← News

ResearchSources: arXiv · 18 Nov 2026

Visibility graphs turn averaged word embeddings into competitive news classifiers

A new arXiv paper reports that features derived from visibility graphs can match or beat embedding-based baselines on AG News topic classification once the vocabulary is large enough.

In plain words

The authors convert document summaries into number sequences and then into graphs that record how the numbers relate to each other. A classifier fed these graph features classified AG News headlines about as accurately as standard word-embedding methods, provided enough words were kept in the vocabulary.

What was announced

The study, posted as an arXiv abstract, tests natural visibility graphs as an intermediate step for news topic classification. Each document's word embeddings are averaged into one fixed-length vector, read as an ordered series, and converted into a graph. That graph reaches the classifier either as node-degree values or as an adjacency matrix, with vocabulary sizes held between 100 and 1,500 selected words.

On the four-class AG News benchmark, the authors report that a fully connected classifier using the degree representation was the most accurate setup from 700 words onward, peaking at 91.24% accuracy with 1,000 words and beating a BERT-based baseline by 0.29 percentage points at 1,500 words. In a separate convolutional test, the adjacency input reached 86.16% at 1,000 words versus 83.12% for the strongest static embedding. The authors caution that the small margins need further controls to establish reliability, and the findings come from one benchmark under controlled vocabulary budgets.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, so methods and results cannot be independently checked.
  • The reported accuracy gaps are small, and the authors themselves note that controls are needed to confirm the reliability of these differences.
  • Results cover a single benchmark (AG News) under controlled vocabulary budgets, so performance on other datasets or unrestricted vocabularies is unknown.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.