Research
When building an “ingest workflow” for scientific papers (PDFs) into my LLM wiki, I realized summaries that contain everything are counterproductive. Instead, I needed to extract only genuinely relevant information. The context “clogs” processing, creating enormous noise. Many details in a paper are simply irrelevant to specific research questions, but the model doesn’t know that on its own.
In the analogue world, the solution is straightforward: linear reading, highlighting, extracting. Time-consuming. Error-prone. Not scalable.
But there is an approach: extracting information through structured parsing using the Discourse Graph methodology.
The Discourse Graph is not merely an ontological knowledge concept—it’s a practical method for systematically extracting and linking knowledge from unstructured sources. The idea: decompose every text into granular, interconnected units rather than treating it as a monolithic block.
I applied the five core components of the Discourse Graph to my wiki as follows:
The primary document itself. Who wrote it? When? In what context? This anchors all downstream information.
All usable information from the source:
Methodological details (research design, sample size, instruments)
Theoretical foundations (concepts, frameworks)
Empirical findings (results, metrics, observations)
Important: Evidence is facts or observations, not yet interpretations.
The conclusions authors draw from evidence: “If evidence X shows, then statement Y follows.” Claims are the scientific findings a paper propagates.
The scientific questions the text answers or raises. These emerge inductively from evidence and claims. They map the research space.
Overarching ideas, theories, constructs linking multiple claims and evidence. They are the “building blocks” of my knowledge graph.
This structure is also called atomization or granularization of knowledge. Each unit is small enough to be understood in isolation, yet large enough to carry information content.
This structure builds on Stephen Toulmin’s argumentation theory (The Uses of Argument, 1958). Toulmin showed that valid arguments don’t consist merely of “thesis + proof,” but of:
Claim (assertion)
Data/Evidence (support)
Warrant (why evidence supports the claim)
Backing (support for the warrant)
Qualifier (limitation, certainty)
Rebuttal (counter-arguments)
The Discourse Graph is a practical application of this model to knowledge networking and proves very useful for analyzing scientific texts.
I developed this workflow iteratively. Rather than simply loading PDFs and generating summaries, I implemented structured parsing in the LLM wiki. As I outlined in my earlier work on the Wiki-LLM to Reasoning Linter, the pipeline looks like this:
PDF Input
↓
[1] Structured Parsing
→ Text sections & segments
→ Each segment indexable & linked
↓
[2] Evidence Extraction
→ Methodological evidence
→ Theoretical evidence
→ Results-based evidence
↓
[3] Claim Identification
→ Claims based on evidence
→ Claims as interpretations/conclusions
↓
[4] Question Generation (inductive)
→ Which questions does this text answer?
→ Which new questions emerge?
↓
[5] Knowledge Graph Assembly
→ Bidirectional links between all components
→ Define relation types
→ Graph storage (Tana, ObsidianMD, Neo4j, etc.)
↓
Output: Queryable Knowledge GraphFig. 1: The Pipeline (Flowchart) (author’s own)
This yields several advantages:
Structures, not summaries: The process doesn’t generate “The paper says X,” but rather “Evidence A supports Claim B, which answers Question C.”
Context reduction through granularity: Each segment is small enough to stay focused on its point.
Linking as quality assurance: If evidence and claim can’t be linked, something’s missing.
Bidirectional connections: I can ask not only “Which claims does this evidence support?” but also “What evidence do I have for this claim?”
I developed a Python parsing pattern (imperfect, but a solid start):
def parse_pdf_to_discourse_graph(pdf_path):
# 1. Decompose PDF into segments
segments = extract_segments(pdf_path)
# 2. For each segment: identify evidence
for segment in segments:
evidences = extract_evidences(segment, types=[
"methodological",
"theoretical",
"empirical"
])
# 3. Derive claims from evidence
claims = derive_claims(segment, evidences)
# 4. Generate questions inductively
questions = generate_questions(segment, evidences, claims)
# 5. Store in graph with relations
save_to_knowledge_graph({
"source": pdf_metadata,
"segment": segment,
"evidences": evidences,
"claims": claims,
"questions": questions,
"links": create_bidirectional_links(...)
})Future note types can extend beyond claims, evidence, questions, and concepts:
Hypotheses: Conjectures derived from the data—not just what the paper says, but what it could mean.
Didactical Patterns: For educational development and learning research, didactical concepts also matter.
Methodological Innovations: New methods emerging from the findings.
I also need more sophisticated relation grammar between knowledge types. Relations should describe not just “these things are linked,” but how they’re linked.
This connects to several other posts in this digital garden:
From PKM to CKM: Collaborative Knowledge Management instead of Personal Knowledge Management—exactly what a knowledge graph enables.
Wikipedia Project: PKM: The idea of structurally curating public knowledge.
Why We Write Publicly: Because public knowledge becomes more valuable through networking.
From Wiki-LLM to Reasoning Linter: The technical realization—automated “linting” of arguments.
Everything Claude Code: The tools to automate this.
Granularization as Standard: What if academic research expected scholars to publish findings in granular form rather than writing books and papers?
Research as Graph Query: Instead of “Write a review,” researchers could frame tasks like: “Create a query on the graph: Which evidence supports thesis X? Which concepts are controversial? Where are the gaps?”
Automated Synthesis: With LLMs, I can not only read papers but query the graph directly and generate syntheses. Rather than writing them manually, a pre-structured text with an evaluation of existing findings can be created as inspiration for new writing.
Interdisciplinary Insight: When I link knowledge graphs across disciplines, new insights emerge simply through the architecture of the graph itself, alongside manual thinking.
All this sounds like information extraction from scientific texts can be fully automated—work machines will eventually handle. Yes, I do hope to offload certain manual tasks. Yet we as researchers remain responsible for quality assurance. Results must be verified in the end. That’s probably an idea for a later post.
Created: 2026-05-17 · v01
â—Ź Seedling
â—Ź Growing
â—Ź Evergreen