Tracing the Ecosystem of Open Science

Data Collection

Our initial dataset was built by querying the OpenAlex database using combinations of phrases such as "open science ecosystem", "ecosystem in open access", and others. After cleaning, 211 scholarly articles remained for analysis.

Methodology Overview

We used GPT-based prompt engineering to extract entity-relation triples centered on the term ecosystem. These were converted into a semantic knowledge graph using NetworkX and Pyvis. Nodes were labeled with a manually refined two-level taxonomy.

The semantic network was further enriched with a citation network, forming a multilayer model that maps both concepts and communities of scholarship.

Results Summary

Our results demonstrate that the term ecosystem acts as a boundary object, linking diverse thematic communities in Open Science. In particular, Community 5 shows its use in techno-economic discourse (e.g., "machine learning", "digital transformation"), whereas Community 4 aligns it with policy and governance contexts. These communities frame different priorities through the same term, confirming its flexible communicative function.

Our semantic and multilayer network visualizations validate the centrality of core terms like open science, sharing, and repository, while also revealing the bridging function of ecosystem as a node with high betweenness centrality. These patterns support our central thesis: ecosystem is not just a metaphor, but an active, structural mediator in the scholarly landscape.

📊 Semantic and Multilayer Networks

Semantic Knowledge Graph

Multilayer Citation + Semantic Network

Taxonomy Network

Two-Level Taxonomy of Thematic Categories

The table below shows the finalized two-level thematic taxonomy used to label nodes in the semantic network. It includes 18 main categories and their corresponding sub-categories extracted and grouped manually after entity cleaning and standardization.

LLM Prompt for Knowledge Extraction

The following is the full prompt provided to the GPT-3.5 Turbo model for the extraction of entities and relations from the text chunks. It was specifically designed to extract only meaningful concepts and semantic relationships centered around the word "ecosystem".

You are an expert in knowledge extraction.
Your task:
- Focus ONLY on text where the word "ecosystem" appears.
- From those parts, extract meaningful, well-formed entities (as nodes) 
  and their relationships (as triples).

Guidelines:
- Include any terms that directly contain the word "ecosystem" 
  (e.g., "Open Science Ecosystem", "Blockchain Ecosystem") as part 
  of the concept list.
- Only include entities that are concrete or conceptual — no vague 
  or generic entries.
- Limit to a maximum of 20 concepts and maximum of 20 relations.
- It is perfectly fine if there are fewer than 20 concepts or 
  relations, depending on the richness of the information.
- Do NOT invent or hallucinate content that is not explicitly or 
  implicitly supported by the text.
- Prioritize clarity, importance, and relevance.
- Avoid very long phrases (keep concept labels under 6-7 words).
- For relations, avoid generic verbs like "is", "are", or 
  conjugated forms like "includes". Use meaningful, specific verbs 
  (e.g., "participate in", "enable", "comprise").

Output Format (strictly):
Return a valid JSON object with exactly these two keys:
- "concepts": a list of important concept strings
- "relations": a list of [subject, relation, object] triples

Example:
{
  "concepts": ["Open Science Ecosystem", "Repositories", 
               "Researchers", "Knowledge Sharing"],
  "relations": [
    ["Researchers", "participate in", "Open Science Ecosystem"],
    ["Repositories", "enable", "Knowledge Sharing"]
  ]
}

Strict instructions:
- Only return the JSON object — no explanations, commentary, or 
  extra text.
- Do not repeat terms unnecessarily.
- Ensure that the JSON is valid (no missing commas, brackets, or 
  quotation marks).