ACOM embedding grid
A study of semantic structure in a discrete document layout.
Points in a two-dimensional document plot can overlap. A discrete grid gives each document a distinct cell, but that constraint can distort relationships.
At ELTE, I implemented an ACOM-style grid optimizer and an evaluation pipeline to study that tradeoff.
The layout is an assignment problem
Documents first become embeddings. The optimizer assigns documents to cells on a fixed grid and improves the arrangement by swapping occupants.
The result is a discrete layout throughout optimization. It is not a continuous projection rounded to grid coordinates afterward.
The grid provides an address for each document. Its capacity must be chosen before placement, which constrains the collection it can hold.
Every method uses the same distance reference
The pipeline computes semantic distances once from the embeddings. ACOM, PCA, t-SNE, and optional UMAP share that reference.
A shared reference makes the comparison about the layouts. Different distance calculations would introduce another source of variation.
The evaluation measures preserved structure
Trustworthiness, neighborhood preservation, and stress assess each layout against the same reference. Runs retain their outputs and figures for inspection.
The work was accepted to ACM ICPS and presented at IntRob Fair 2026 with the ELTE AI Department under Prof. László Gulyás.
The study compares a specific corpus and embedding model. Its results do not establish that the same layout works best for other collections.
Discrete placement has a cost
A grid restricts where a document can be placed. Some distance information must change when continuous relationships are mapped to fixed cells.
The useful question is how much structure remains under that restriction. The project does not claim that grids should replace continuous projections.