Modern machine learning relies heavily on distributed representations. Whether dealing with word vectors, large language model (LLM) hidden states, or protein sequence embeddings, concepts are mapped into dense, high-dimensional vector spaces. While these spaces capture rich semantic relationships, they are fundamentally uninterpretable to humans. In a 768-dimensional space, we cannot easily see where clusters form, how concepts overlap, or where a model's understanding breaks down.
To bridge this gap, researchers use semantic maps—visual representations that translate abstract mathematical distances into spatial proximity. Early work by Qing Ma et al. in Self-Organizing Chinese and Japanese Semantic Maps demonstrated that spatial layouts offer a far more intuitive grasp of vocabulary structures than traditional flat clustering. The core intuition is simple: if two concepts are semantically related, they should be geometrically close in a visual space.
The Mechanics of Projection
How do we compress hundreds of dimensions into two or three without losing the underlying structure? The answer lies in non-linear dimensionality reduction techniques like $t$-Distributed Stochastic Neighbor Embedding ($t$-SNE), Uniform Manifold Approximation and Projection (UMAP), and Multidimensional Scaling (MDS).
These algorithms operate by preserving topologies. For instance, $t$-SNE converts Euclidean distances between high-dimensional data points into conditional probabilities that represent similarities. It then minimizes the divergence between these probabilities in the high-dimensional and low-dimensional spaces. This approach excels at preserving local structure, ensuring that close neighbors in the embedding space remain close in the 2D projection.
This local preservation makes $t$-SNE highly effective for mapping linguistic and behavioral nuances. For example, Siobhán Grayson et al. used $t$-SNE in Novel2Vec: Characterising 19th Century Fiction via Word Embeddings to visualize character clusters and stylistic patterns in literature. Similarly, Supun Nakandala applied $t$-SNE in Gendered Conversation in a Social Game-Streaming Platform to reveal how user interaction patterns naturally cluster based on social dynamics.
When global relationships or thematic distances are more critical than fine-grained local neighborhoods, techniques like MDS are preferred. This is demonstrated by Miaoyutian Jia et al. in Users' Feedback on COVID-19 Lockdown Documentary: An Emotion Analysis and Topic Modeling Analysis, which maps the macro-relationships and overlaps between different discussion themes.
Spatial Interfaces for Discovery and Navigation
When high-dimensional embeddings are projected onto a 2D plane, the resulting map can serve as an interactive interface for exploring massive datasets. Instead of relying on literal keyword searches, users can navigate data via spatial proximity.
Consider the challenge of exploring millions of AI prompts or chat logs. Krzysztof Adamkiewicz et al. developed PromptMap: An Alternative Interaction Style for AI-Based Image Generation, using UMAP to project millions of generative prompts into a geographic map metaphor. Users explore the model's capabilities by panning and zooming through semantic clusters. A similar approach is taken by Yuntian Deng et al. in WildVis: Open Source Visualizer for Million-Scale Chat Logs in the Wild, where projecting the first turn of conversations allows researchers to discover high-level data trends visually.
This spatial navigation metaphor extends to multimodal data as well. In VCR: Video representation for Contextual Retrieval, Oron Nir et al. map video archives into a 2D "Topics-Map," allowing users to discover content by moving through conceptual clusters rather than exact text matches. By turning abstract vectors into physical coordinates, these systems transform search from a query-matching problem into an exploratory journey.
Probing and Debugging Model Representations
Beyond navigation, visualization serves as a diagnostic tool to audit what a model has actually learned. If a model understands a domain, its internal embeddings should form distinct, logical clusters when projected.
For example, Jiayi Liao et al. applied dimensionality reduction in Concept Understanding in Large Language Models: An Empirical Study to show that LLMs cleanly separate concrete objects but struggle to organize abstract concepts hierarchically. In the realm of interpretability, Ekaterina Loginova et al. demonstrated in An Interactive Web-Interface for Visualizing the Inner Workings of the Question Answering LSTM that projecting hidden states can visually isolate clusters of correct versus incorrect predictions, exposing the model's decision boundaries.
This probing technique is highly cross-disciplinary. In computational biology, Dan Ofer et al. used $t$-SNE in Protein language models expose viral mimicry and immune escape to evaluate how protein language models represent distinct protein classes, revealing where viral mimicry causes the model to misclassify sequences.
Interactive Refinement and Alternative Embeddings
Visualization does not have to be a passive, one-way street; it can act as a bridge for human-in-the-loop alignment. In Semantic Concept Spaces: Guided Topic Model Refinement using Word-Embedding Projections, Mennatallah El-Assady et al. introduced a framework where users interact directly with 2D word-embedding projections. By manually adjusting the positions of words on the map, users inject domain knowledge back into the model, refining the underlying topic distributions.
While post-hoc dimensionality reduction (like $t$-SNE or UMAP) is the dominant paradigm, it can introduce distortion because it forces high-dimensional geometries into a lower-dimensional container. An alternative approach is to change how the embeddings are learned from the beginning.
Charlie Frogner et al. proposed a novel solution in Learning Embeddings into Entropic Wasserstein Spaces. Instead of embedding data as single points in a high-dimensional Euclidean space, they embed data as probability distributions directly into a low-dimensional (2D or 3D) ground space. Because the representation itself is inherently low-dimensional, it can be visualized directly without external reduction techniques. This approach captures complex, multi-faceted relationships like polysemy through overlapping probability mass, bypassing the distortion entirely.
Key Foundational Work
Monolingual and Bilingual Concept Visualization from Corpora (2003) — contributed 1 concept elements to this direction.
Self-Organizing Chinese and Japanese Semantic Maps (2002) — contributed 1 concept elements to this direction.