A Pragmatist’s Guide to Building Knowledge Graphs from Unstructured Data

A Pragmatist’s Guide to Building Knowledge Graphs from Unstructured Data

by Alessandro Pireno

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQEExpert systems
🎙 Alessandro Pireno 👥 5K 📅 October 21, 2025 ⏱ 25 min 👁 155 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

knowledge graphunstructured dataLLMNLPentity resolution

Summary

Alessandro Pireno presents a pragmatic approach to building knowledge graphs from unstructured data, drawing on his experience at a startup. He introduces the concept of KG-ETL, treating knowledge graphs as a first-class output of data pipelines. He compares three architectures: LLM-based, traditional NLP-based, and hybrid vector search-based. LLM-based methods are powerful but costly, slow, and noisy. NLP-based methods are deterministic, explainable, and fast, but brittle and require upfront knowledge. Hybrid approaches combine techniques for better precision and recall. He details a novel technique using FastText embeddings for entity resolution, treating it as a classification problem to link misspellings to canonical entities. He emphasizes the importance of scoring and confidence in NLP outputs, and suggests using SLMs (small language models) to refine results. He also discusses using LLMs offline to improve rules and search criteria, creating a self-improving system. The talk concludes with a decision framework: use NLP/SLMs for cost-effective scalability and high precision, and LLMs for high-value tasks or training the system. He answers a question about open-source code and throughput, noting that LLM latency is unpredictable while NLP is more scalable.

182 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable practical insights into building knowledge graphs, especially the comparison of LLM vs. NLP approaches and the novel use of FastText embeddings for entity resolution. The argumentation is based on personal experience and logical reasoning, but lacks quantitative benchmarks or formal evaluations. The speaker clearly explains trade-offs in cost, latency, and precision, and offers a decision framework that is useful for practitioners. However, the claims about scalability and performance are anecdotal and not backed by rigorous data.

Scientific Rigor, Source Quality, Title Accuracy

The talk is an expert opinion based on the speaker’s professional experience, but it does not cite specific sources or references. The only external link provided is to the MLOps World conference, which is not a source for the technical content. The title accurately reflects the content, which is a pragmatic guide. The lack of citations and empirical evidence limits the scientific rigor, but the practical insights are valuable for practitioners.

166 words

Title / Content Match

The title accurately reflects the content, which is a pragmatic guide based on the speaker's hands-on experience.

Quality & Reliability

7/10

The talk provides a practical, experience-based overview of knowledge graph construction from unstructured data, with clear comparisons of techniques and trade-offs. However, it lacks formal citations, empirical benchmarks, and detailed reproducibility, limiting its scientific rigor.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • MLOps World — Conference where the talk was presented, providing context for the talk.

Concurring Sources

  • Knowledge graph — General concept of knowledge graphs.
  • FastText — Library for the embedding technique mentioned.
  • spaCy — NLP library used in the talk.

Contribution & Novelties

The talk offers a pragmatic, experience-based framework for building knowledge graphs from unstructured data, comparing LLM, NLP, and hybrid approaches. It introduces a novel technique using FastText embeddings for entity resolution, treating it as a classification problem rather than semantic search. The emphasis on using SLMs for refinement and LLMs for offline self-improvement is a practical contribution.

Pour aller plus loin :

  • Knowledge graph — Foundational concept.
  • FastText — Official library for the embedding technique mentioned.
  • spaCy — NLP library used in the talk for pattern matching.
  • Graph RAG — Related approach for using knowledge graphs with LLMs.

98 words

Radar Profile

The radar profile shows a balanced but moderate performance across all dimensions, with slightly higher scores in information quantity and quality, and lower in technical level and reliability. This reflects a talk that is informative and practical but lacks deep technical depth and rigorous sourcing.

Reliability 6/10

💬 Sur les 0 commentaires analysés, aucune tendance n'est disponible.