taxonomy-structural-matrix-1024-mxbai

Taxonomy Structural Matrix (1024-D, mxbai) — Semantic-Swingers @ LLMs4OL 2026 (Task C)

A trained bilinear structural adjacency matrix W used by the Semantic-Swingers taxonomy-discovery learner for the 3rd LLMs4OL Challenge (2026), Task C (Taxonomy Discovery). It powers a scalable alternative pipeline (SemanticSwingersMatrixTaxonomyLearner) to the team's primary LLM taxonomy selector — a scale-aware path that bypasses per-node LLM inference on very large ontologies. Trained via contrastive margin-ranking to suppress sibling edges.

What this is

A single torch.nn.Linear(1024, 1024, bias=False) weight (W.weight, shape [1024, 1024], float32) that maps a child embedding into an asymmetric parent space. Given L2-normalised mixedbread-ai/mxbai-embed-large-v1 embeddings of a child and a candidate parent, the score is ⟨W·child, parent⟩. It provides a fast, direction-aware structural signal over candidate is-a edges, which the hybrid learner combines with an LLM parent-selector (and a high-speed matrix-only bypass for very large type vocabularies).

• File: structural_matrix_w_1024_mxbai.pt (state dict: {"W.weight": [1024, 1024]})
• Base encoder: mixedbread-ai/mxbai-embed-large-v1 (1024-D)
• Task: LLMs4OL 2026 Task C — Taxonomy Discovery

Usage

Loaded automatically by the SemanticSwingersMatrixTaxonomyLearner in our OntoLearner integration — no manual download needed:

from ontolearner.learner.taxonomy_discovery import SemanticSwingersMatrixTaxonomyLearner

learner = SemanticSwingersMatrixTaxonomyLearner(
    embedding_model="mixedbread-ai/mxbai-embed-large-v1",
    top_k=10,
    selector="openai",   # champion; "embedding"/"ollama" for offline reproduction
)
learner.load()   # fetches this matrix from the HF registry if not present locally

Or load the raw weights directly:

import torch, torch.nn as nn
from huggingface_hub import hf_hub_download

W = nn.Linear(1024, 1024, bias=False)
W.load_state_dict(torch.load(
    hf_hub_download("datagero/taxonomy-structural-matrix-1024-mxbai",
                    "structural_matrix_w_1024_mxbai.pt"),
    map_location="cpu"))

Links

• OntoLearner integration PR: https://github.com/sciknoworg/OntoLearner/pull/338
• Team project repo: https://github.com/matias-vizcaino/OntoLearner (public fork)
• Challenge: https://sites.google.com/view/llms4ol2026

License

MIT.