{"path":"research/argsbase-argument-web.md","content":"# ARG-tech, Argument Web, and Structured Deliberation Research\n\n**Research date**: March 27, 2026\n**For**: Deliberus project — understanding the computational argumentation ecosystem\n\n---\n\n## 1. Chris Reed & ARG-tech (Centre for Argument Technology)\n\n**Who**: Professor Chris Reed, University of Dundee, Scotland. Heads the Centre for Argument Technology (ARG-tech) — the world's leading academic group in computational argumentation. 20+ years at the overlap of argumentation theory and AI.\n\n**Scale**: Over £13m in funding, 200+ refereed papers, 5 books. Tens of thousands of users across their tools.\n\n**Core research areas**:\n- Argument mining (automatic extraction of reasoning from text)\n- Reason checking (argumentation-based fact-checking)\n- Public deliberation (large-scale democratic debate)\n- Inference Anchoring Theory (IAT) — linking logical inferences to dialogue structure\n- Argumentation schemes (Walton-style stereotypical reasoning patterns)\n\n### Key People\n\n| Person | Role |\n|--------|------|\n| **Chris Reed** | Professor, director. Argumentation theory + AI |\n| **John Lawrence** | Key collaborator. Argument mining, AIFdb infrastructure, OVA tools |\n| **Debela Gemechu** | First employee of Arg Technica Ltd. External knowledge + argument mining |\n| **Kamila Gorska** | First employee of Arg Technica Ltd. FORECAST2023 corpus |\n| **Ramon Ruiz-Dolz** | DialAM shared task, multilingual argumentation |\n\n### Currently Funded Projects (2024-2028)\n\n| Project | Funder | Dates | Focus |\n|---------|--------|-------|-------|\n| **AI4Deliberation** | EU Horizon | 2024-2027 | AI tools for democratic deliberation |\n| **HDR-UK** | HDR UK | 2024-2027 | Argumentation in distributed data environments |\n| **CLARUS** | EU | 2025-2028 | Digital forensic investigation |\n| **SIGNATURE** | IARPA | 2022-2026 | Protecting identity of online discourse participants ($2.5m) |\n| **TITAN** | EU | 2022-2025 | Countering disinformation |\n| **Deliberation Laboratory** | VW Foundation | 2021-2026 | Building deliberation infrastructure |\n| **AMECC** | Swiss NSF | 2021-2025 | Financial argument structure in earnings calls |\n\n### Notable Completed Projects\n\n- **Trajectories of Conflict** (AHRC, 2021-2024) — worked with UN Security Council on conflict analysis\n- **Fake News Immunity** (ESRC, 2020-2021) — combating COVID infodemic\n- **SERAPIS** (DSTL, 2020-2023) — argument technology for defence/intelligence\n- **ADD-up** (VW Foundation, 2017-2021) — argument technology in government participation\n- **Argument Mining** (EPSRC, 2016-2019) — automatic reasoning extraction from text\n\n---\n\n## 2. AI4Deliberation EU Horizon Project\n\n**Grant**: €2,999,500 (100% EU-funded)\n**ID**: 101178806 (HORIZON-RIA)\n**Period**: November 2024 — October 2027\n**Coordinator**: Uni Systems (Athens, Greece)\n\n### Objective\n\nDeliver robust and ethical AI tools alongside detailed guidance to support governments in institutionalising, implementing, and evaluating multimodal, gamified, large-scale deliberative processes. Addresses declining trust in institutions, political polarization, and disinformation.\n\n### Consortium (12 partners, 8 countries)\n\n| Partner | Country | EU Contribution | Notes |\n|---------|---------|-----------------|-------|\n| Uni Systems | Greece | €439,250 | Coordinator |\n| TU Delft | Netherlands | €402,125 | |\n| CERTH | Greece | €395,312 | |\n| Konnekt Able Technologies | Ireland | €350,000 | SME |\n| Otto-Friedrich-Universität Bamberg | Germany | €222,375 | |\n| University of Dundee (ARG-tech) | UK | €220,375 | Argument mining lead |\n| Institut za Kriminologijo | Slovenia | €220,000 | |\n| Thoughtgraph Ltd | UK | €200,250 | |\n| Dembrane BV | Netherlands | €150,250 | SME, deliberation platform |\n| Stadt Bamberg | Germany | €101,437 | Pilot site |\n| ActionAid Italia | Italy | €100,625 | |\n| Universität Zürich | Switzerland | — | Associate (no-cost) |\n\n### Key Deliverables\n\n- Summarization, moderation, fact-checking, and hate speech detection capabilities\n- Four large-scale pilots (city to international level) on climate change and long COVID\n- Open-source AI toolkit for deliberative democracy\n- Framework with practical guidelines for institutional adoption\n\n### Dembrane BV (Consortium Member)\n\nDutch AI startup focused on \"making sense of big, messy conversations.\" Won OpenAI's \"Democratic Inputs to AI\" grant with the DeliberationAtScale team from Eindhoven. Building open-source tools for mass deliberation integrating AI with democratic innovations (cross-pollination, sortition). Part of the bridge between ARG-tech's academic work and practical deployment.\n\n**Deliberus relevance**: Dembrane is a potential peer/competitor in the deliberation space. Their open-source commitment and EU backing make them worth watching.\n\n---\n\n## 3. AIF (Argument Interchange Format)\n\n### What It Is\n\nAn international standard for representing argumentation structures, originating from a 2005 Budapest colloquium. Think of it as \"HTML for arguments\" — a common ontology that lets different tools exchange argument data.\n\n### Core Ontology\n\n| Node Type | Code | Purpose |\n|-----------|------|---------|\n| **I-node** (Information) | `I` | Propositional content (claims, premises, data) |\n| **S-node** (Scheme) | — | Inferential links between propositions |\n| **RA-node** (Rule of Inference Application) | `RA` | Support relationships |\n| **CA-node** (Conflict Application) | `CA` | Attack/conflict relationships |\n| **PA-node** (Preference Application) | `PA` | Preference orderings |\n\n### AIF+ (Dialogue Extension via IAT)\n\nInference Anchoring Theory (IAT), developed by Reed and colleagues, extends AIF to handle dialogical argumentation. Adds:\n- **Locution nodes** (L-nodes) — what was actually said\n- **Illocutionary connections** — linking utterances to their propositional content\n- **Transition nodes** — dialogue moves between speakers\n\nThis is critical for Deliberus: real-world deliberation is dialogue, not monologue. Standard AIF only captures argument structure; AIF+/IAT captures the conversation that produced it.\n\n### Available Serializations\n\n| Format | Purpose |\n|--------|---------|\n| OWL-DL | Semantic Web formal ontology |\n| RDF-S (AIF-RDF) | Semantic Web representation |\n| SQL | Database schema (used by AIFdb) |\n| JSON | Web API interchange |\n| DOT | Graph visualization |\n\n**No JSON-LD specification exists yet**, which is a gap. Modern graph databases and APIs expect JSON-LD or similar.\n\n### Who Actually Uses AIF?\n\n**Active users**:\n- **AIFdb** — the canonical implementation (ARG-tech)\n- **OVA/OVA3** — annotation tool (ARG-tech)\n- **Arvina** — dialogue system (ARG-tech)\n- **ArgDF** — Semantic Web implementation using RDF\n- **Carneades** — legal argumentation system (import/export)\n- **Rationale** — argument mapping tool (import/export)\n- **Araucaria** — argument diagramming tool (import/export)\n- **TOAST** — ASPIC+ implementation for computing semantics\n\n**Not using AIF**: Kialo, Pol.is, Consider.it, Habermas Machine, most modern deliberation platforms. This is significant — the academic standard and the practitioner tools occupy different worlds.\n\n### Assessment for Deliberus\n\n**AIF is a living standard but narrowly adopted.** It's the de facto format in computational argumentation research, but no major consumer-facing deliberation platform uses it. The format is well-designed for representing argument structure, but:\n- It's verbose and academic-oriented\n- No JSON-LD serialization\n- The tooling ecosystem is entirely ARG-tech's\n- Modern alternatives (knowledge graphs, LLM-native extraction) may be more practical\n\n**Recommendation**: Design Deliberus's internal ontology independently, but provide AIF export capability for interoperability with the research community. Don't adopt AIF as the primary data model — it would constrain the design without providing practical benefits for end users.\n\n---\n\n## 4. AIFdb and the Argument Web\n\n### What It Is\n\nAIFdb is the database backbone of the \"Argument Web\" — a vision of interconnected argument data analogous to how the World Wide Web connects documents. Built by John Lawrence and Chris Reed.\n\n### Scale\n\n- **Millions of words** annotated across **thousands of arguments**\n- **Tens of thousands of users** (via OVA)\n- The largest structured argumentation corpus in existence\n- Multiple curated sub-corpora available at [corpora.aifdb.org](https://corpora.aifdb.org/)\n\n### API (ws.arg.tech)\n\nREST API with JSON responses. Key endpoints:\n\n| Endpoint | Method | Purpose |\n|----------|--------|---------|\n| `GET nodes/[NodeID]` | GET | Retrieve node (text, type, timestamp) |\n| `GET edges/[EdgeID]` | GET | Retrieve edge (from, to, scheme) |\n| `GET edges/to/[NodeID]` | GET | All edges pointing to a node |\n| `GET edges/from/[NodeID]` | GET | All edges from a node |\n| `GET schemes/all` | GET | All argumentation schemes |\n| `POST json/` | POST | Upload argument map (auth required) |\n| `GET json/[nodeSetID]` | GET | Export argument map as JSON |\n\n**Import/export formats**: JSON, LKIF (Carneades), RTNL (Rationale), RDF, DOT, PNG/SVG.\n\n**Authentication**: POST operations require username/password.\n\n### Example JSON Structure\n\n```json\n{\n  \"nodes\": [\n    {\"nodeID\": \"1\", \"text\": \"Britain should disarm\", \"type\": \"I\", \"timestamp\": \"2012-01-10 14:35:50\"},\n    {\"nodeID\": \"2\", \"text\": \"Nuclear weapons are immoral\", \"type\": \"I\", \"timestamp\": \"2012-01-10 14:36:12\"},\n    {\"nodeID\": \"3\", \"text\": \"\", \"type\": \"RA\", \"timestamp\": \"2012-01-10 14:36:30\"}\n  ],\n  \"edges\": [\n    {\"edgeID\": \"1\", \"fromID\": \"2\", \"toID\": \"3\"},\n    {\"edgeID\": \"2\", \"fromID\": \"3\", \"toID\": \"1\"}\n  ]\n}\n```\n\n### Assessment for Deliberus\n\nAIFdb data could serve as training/evaluation data for Deliberus's argument extraction pipeline. The API is functional but dated (no pagination, no GraphQL, basic auth). Bulk data access would likely require direct collaboration with ARG-tech rather than API scraping.\n\n---\n\n## 5. ARG-tech Software Ecosystem\n\n| Tool | Status | Purpose |\n|------|--------|---------|\n| **OVA3** | Active (2022) | Browser-based argument analysis with collaborative editing, PDF annotation. Open source (GitHub) |\n| **AIFdb** | Active | Argument Web database + search + web services |\n| **AIFdb Corpora** | Active | Corpus management for sets of argument maps |\n| **TOAST** | Active | ASPIC+ implementation for computing acceptability semantics |\n| **OVAgen** | Released | Drawing abstract argumentation frameworks (Dung-style) |\n| **Arvina** | Development | Mixed-initiative dialogue system with protocol specifications |\n| **ArgAnalysisWall** | Released (2012) | Large touch-screen collaborative analysis for broadcast debate |\n| **ArguBlogging** | Released (2012) | Argument-aware blogging platform |\n\n---\n\n## 6. Arg Technica Ltd (Commercial Spinoff, 2025)\n\n**Company**: Arg Technica Ltd, registered in Scotland (SC649381). Cyber Essentials certified.\n\n**Focus**: Defence and security — \"where transparent reasoning, structured analysis, and rapid situational awareness are mission-critical.\"\n\n### Products/Services\n\n1. **Real-Time Reasoning Analysis** — structures ongoing discussions into synchronized rationale overviews\n2. **Risk & Assumption Detection** — surfaces implicit/explicit assumptions automatically\n3. **Bias-Aware Decision Support** — Analysis of Competing Hypotheses (ACH) matrices\n4. **Fact & Evidence Verification** — proposition classification + fact-checking (with offline capability)\n5. **Knowledge Graphs & Argument Maps** — interactive maps of ideas, evidence, conclusions\n6. **Multi-Source Intelligence** — mines external documents, links to live discussion points\n\n### Implications for the Open-Source Ecosystem\n\nThe commercialization focuses on defence/security, which is unlikely to compete with Deliberus's civic/democratic deliberation focus. However, it signals that ARG-tech's most advanced tooling may become proprietary rather than open-source. OVA3 and AIFdb remain open, but new capabilities may land in Arg Technica first.\n\n**Key question for Deliberus**: Will AI4Deliberation deliverables be open-source (as the EU project mandates), or will the commercially interesting parts flow through Arg Technica?\n\n---\n\n## 7. Recent Publications from Reed's Group (2024-2025)\n\n### Argument Mining + LLMs\n\n- **Gemechu & Reed (EMNLP 2024)**: \"External Knowledge-Driven Argument Mining: Leveraging Attention-Enhanced Multi-Network Models\" — using external knowledge to improve argument relation identification\n- **Gemechu, Ruiz-Dolz & Reed (ArgMining 2024)**: \"ARIES: A General Benchmark for Argument Relation Identification\" — standardizing evaluation of argument relation extraction\n- **Ruiz-Dolz & Lawrence (ArgMining 2025)**: \"Detecting Argumentative Fallacies in the Wild: Problems and Limitations of Large Language Models\" — exploring LLM limitations for fallacy detection\n- **Gemechu, Ruiz-Dolz, Lawrence & Reed (ArgMining 2025)**: \"Practical Solutions to Practical Problems in Developing Argument Mining Systems\"\n\n### Corpora and Shared Tasks\n\n- **DialAM-2024** (ACL ArgMining Workshop): First shared task in dialogical argument mining. Six teams participated. Best F1: 67.05% for complete argument map reconstruction. Data from QT30 corpus of broadcast political debate. Uses AIF+/IAT representation with both propositional relations and illocutionary relations.\n- **FORECAST2023** (LREC-COLING 2024): Corpus linking argumentation structure to forecasting accuracy. Explores what features of reasoning lead to better predictions — effectively measuring \"argument quality\" objectively.\n- **RIP Corpus** (LREC-COLING 2024): Collaborative hypothesis-making corpus\n- **NLAS-multi** (Data in Brief): Multilingual corpus of automatically generated natural language argumentation schemes\n\n### Other Notable 2024 Work\n\n- **Ruosch et al. (COMMA 2024)**: \"Toward the Argument Web of Science\" — extending the Argument Web concept to scientific literature\n- **Seyfried et al. (COMMA 2024)**: Psycholinguistic evidence for defining argumentative discourse units as clauses\n- **D'Agostino et al. (IJCAI 2024)**: Question analysis in financial earnings calls\n\n### Key Trend\n\nReed's group is actively combining LLMs with formal argumentation frameworks, but with a cautious empirical approach — testing where LLMs succeed and fail against formal benchmarks rather than wholesale replacing formal methods with LLMs. The DialAM shared task is particularly relevant: it frames argument mining as a structured prediction problem over AIF graphs, not just text classification.\n\n---\n\n## 8. \"From Debate to Deliberation\" (DCI Framework)\n\n**Paper**: \"From Debate to Deliberation: Structured Collective Reasoning with Typed Epistemic Acts\"\n**Author**: Sunil Prakash\n**Date**: March 12, 2026 (arXiv: 2603.11781)\n**Note**: This is NOT from ARG-tech. Independent work, but highly relevant.\n\n**Note on ARGSBASE**: The original research request referenced \"ARGSBASE (EACL 2026)\" as a multi-agent interface for structured human-AI deliberation. No paper by that exact name was found at EACL 2026 or elsewhere. The DCI paper is the closest match to the described concept in the 2025-2026 literature. ARGSBASE may be a working title, a misremembered name, or a paper not yet indexed.\n\n### Architecture\n\nDCI introduces **Deliberative Collective Intelligence** — a structured multi-agent framework where differentiated LLM agents exchange typed reasoning moves.\n\n#### Four Reasoning Archetypes\n\n| Archetype | Symbol | Role |\n|-----------|--------|------|\n| **Framer** | δF | Defines the real problem. Clarifies ambiguity, identifies hidden dimensions, decomposes mixed issues |\n| **Explorer** | δE | Generates novel possibilities and expansions before narrowing |\n| **Challenger** | δC | Pressure-tests everything. Searches for hidden assumptions, weak logic, risks, blind spots |\n| **Integrator** | δI | Combines group thinking into coherent direction. Identifies patterns, synthesizes positions |\n\n#### 14 Typed Epistemic Acts (6 Families)\n\n| Family | Acts | Function |\n|--------|------|----------|\n| **Orienting** | frame, clarify, reframe | Problem definition |\n| **Generative** | propose, extend, spawn | Broaden solution space |\n| **Critical** | ask, challenge | Test assumptions |\n| **Integrative** | bridge, synthesize, recall | Connect ideas |\n| **Epistemic** | ground, update | Anchor confidence |\n| **Decisional** | recommend | Enable closure |\n\nEach move has three layers: speech mode (stance), interaction act (move type), and intent (purpose).\n\n#### DCI-CF Algorithm (8 Stages)\n\n0-2: Initialization, independent proposals, canonicalization\n3-6: Deliberation loop (max R rounds) — evidence collection, revision, scoring, convergence testing\n7: Fallback resolution if no natural convergence (outranking → minimax regret → robust satisficing)\n8: Finalization with structured decision packet\n\n#### Shared Workspace (6 Sections)\n\nOrganizes collective thought (not a transcript accumulator):\n- Problem View\n- Key Frames\n- Emerging Ideas\n- **Tensions** (preserved as first-class objects)\n- Synthesis in Progress\n- Next Actions\n\n#### No Central Moderator\n\nDCI distributes facilitation: the Integrator guides synthesis, the algorithm enforces phase progression, the workspace provides structure. This differs fundamentally from centralized moderation (Habermas Machine) or unstructured debate.\n\n### Evaluation Results\n\n| Metric | Value | Notes |\n|--------|-------|-------|\n| DCI vs. debate (non-routine, n=40) | **+0.95** | 95% CI [+0.41, +1.54], statistically significant |\n| DCI on hidden-profile tasks | **9.56/10** | Best of any system |\n| DCI on routine tasks | **5.39/10** | Worst of all systems |\n| DCI vs. single agent (overall) | **-0.60** | Single agent still wins on average quality |\n| Decision packet completeness | **100%** | |\n| Minority report presence | **98%** | |\n| Token cost | **~237K** | ~62x single agent |\n\n### Comparison to Habermas Machine\n\n| Dimension | DCI | Habermas Machine |\n|-----------|-----|------------------|\n| **Interaction** | Multi-agent typed dialogue | Caucus mediation (no participant interaction) |\n| **Structure** | 14 typed acts + phases | Unstructured opinion → synthesis → ranking |\n| **Dissent** | Preserved (minority reports, tensions) | Flattened into consensus |\n| **Participants** | LLM agents with differentiated roles | Human participants + LLM mediator |\n| **Output** | Decision packet + residual objections + reopen conditions | Single consensus statement |\n| **Theoretical basis** | Habermas communicative action (implemented) | Habermas (named, loosely applied) |\n\n### Relevance for Deliberus\n\nDCI's typed epistemic acts map directly onto argumentation theory concepts. The 14 acts could serve as a pragmatic middle ground between:\n- Full AIF formalism (too complex for real-time interaction)\n- Unstructured LLM debate (too loose for accountability)\n\nThe \"tensions as first-class objects\" principle is particularly valuable — most deliberation tools try to resolve disagreements; DCI preserves them as structural elements. This aligns with Deliberus's vision of making the structure of disagreement visible.\n\n**Token cost concern**: 62x single-agent cost means DCI is expensive. For human-AI deliberation (vs. pure AI-AI), the bottleneck shifts to human response time, making token cost less critical.\n\n---\n\n## 9. Other Relevant Papers (2026)\n\n### \"Learning to Negotiate: Multi-Agent Deliberation for Collective Value Alignment in LLMs\"\n\n**Authors**: Panatchakorn Anantaprayoon et al.\n**Date**: March 11, 2026 (arXiv: 2603.10476)\n\nMulti-agent negotiation framework where two LLM instances with opposing personas engage in structured turn-based dialogue to synthesize mutually beneficial solutions to moral dilemmas. Uses RLAIF with GRPO and external LLM reward model. Achieves collective agency alignment while improving conflict-resolution.\n\n**Relevance**: Shows that structured adversarial dialogue between AI agents can improve value alignment — a mechanism Deliberus could use for testing argument robustness.\n\n### PAKT: Perspectivized Argumentation Knowledge Graph\n\n**Authors**: Plenz, Heinisch, Frank, Cimiano (Heidelberg + Bielefeld)\n**Date**: 2024 (arXiv: 2404.10570)\n\nKnowledge graph for deliberation analysis built on Neo4j. Structures arguments by premises/conclusions, annotates for stances/framings/values, links to ConceptNet. Uses MediaFrames taxonomy and Schwartz value continuum.\n\n**Relevance**: Similar graph-based approach to Deliberus's envisioned architecture. Uses Neo4j (Deliberus uses FalkorDB, which supports Cypher). Does NOT use AIF — chose its own ontology instead. This validates the approach of designing a custom ontology rather than adopting AIF wholesale.\n\n### \"Toward an artificial deliberation? On Google DeepMind's Habermas Machine\"\n\n**Published**: Ethics and Information Technology (Springer), 2025\n\nCritical analysis questioning whether the Habermas Machine truly facilitates deliberation in a Habermassian sense, or merely aggregates and refines existing common ground without fostering deeper deliberative engagement. Raises concerns about whether AI-mediated consensus actually involves reasoned argumentation.\n\n**Relevance**: Validates Deliberus's bet that structured argumentation (not just opinion synthesis) is necessary for genuine deliberation.\n\n---\n\n## 10. Discussion Tracker Corpus\n\n**Paper**: \"The Discussion Tracker Corpus of Collaborative Argumentation\" (LREC 2020)\n**Scale**: 29 multi-party discussions, 985 minutes of audio, 3,261 student turns, 2,128 argumentative turns\n**Domain**: American high school English literature discussions\n\n### Annotations (Three Dimensions)\n\n1. **Argument moves**: Claims (65.3%), Evidence (24.3%), Explanations (10.4%)\n2. **Specificity**: Low, Medium, High\n3. **Collaboration**: Extensions, disagreements about others' ideas\n\n### Training Utility\n\n- Provides benchmarks for multi-task learning across argument dimensions\n- Code available via [discussiontracker.cs.pitt.edu](https://discussiontracker.cs.pitt.edu/)\n- One of the few corpora for synchronous, multi-party argumentation in a real setting\n\n### Assessment for Deliberus\n\n**Moderate utility.** The domain (high school literature) is narrow, but the annotation scheme (claims + evidence + specificity + collaboration) is generalizable. The multi-party, spoken argumentation setting is closer to Deliberus's target than most argument mining corpora (which analyze essays or debate.org posts). Could be useful for training argument move classification, but Deliberus will need its own annotation scheme for the deliberation context.\n\n---\n\n## 11. Practical Implications for Deliberus\n\n### Should Deliberus adopt AIF?\n\n**As the primary format: No.** AIF is designed for academic argument analysis, not for powering a real-time deliberation platform. It's verbose, lacks modern serialization (no JSON-LD), and the tooling ecosystem is entirely ARG-tech's.\n\n**As an export format: Yes.** Providing AIF export would:\n- Enable interoperability with the research community\n- Allow Deliberus data to feed into AIFdb and be used in research\n- Position Deliberus as a contributor to the Argument Web\n- Enable comparison with existing annotated corpora\n\n### Should Deliberus leverage AIFdb data?\n\n**Yes, selectively.** AIFdb's thousands of annotated argument maps are valuable for:\n- Training/evaluating argument extraction models\n- Understanding real argumentation patterns and scheme distributions\n- Benchmarking against the DialAM shared task format\n\nAccess requires collaboration with ARG-tech (API is functional but basic).\n\n### Should Deliberus aim for Argument Web interoperability?\n\n**Eventually, but not as a priority.** The Argument Web vision is compelling but the ecosystem is small. Deliberus should:\n1. Design its own ontology first (informed by AIF but not constrained by it)\n2. Build AIF export as a second-tier feature\n3. Engage with AI4Deliberation consortium — potential collaboration, not competition\n\n### Key Design Lessons from This Research\n\n1. **IAT (AIF+) matters more than base AIF** — real deliberation is dialogue, so the dialogue extension is essential\n2. **Typed epistemic acts** (DCI's 14-act taxonomy) are a more practical unit than AIF's fine-grained node types for real-time interaction\n3. **Tensions/disagreements as first-class objects** (DCI) — don't just model consensus, model the structure of disagreement\n4. **Knowledge graph backend** (PAKT's Neo4j approach) validates Deliberus's FalkorDB choice\n5. **Argumentation schemes** (Walton's taxonomy, used throughout ARG-tech's work) remain the gold standard for classifying reasoning patterns\n6. **LLMs + formal argumentation = the current frontier** — Reed's group is the leading practitioner of this combination\n7. **The DialAM task** (67% F1 for full argument map reconstruction) shows the state of the art — there's significant room for improvement, which means an opportunity for Deliberus to contribute\n\n### Strategic Positioning\n\nDeliberus occupies a unique position:\n- **ARG-tech** builds research tools for academics (OVA, AIFdb) and defence (Arg Technica)\n- **AI4Deliberation** builds institutional tools for governments\n- **Habermas Machine** builds consensus-finding for polls\n- **Dembrane** builds conversation-sense-making for communities\n- **DCI** (Prakash) builds AI-AI deliberation frameworks\n\n**Deliberus's gap**: A platform that makes structured argumentation accessible to citizens for genuine epistemic deliberation — not academic annotation, not opinion polling, not AI-only reasoning. The \"Lean for all knowledge and decision-making\" vision is orthogonal to all of these.\n\n---\n\n## Sources\n\n- [ARG-tech Centre for Argument Technology](https://www.arg.tech/)\n- [Chris Reed's publications](http://arg-tech.org/people/chris/publications/showbyyear.php)\n- [AI4Deliberation CORDIS fact sheet](https://cordis.europa.eu/project/id/101178806)\n- [Arg Technica Ltd](https://www.argtechnica.com/)\n- [AIFdb Web Services](https://www.arg.tech/index.php/research/aifbdb-user-guide/)\n- [AIFdb Corpora](https://corpora.aifdb.org/)\n- [AIF Wikipedia](https://en.wikipedia.org/wiki/Argument_Interchange_Format)\n- [Argument Web Wikipedia](https://en.wikipedia.org/wiki/Argument_Web)\n- [OVA3 (ARG-tech)](https://www.arg.tech/index.php/research/online-argument-visualisation/)\n- [DialAM-2024 overview](https://aclanthology.org/2024.argmining-1.8/)\n- [FORECAST2023 (LREC-COLING 2024)](https://aclanthology.org/2024.lrec-main.652/)\n- [Prakash, \"From Debate to Deliberation\" (arXiv 2603.11781)](https://arxiv.org/abs/2603.11781)\n- [Anantaprayoon et al., \"Learning to Negotiate\" (arXiv 2603.10476)](https://arxiv.org/abs/2603.10476)\n- [PAKT (arXiv 2404.10570)](https://arxiv.org/abs/2404.10570)\n- [Discussion Tracker Corpus (ACL Anthology)](https://aclanthology.org/2020.lrec-1.130/)\n- [Dembrane / AI4Deliberation blog](https://www.dembrane.com/blog/eu-research-ai4deliberation)\n- [\"Toward an artificial deliberation?\" (Ethics & IT, 2025)](https://link.springer.com/article/10.1007/s10676-025-09854-1)\n- [ARG-tech funding page](https://arg-tech.org/index.php/funding/)\n- [LLMs in Argument Mining survey (arXiv 2506.16383)](https://arxiv.org/html/2506.16383v3)\n- [Habermas Machine critique (TRISE)](https://www.trise.net/2026/03/17/habermas-as-a-thinker-against-googles-habermas-machine/)\n"}