordinates needed manual rescaling to map to room-scale (0.5–2m spread).

Different visualizations support different kinds of understanding

→ 2D = clarity & separability, VR = spatial intuition & structure.

Designing fair comparisons between 2D and VR tasks

→ Same concept (clusters) feels different across representations, making evaluation non-trivial.

Thank You

Questions?