Noah Meissner

The Joint Ontology Workshops (JOWO) have been held annually since 2015 as a satellite event of FOIS (Formal Ontology in Information Systems), bringing together a broad spectrum of ontology-related workshops under one roof. Over the past decade, the ontology research landscape has shifted considerably — driven by the rise of knowledge graphs, large language models, and applied AI. This raises the question of whether JOWO's thematic focus has followed suit.

Research Questions:

RQ1: How has JOWO evolved in terms of topics over the past ten years? Are there emerging trends, or topics that have become less prominent?

RQ2: How does JOWO's thematic profile compare to FOIS — the more selective, formally oriented main conference it accompanies?

Paper metadata (title, authors, year, URL) was collected from DBLP using the following conference scope:

conference_dict = {
    "JOWO": [2015, 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025],
    "FOIS": [2016, 2018, 2020, 2021, 2023, 2024, 2025]
}

Since DBLP does not provide abstracts, a custom PDF Extraction Pipeline was used to extract abstracts and keywords directly from the papers (see 01_Extraction.ipynb).

A multi-stage pipeline:

1. Download & Parsing: PDFs are downloaded from CEUR-WS / IOS Press and the first 5 pages are parsed using PyMuPDF.
2. Regex Extraction (fast path): Regular expressions attempt to locate abstract and keyword sections by matching structural markers (e.g., "Abstract", "Keywords:").
3. LLM Fallback (slow path): If regex fails, the extracted text is passed to a local LLM (phi4-mini via Ollama) with a structured prompt to recover the missing fields.
4. Incremental Storage: Results are written to CSV after each paper, enabling seamless resumption after interruptions.

To assess pipeline quality, a sample of 20 papers was annotated manually against ground-truth keywords (collected via a separate CSV), and the following metrics were computed (see evaluate_extraction.ipynb): (!TODO to show)

To identify high-level topics, extracted keywords were clustered using the following pipeline (see 02_Cluster.ipynb):

1. Semantic Embeddings: Each unique keyword was embedded using the embeddinggemma model (via Ollama), capturing distributional meaning.
2. Dimensionality Reduction: Embeddings were L2-normalized and reduced to 30 dimensions with PCA.
3. Optimized K-Means: K-Means was applied for $k \in [5, 24]$; the Silhouette Score selected $k = 5$ as optimal. The 5 raw clusters were then merged by hand into 5 interpretable macro-clusters based on the most frequent keywords in each:

Interactive visualizations were built with Plotly in 03_Analysis.ipynb:

JOWO 2016 is listed in DBLP but no papers could be extracted. Analysis covers 492 papers across 9 editions (2017–2025).

FOIS coverage: 7 editions (2016, 2018, 2020, 2021, 2023, 2024, 2025), 145 papers. Comparison uses average cluster shares across all available editions. Direct year-by-year comparison is not used due to FOIS's biennial publication schedule.