{"path":"research/semantic-disambiguation-and-concept-tracking.md","content":"# Semantic Disambiguation and Concept Tracking\n\n**Date**: March 28, 2026\n**Origin**: /spec session conversation exploring what goes in structure vs language, the layers of meaning in an argument graph, and the 2013 \"concept tracking\" sketch.\n\n---\n\n## The Problem: Words Hide Disagreement\n\nOne of the primary sources of disagreement in the world is that the same words mean different things to different people. This was identified by Emanuel Kumlien (linguistics student) in a 2013 voice memo conversation with Fredrik ([voice-memo-emanuel-sofia.md](voice-memo-emanuel-sofia.md)):\n\n- Two people can give the SAME definition of a concept, genuinely believe they mean the same thing, yet in discursive practice draw completely opposite conclusions\n- The Swedish \"barnfattigdom\" (child poverty) debate was entirely semantic — everyone agreed on the facts, but different operative definitions generated a public fight about nothing\n- Even when semantic agreement is achieved, discourse-level meaning can diverge\n\nThis is not an edge case — it is central to contested concepts like \"freedom,\" \"justice,\" \"equality,\" \"rights,\" \"capitalism,\" \"democracy.\" These are the concepts that matter most to deliberation, and they are the ones most afflicted by semantic multiplicity.\n\n## Three Layers of What Words Do in an Argument Graph\n\n### Layer 1: What CAN be expressed in language\nPropositional claims, evidence links, logical relationships. This is the graph's natural territory. \"Factory farming causes animal suffering\" is a claim that language handles well.\n\n### Layer 2: What language HIDES\n\"Freedom matters\" — a libertarian means freedom FROM government; a socialist means freedom FROM deprivation; a Buddhist means freedom FROM attachment. Same words, same embedding vector (the words are identical!), completely different operational meaning. Emanuel's 2013 insight: semantic agreement masking real disagreement.\n\nLLMs capture \"meaning as use\" (Wittgenstein) through statistical patterns. They CAN distinguish senses in context, but for bare contested claims, they embed the AVERAGE meaning across training data — flattening precisely the disagreements that matter.\n\n### Layer 3: What CANNOT be put into language at all\nPolanyi's \"tacit knowledge\" — we know more than we can tell. Embodied intuitions, aesthetic judgments, the felt sense that a worldview is right before articulation. This layer may be fundamentally outside Deliberus's scope — or it may be what the worldview filter approximates non-propositionally.\n\n## Four Node Types for the Graph Ontology\n\nThe graph needs to distinguish at least four types of nodes:\n\n| Type | Example | Verifiable? | Role in worldview |\n|------|---------|-------------|-------------------|\n| **Empirical claim** | \"Global temperatures rose 1.1°C since 1900\" | Yes — evidence, NLI, calibration | Factual substrate — same for all worldviews |\n| **Normative claim** | \"We should reduce emissions drastically\" | No — decomposable into value axiom + empirical premise | Where worldviews diverge in CONCLUSIONS |\n| **Definitional claim** | \"By 'freedom' I mean absence of state coercion\" | No — but CRITICAL for disambiguation | Where worldviews diverge in MEANING |\n| **Value premise** (currently undecomposed) | \"Individual liberty outweighs collective welfare\" | No — but ALWAYS decomposable further (WHY?) | Appears foundational in context; system invites deeper examination. See [extraction-virtual-run.md](extraction-virtual-run.md) §No Copout Axioms |\n\n**Key insight**: Definitional claims are themselves arguments. \"By freedom I mean X\" is not neutral metadata — it's a MOVE in the argument. Different definitions of the same word lead to different conclusions from the same evidence. The system must treat definitions as first-class nodes in the graph, not as disambiguation preprocessing.\n\n## Concept Tracking: The Crowdsourced Solution (2013 Sketch)\n\nFredrik sketched \"Concept tracking (experimental)\" in 2013, directly influenced by Emanuel Kumlien's linguistic insights. The sketch shows:\n\n- **Word disambiguation**: \"Sweet\" → \"Sweet (sugar)\" / \"Sweet (nice)\" / \"Sweet (?)\" — a single word decomposed into distinct sense-nodes\n- **Concept branching**: \"humans\" tracked across different conceptual uses with branching meanings\n- **Crowdsourced operations**: \"Meta: merges/splits\" — the community can merge equivalent concepts or split ambiguous ones\n- **\"Rösta/argumentera\"** (Vote/argue) — voting and argumentation as the mechanisms for resolving disambiguation disputes\n- **\"axiom 100% (grundläggande)\"** — axioms labeled as \"fundamental\" with 100% confidence\n- **Lifecycle states**: \"Stabilisera / Komplett\" (Stabilize / Complete) — concepts progress through disambiguation stages\n\nThis is primary source evidence that concept tracking — semantic disambiguation as first-class graph structure — has been part of the Deliberus vision since 2013. It was conceived as experimental and crowdsourced, not top-down.\n\n## The Extraction Pipeline Implication\n\nWhen the extraction pipeline encounters a contested concept, it should:\n\n1. **Flag the ambiguity** (\"this claim uses 'freedom' — which sense?\")\n2. **Offer alternative formulations** (LLM-generated sense disambiguation)\n3. **Store the disambiguation as a definitional claim node** in the graph\n4. **Link downstream arguments to the SPECIFIC definition they depend on**\n\nThis way, two people who both say \"freedom matters\" but mean different things get DIFFERENT nodes in the graph, connected to different downstream conclusions. The semantic disagreement becomes VISIBLE in the structure rather than hidden in the language.\n\n**Implementation update (Apr 3, 2026):** This is no longer purely aspirational. Claim pages now let users clarify contested terms directly, those clarifications create definitional claims, concept pages accumulate shared sense memory, extraction pages now surface an earlier interpretation checkpoint for ambiguous source usages, and the landing page now runs a lightweight preflight for text/URL input before full extraction settles. Concept pages and APIs now also surface emergent lifecycle states like underdefined, emerging, bifurcated, fragmented, and locally stable, so the graph can start saying not only \"here are the senses\" but \"this concept is splitting\" or \"this concept is settling.\" There is also now a browseable `/concepts` index so this layer can be inspected across the graph rather than only one term at a time. For first-time source material, where graph memory has no reusable senses yet, the preflight can still offer provisional sense candidates from contested-concept detection rather than asking the user to type into an empty field. The remaining gap has moved deeper still: file-upload entry, voice-first continuity into sense selection, and fuller merge/split governance beyond these first lifecycle signals.\n\n**Governance implication (Apr 3, 2026 evening):** The next sensible step for sense governance is probably not literal hard merge but **soft canonical clustering**. Concept senses that look highly similar may be treated as belonging to one provisional operative cluster for retrieval, feeds, and display, while the underlying nodes remain distinct and provenance-preserving. This is more compatible with Deliberus's actual purpose than aggressive merge, because concept-sense false merges erase precisely the distinctions the platform is meant to reveal. So the likely progression is: candidate merge → locally stable cluster → explicit split if false sameness appears, rather than silent collapse into one canonical sense too early. See [soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md).\n\n**The \"aha moment\"**: \"Oh, we've been using the same word to mean different things, and THAT'S why we disagree\" — Deliberus dissolving a false disagreement through structural clarity.\n\n## What Should Be STRUCTURE vs LANGUAGE\n\n- **Language**: the content of claims (what people actually said/meant)\n- **Structure**: the relationships between claims (supports, attacks, depends-on, defines-term-for), the claim types (empirical/normative/definitional/axiological), and the disambiguation links (this version of \"freedom\" → that set of conclusions)\n\n**The structure makes visible what language hides. That's the whole point of Deliberus.**\n\n## The Bottom-Up vs Top-Down Resolution\n\nThe worldview \"dials\" (what changes when you put on different glasses) should NOT be defined top-down through a predetermined taxonomy (Haidt's moral foundations, political compass axes). This would impose the designer's epistemological framework — the \"neutrality illusion.\"\n\nInstead: build the graph, let people evaluate claims (including two-axis voting), and let the worldview groupings EMERGE from patterns of agreement and disagreement. The minimum structural requirement for day one: fact/value classification in the extraction pipeline. The full worldview filter is a Phase 2 discovery, not a Phase 1 design decision.\n\nDennett's \"Universal Acid\" applies: reason, once applied to any topic, dissolves assumptions and reveals structure regardless of starting point. The specific seed topic is almost arbitrary because the value is in the method.\n\n## The Spec Session Resolution (What's Needed for MVP)\n\n**Day one minimum**:\n- Claim extraction with fact/value classification\n- Definitional claim nodes as first-class entities (concept tracking)\n- Two-axis evaluation primitive (agree/disagree + well-argued/poorly-argued)\n\n**Phase 2 (emerges from accumulated data)**:\n- Worldview clusters (discovered bottom-up from evaluation patterns)\n- The full worldview filter (perspective-shifting based on discovered clusters)\n- Bridging arguments (cross-cluster reasoning quality detection)\n\n## Cross-References\n\n- [voice-memo-emanuel-sofia.md](voice-memo-emanuel-sofia.md) — Emanuel Kumlien's 2013 discourse-layer insight\n- [steelmanned-critiques.md](steelmanned-critiques.md) §Formalization Paradox — Wittgenstein's \"meaning is use\"\n- [conceptual-threads.md](../conceptual-threads.md) §Thread 3 (Fact/Value Boundary) — the 30°C example\n- [conceptual-threads.md](../conceptual-threads.md) §Thread 4 (Dedup) — false equivalence at discourse level\n- [sketches.md](../sketches.md) §Concept Tracking Sketch — 2013 primary source\n- [vision.md](../vision.md) §The Core Dialectic — analysis↔attunement, bottom-up worldview discovery\n"}