Personal Knowledge Graphs: LLM Triple Extraction with Qwen, Gemma

Abhirup Dasgupta, Fernando Spadea and Oshani Seneviratne submitted a paper on 18 Apr 2026 (arXiv:2607.00003) that presents a reproducible pipeline using lightweight Large Language Models to extract structured user-preference triples from conversational data for Personal Knowledge Graphs. The authors evaluate Qwen- and Gemma-based models on producing RDF-compliant triples linked to Wikidata identifiers and test how those graphs perform in a downstream recommendation task.

What did the paper evaluate?

The paper evaluated Qwen- and Gemma-based models for two linked goals: fidelity of semantic triple extraction and the utility of the resulting graphs in recommendations. The authors measured how well lightweight LLMs convert unstructured conversational "strings" into RDF-compliant, Wikidata-identified "things" and then assessed whether those triples improve a downstream recommendation task.

The evaluation therefore covers both extraction quality and end-to-end usefulness. The abstract specifically states the models were judged on their ability to extract RDF-compliant triples linked to Wikidata identifiers from conversational data for Personal Knowledge Graph construction, and that downstream recommendation performance was compared to triple-extraction performance.

How does the pipeline turn strings into things?

The paper presents a reproducible pipeline that uses lightweight Large Language Models to extract structured user-preference triples from decentralized conversational inputs. At a high level, the pipeline converts conversational text into RDF-compliant triples and links entities to Wikidata identifiers so the triples become machine-interpretable nodes in a Personal Knowledge Graph.

The authors frame this pipeline as a bridge between conversational "strings" and semantic "things," emphasizing extraction of user-preference triples suitable for PKG construction. The arXiv entry lists sections for associated code, data and media for the article, indicating the authors intended the work to be reproducible and inspectable beyond the paper itself.

How did the models perform?

Certain models performed well, and those models showed proportionally high downstream performance relative to their triple extraction performance. The abstract reports a direct relationship between extraction fidelity and recommendation utility: models that produced higher-quality triples tended to yield better downstream recommendation results.

Beyond that core result, the paper compares Qwen- and Gemma-based approaches as representative lightweight LLM families for this task. The evaluation explicitly covers semantic extraction fidelity and the utility of resulting graphs in a recommendation pipeline, tying model-level output quality to application-level gains.

Why it matters

Personal Knowledge Graphs offer a privacy-preserving way to model user preferences, but constructing them from unstructured, decentralized conversational data is challenging. Demonstrating a reproducible path from conversational text to RDF-compliant, Wikidata-linked triples shows that lightweight LLMs can feasibly populate PKGs. The reported proportionality between extraction fidelity and downstream gains implies that improving triple extraction is not just an academic metric: it directly affects recommendation quality.

That linkage matters for teams building on-device or privacy-focused recommenders: investment in better extraction pipelines can translate to tangible application improvements without requiring centralization of raw conversational data.

What to watch

Check the paper's associated materials listed under "Code, Data and Media" on the arXiv page for implementation details and artifacts that enable replication. Subsequent work should show whether the proportional relationship the authors report holds across larger conversational datasets, other lightweight LLM families, or in live recommendation deployments.

Citation and metadata: From "Strings" to "Things" for Personal Knowledge Graphs: Evaluating LLM Triple Extraction for Recommendation Systems, Abhirup Dasgupta, Fernando Spadea, Oshani Seneviratne; arXiv:2607.00003, submitted 18 Apr 2026.