Metraipolis › archive › dc_ef4833a3dff7

Lightweight Semantic Embeddings Enable Redundancy-Sensitive Knowledge Graph Construction for LLM-Based Document Processing (2026) · DOI 10.21203/rs.3.rs-10053420/v1

Source
Metraipolis archive
Original
https://doi.org/10.21203/rs.3.rs-10053420/v1
License (as archived)
CC-BY-4.0
Archived
2026-10-08T07:03:09.695Z
SHA-256 of the archived text
9098f0c50c183780cd5172e1f5334f4234cf631e2ac5a54acfcdd5cbfdaace65
Stable URL of this page
https://metraipolis.com/documento/dc_ef4833a3dff7

How to cite

Lightweight Semantic Embeddings Enable Redundancy-Sensitive Knowledge Graph Construction for LLM-Based Document Processing (2026) · DOI 10.21203/rs.3.rs-10053420/v1 — Metraipolis archive, https://doi.org/10.21203/rs.3.rs-10053420/v1 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_ef4833a3dff7 (Metraipolis archive dc_ef4833a3dff7, sha256 9098f0c50c183780cd5172e1f5334f4234cf631e2ac5a54acfcdd5cbfdaace65)
BibTeX
@misc{dcef4833a3dff7,
  title = {Lightweight Semantic Embeddings Enable Redundancy-Sensitive Knowledge Graph Construction for LLM-Based Document Processing (2026) · DOI 10.21203/rs.3.rs-10053420/v1},
  howpublished = {Metraipolis archive},
  url = {https://doi.org/10.21203/rs.3.rs-10053420/v1},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_ef4833a3dff7; sha256 9098f0c50c183780cd5172e1f5334f4234cf631e2ac5a54acfcdd5cbfdaace65},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.21203/rs.3.rs-10053420/v1. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Articolo recente con licenza aperta: scheda dai metadati Crossref (posted-content).
Titolo: Lightweight Semantic Embeddings Enable Redundancy-Sensitive Knowledge Graph Construction for LLM-Based Document Processing
Autori: Sedar Olmez, Maxim Smilovitskiy, Koichi Yokota
Pubblicato: 2026-09-11 · editore: Springer Science and Business Media LLC
DOI: https://doi.org/10.21203/rs.3.rs-10053420/v1
Licenza dell'articolo: CC-BY-4.0

Abstract (testo degli autori, CC-BY-4.0):
Abstract Large language model (LLM) pipelines for constructing knowledge graphs from long technical documents can be costly because they often process every section independently, even when sections repeat the same semantic content. We propose an ML-augmented Automated Mapping of Data Definitions (AMODD) pipeline that adds a lightweight Semantic Embedding Layer for pre-LLM section deduplication, post-LLM triple clustering and semantic retrieval over the final graph. The layer uses a locally deployed all-MiniLM-L6-v2 sentence-transformer model to identify near-duplicate sections before extraction, consolidate semantically equivalent graph entities after extraction and rank triples by cosine similarity at query time. We evaluate AMODD using a controlled A/B experiment on synthetic hydrogen-sector regulatory documents with low, moderate and high redundancy. On the non-redundant document, the embedding layer introduced a 6-10% overhead. On moderately and highly redundant documents, it reduced token consumption by 56.7% and 84.9%, LLM calls by 57.7% and 85.4%, and wall-clock latency by 29.9% and 58.1%, while preserving final graph size above a redundancy-adjusted retention bound. Semantic retrieval also improved broad-query coverage over keyword overlap. These results show that lightweight embeddings can make LLM-based knowledge-graph construction redundancy-sensitive rather than document-length-sensitive.

Fonte: https://api.crossref.org/works/10.21203%2Frs.3.rs-10053420%2Fv1 — metadati Crossref, consultati il 2026-10-08. Interrogazione: «knowledge graph construction language models».