{"path":"research/bridging-arguments.md","content":"# Bridging Arguments: The Polis+Argumentation Intersection\n\n*Research compiled March 2026 for the Deliberus project. This document examines the theoretically novel territory at the intersection of opinion-space bridging (Polis) and argument-structure analysis (Kialo/AIF), and develops the concept of \"bridging arguments\" — potentially Deliberus's most distinctive theoretical contribution.*\n\n---\n\n## Table of Contents\n\n1. [Polis Bridging: Current State of the Art](#1-polis-bridging-current-state-of-the-art)\n2. [The Novel Concept: Bridging Arguments](#2-the-novel-concept-bridging-arguments)\n3. [Detecting Bridging Arguments Computationally](#3-detecting-bridging-arguments-computationally)\n4. [Cross-Cutting Arguments in Political Science and Deliberation Research](#4-cross-cutting-arguments-in-political-science-and-deliberation-research)\n5. [Shared Premises as Bridges: The Value Divergence Pattern](#5-shared-premises-as-bridges-the-value-divergence-pattern)\n6. [Technical Architecture for Bridging Arguments](#6-technical-architecture-for-bridging-arguments)\n7. [The Deliberus Synthesis](#7-the-deliberus-synthesis)\n\n---\n\n## 1. Polis Bridging: Current State of the Art\n\n### 1.1 The Mechanics of Bridging Statements\n\nPolis identifies bridging statements through a straightforward extension of its clustering algorithm. After K-means clustering projects participants into opinion groups, the system computes per-group agreement rates for every statement in the conversation. A statement \"bridges\" when its *minimum* per-group agreement rate is high — meaning no identified group strongly rejects it. The key design insight is that this is **group-aware consensus**, not raw popularity: a statement that 60% of everyone agrees with but that one cluster rejects 80% is *not* bridging, even though it is popular. A statement that all clusters agree with at 45% *is* bridging, even if it would look mediocre by raw count.\n\nThis matters because bridging statements are often the most *actionable* product of a Polis deliberation. In Taiwan's vTaiwan Uber consultation (925 participants, 31,115 votes), the bridging statement \"passenger safety must be guaranteed\" achieved ~95% cross-group agreement despite fierce disagreement on almost everything else. That single bridging finding became the policy anchor: new ride-hailing regulations centered on driver ID, fare display, and rating systems — implementing the shared value rather than adjudicating the contested positions.\n\n### 1.2 The Blair et al. (2025) Critique: \"The Structure of Bridging\"\n\nA rigorous analysis by Blair, de Raaij, Procaccia and colleagues at Harvard and the University of Toronto ([Blair et al., 2025](https://www.cs.toronto.edu/~nisarg/papers/bridging.pdf)) identifies a fundamental limitation in Polis's standard bridging metric: it only measures bridging relative to a *fixed partition* of participants. If the clustering produces groups A and B, bridging is computed for that specific split. But different clustering parameters, or different moments in the conversation's evolution, produce different partitions — and a statement that bridges A vs. B might not bridge a different meaningful cleavage C vs. D.\n\nTheir proposed improvements:\n- **Pairwise disagreement bridging**: Instead of measuring relative to the computed clusters, measure across all possible pairings of participants. A statement bridges participant i and participant j if both agree on it while disagreeing on most other things. This extends naturally to a continuous measure of \"how much does this statement bridge genuinely disagreeing people.\"\n- **p-mean bridging**: A family of metrics parameterized by p that allows tuning between \"proportionality\" (bridging proportional to the number of people bridged) and \"connectivity\" (credit for bridging any cleavage at all, even between small groups). The axiomatic foundations are proven stable and interpretable even with sparse vote data.\n\nBoth metrics are more computationally expensive than the standard approach but remain tractable at Polis scale. The deeper contribution is conceptual: they show that \"bridging\" is not a single thing but a family of related properties with different normative justifications.\n\n### 1.3 Ovadya & Thorburn (2023): \"Bridging Systems\"\n\nAviv Ovadya and Luke Thorburn's framework ([Bridging Systems, Knight First Amendment Institute, 2023](https://knightcolumbia.org/content/bridging-systems)) situates Polis within a broader ecosystem of systems designed to reduce destructive divisiveness. They define bridging as producing \"increased mutual understanding and trust across divides, creating space for productive conflict, deliberation, or cooperation.\"\n\nThree signal types operationalize bridging:\n1. **Motifs**: Interaction patterns showing diverse approval — when people who disagree on most things both respond positively to something\n2. **Surveys**: Direct measurement of affective polarization reduction before and after exposure\n3. **Content analysis**: Automated detection of dehumanizing or polarizing language, with bridging systems rewarding content that avoids these signals\n\nThe framework positions Polis as a \"space-based relation model\" — it compresses opinion into 2D and surfaces statements that occupy favorable positions relative to group structure. Community Notes on X (Twitter) is another bridging system in this taxonomy: its [open-source bridging algorithm](https://github.com/twitter/the-algorithm/tree/main/sourcegraph/com/twitter/cr_mixer/src/main/scala/com/twitter/cr_mixer/similarity_engine) requires cross-partisan agreement before displaying a fact-check note. Only ~11% of submitted notes achieve \"helpful\" status by reaching cross-partisan consensus — a brutally high bar that, in practice, may leave high-conflict posts unmoderated ([Birdwatch to Community Notes analysis, 2024](https://arxiv.org/html/2510.09585v2)).\n\n### 1.4 The Fundamental Gap: Bridging Opinions ≠ Bridging Reasoning\n\nThis is the critical limitation for Deliberus. Polis can tell you THAT a statement gets cross-group agreement. It cannot tell you WHY — whether the agreement reflects:\n\n- **Genuine shared values**: Both groups care about X for the same reason\n- **Ambiguity arbitrage**: Both groups interpret the statement differently, each reading their preferred meaning into it\n- **Lowest-common-denominator consensus**: The statement is vague enough that nobody objects, but it doesn't actually commit to anything\n- **Bridging reasoning**: Both groups find the *reasoning structure* compelling, even from different starting positions\n\nThe vTaiwan Uber case illustrates the ambiguity problem. \"Passenger safety\" bridges taxi drivers and ride-hailing advocates — but taxi drivers care about safety as an argument *against* deregulation (unvetted drivers are dangerous), while ride-hailing advocates care about it as an argument *for* app-based ratings and transparency (better accountability than taxis). They agree on the statement but deploy it in opposite directions logically. Polis sees the agreement; it cannot see the divergent reasoning structures underneath.\n\nThis is the gap Deliberus can close.\n\n---\n\n## 2. The Novel Concept: Bridging Arguments\n\n### 2.1 Definition\n\nA **bridging argument** is an argument whose *reasoning structure* — not merely its conclusion — is found compelling by participants across multiple opinion groups.\n\nThis is a strictly stronger condition than a bridging statement. A bridging *statement* requires only that different groups agree with a claim. A bridging *argument* requires that different groups find the *inference from premises to conclusion* compelling, even when starting from different positions or holding different values.\n\nThree distinct types of bridging are worth distinguishing:\n\n**Bridging conclusions**: \"We agree on WHAT.\" Both groups endorse the same claim. This is what Polis measures. Example: Both climate-action advocates and free-market conservatives agree that \"energy efficiency improvements can reduce costs.\" They agree on the conclusion but may have arrived at it through completely different reasoning.\n\n**Bridging premises**: \"We agree on WHY.\" Different groups share key premises even when they disagree on ultimate conclusions. Example: Both progressive and conservative participants believe that \"government programs often create unintended incentives\" — but progressives use this premise to argue for better program design, while conservatives use it to argue for fewer programs. The premise bridges despite the conclusions diverging.\n\n**Bridging reasoning patterns**: \"We agree on HOW to think about this.\" Different groups accept the same logical or inferential structure, the same framework for evaluating evidence, or the same type of argument. Example: Both groups accept consequentialist cost-benefit reasoning as a valid method for evaluating policy — they disagree on which costs and benefits matter most, but they agree that the argument form is legitimate.\n\n### 2.2 A Worked Example: Carbon Taxes\n\nConsider a carbon tax debate. A climate-focused participant might argue:\n\n> *Carbon taxes are effective because they use price signals to redirect behavior — people respond to costs, so making carbon expensive will reduce carbon-intensive choices without requiring regulators to dictate which specific changes to make.*\n\nA free-market participant would typically be expected to *oppose* carbon taxes. But look at the reasoning structure: it appeals to price signals, behavioral response to costs, and the superiority of market mechanisms over regulatory dictates. This is an argument built from the conceptual vocabulary of free-market economics. A free-market participant who opposes carbon taxes on grounds of government intervention might nonetheless find this reasoning pattern *compelling on its own terms* — and might even say \"I disagree with your conclusion but I find your argument the most honest version of the pro-tax case.\"\n\nThis \"I disagree with your conclusion but find your reasoning compelling\" signal is the core of bridging arguments. It is invisible to Polis (which only sees the vote on the conclusion) but is potentially the richest signal for productive deliberation. It means:\n- The arguer has genuinely engaged with the other side's conceptual framework\n- There may be a more targeted disagreement about premises (the free-market participant might say \"yes, but carbon taxes create bureaucracy overhead\" — an empirical dispute — or \"the right response to carbon externalities is liability law, not taxation\" — a more principled disagreement that uses the same reasoning type)\n- The disagreement is potentially more tractable than it appears, because it is located at a specific, debatable point rather than in a wholesale framework rejection\n\n### 2.3 The \"Argument Quality\" vs. \"Argument Conclusion\" Distinction\n\nThe distinction maps onto a well-known finding in argumentation theory and social psychology: **people reliably distinguish between finding an argument well-reasoned and finding an argument's conclusion correct**. This is sometimes called \"argument strength\" vs. \"argument agreement.\"\n\nAn argument can be:\n- **Well-reasoned + conclusion accepted**: Strengthens conviction\n- **Well-reasoned + conclusion rejected**: Creates cognitive tension, the most productive state for genuine opinion change — \"I can't immediately refute this, which makes me reconsider\"\n- **Poorly reasoned + conclusion accepted**: Preaching to the choir; no epistemic value\n- **Poorly reasoned + conclusion rejected**: Easy dismissal; no deliberative value\n\nThe \"well-reasoned + conclusion rejected\" quadrant is where Deliberus operates. It is where honest intellectual engagement with opposing views happens. Polis cannot distinguish this from any of the other three — it only sees \"disagree on conclusion.\" Deliberus, by separately recording both dimensions, can identify arguments in this quadrant and surface them as the richest sites of genuine deliberation.\n\n---\n\n## 3. Detecting Bridging Arguments Computationally\n\n### 3.1 The Basic Architecture: Combining Opinion Matrices with Argument Graphs\n\nThe natural approach is to combine Polis-style opinion clustering with a QBAF-style argument graph. This requires:\n\n1. **Build an opinion matrix** (participants × positions) and run PCA + clustering to identify opinion groups, exactly as Polis does\n2. **Build an argument graph** (claims, premises, attack/support relations, evidence) using structured argumentation\n3. **For each argument in the graph**, compute cross-cluster \"logical compellingness\" scores — distinct from agreement scores on the conclusion\n4. **Surface arguments** with high cross-cluster compellingness scores as bridging arguments\n\nThe critical new element is step 3: a separate compellingness voting axis.\n\n### 3.2 Separate Voting Axes: The Key Design Decision\n\nDeliberus must implement dual voting axes for each argument:\n\n- **Axis 1 (Conclusion)**: Do you agree or disagree with this claim?\n- **Axis 2 (Reasoning)**: Regardless of whether you agree with the conclusion, do you find this argument well-reasoned or poorly-reasoned?\n\nThis separation is the mechanism that makes bridging arguments detectable. A participant in Group A (pro-climate-action) who votes \"well-reasoned\" on a carbon-tax argument built from free-market principles is providing the same signal as a Group B (free-market) participant who votes \"well-reasoned\" on it — they both find the *reasoning pattern* compelling. The cross-group agreement on reasoning quality is the bridging signal, even when they diverge on the conclusion vote.\n\nThis design is not merely theoretical. Kialo uses a binary pro/con structure that implicitly conflates \"I think this argument supports the position\" with \"I agree with the argument.\" Some deliberation platform designs have explored multi-dimensional rating, but none have implemented the specific separation of conclusion vs. reasoning quality as a first-class data structure aimed at bridging detection.\n\n### 3.3 The Argument Convincingness Literature\n\nThe NLP community has studied argument convincingness prediction since the mid-2010s. Habernal and Gurevych's foundational work established the UKPConvArg corpus ([ACL 2016](https://aclanthology.org/P16-1150/)) — 16,000 pairwise argument comparisons annotated for which argument is \"more convincing\" — and demonstrated that bidirectional LSTM models could predict convincingness with 0.76-0.78 accuracy across topics. Their follow-up work ([EMNLP 2016](https://aclanthology.org/D16-1129/)) identified key attributes of convincingness: specific evidence, concrete examples, acknowledgment of counterarguments, and avoidance of ad hominem.\n\nThe crucial limitation of this body of work for Deliberus: convincingness is measured as a crowd-aggregate judgment, not a cross-group judgment. If one ideological group dominates the annotation pool, the \"convincing\" label will reflect their standards. Durmus and Cardie ([NAACL 2018](https://aclanthology.org/N18-1094/)) showed that prior beliefs matter more than language features for persuasion — reader ideology predicts persuasion outcomes better than argument text features alone. This is both a warning and a pointer: **the same argument will be rated \"convincing\" or \"unconvincing\" by different groups for different reasons**. A system that ignores group structure conflates these signals.\n\nMore recently, El Baff, Al Khatib, et al. ([EMNLP 2024](https://aclanthology.org/2024.findings-emnlp.265/)) studied how to computationally improve argument effectiveness across ideologies using instruction-tuned LLMs. They found that LLMs can rewrite arguments to be more effective for specific ideological audiences — a capability that cuts both ways for Deliberus. Used constructively, LLMs could help users identify which *framings* of their arguments appeal across group lines. Used destructively, it enables targeted rhetorical manipulation. The design question for Deliberus is whether this LLM capability should be a tool for individual users or a system-level function for surfacing cross-group framings.\n\n### 3.4 The PAKT Framework: Perspectivized Argumentation Knowledge Graph\n\nThe most directly relevant recent research for Deliberus's architecture is PAKT (Perspectivized Argumentation Knowledge Graph and Tool), published by Plenz, Heinisch, Frank, and Cimiano at Heidelberg in 2024 ([arxiv:2404.10570](https://arxiv.org/abs/2404.10570)). PAKT structures the argumentative space across topics by segmenting arguments into *premises and conclusions*, annotating them for *stances, framings, and underlying values*, and connecting them to background knowledge.\n\nThe PAKT data model is:\n```\nArgument → {\n  premises: [Claim],\n  conclusion: Claim,\n  frames: [Frame],         # rhetorical frames (e.g., economic, safety, liberty)\n  values: [Value],          # underlying value commitments\n  stance: Polarity,\n  stakeholder_group: Group\n}\n```\n\nThis is essentially a formalization of the \"bridging premises\" concept. By annotating which values and frames each argument relies on, PAKT enables queries like: \"Find arguments from Group A and Group B that share value annotations V and frame annotations F despite having opposing stances on conclusion C.\" Those are the bridging argument candidates.\n\nPAKT does not yet implement cross-group bridging as a first-class search function — it provides the data model that would make this possible. Deliberus could extend this approach by:\n- Running opinion clustering on PAKT-annotated argument data\n- Computing cross-cluster overlap in value and frame annotations per argument\n- Surfacing arguments with high cross-cluster frame/value overlap as bridging candidates\n\n### 3.5 The Deliberation Knowledge Graph (2025)\n\nClosely related is the Deliberation Knowledge Graph ([EGOVIS 2025](https://link.springer.com/chapter/10.1007/978-3-032-02225-7_9)) — a system integrating deliberation processes, arguments, and participants across institutional and civic spheres (European Parliament, civic platforms, public forums). It provides a Deliberation Ontology as a shared data model and demonstrates cross-platform argument integration.\n\nThe interest for Deliberus is primarily architectural: the Deliberation Knowledge Graph shows that argument-level knowledge graphs are technically feasible at parliamentary scale, and that connecting argument nodes across different deliberative contexts (parliamentary and civic) is achievable. The approach to cross-institutional bridging in that work parallels what Deliberus needs for cross-group bridging.\n\n---\n\n## 4. Cross-Cutting Arguments in Political Science and Deliberation Research\n\n### 4.1 Deliberative Polling: The Best Evidence We Have\n\nJames Fishkin's Deliberative Polling is the most rigorous empirical test of whether structured exposure to arguments changes opinions across political groups. The basic design: a representative random sample of citizens (typically 200-500) receives balanced briefing materials, deliberates in small groups, hears from competing experts, and takes pre- and post-surveys.\n\nThe consistent findings are striking:\n- **Opinion movement is substantial**: In Fishkin's \"America in One Room\" (2019), participants moved toward consensus on 22 of 26 highly polarized policy proposals; movements were statistically significant in 19 of these\n- **Affective polarization decreases**: Democrats rated Republicans 13 points higher on a 100-point thermometer after deliberation; Republicans rated Democrats 14 points higher\n- **~70% of participants changed their positions** on at least one major issue after exposure to balanced arguments and expert Q&A\n\nWhat drives these changes? Fishkin's careful design isolates the mechanism: exposure to *balanced information* and *arguments from expert representatives of opposing positions*, combined with facilitated small-group discussion. The key is not merely hearing opposing views but engaging with the *best versions* of opposing arguments, not strawmen.\n\nThis is exactly what a bridging argument architecture enables: it surfaces the opposing group's *strongest arguments* (those rated well-reasoned even by the opposition), not their weakest ones. The Deliberative Polling mechanism — structured engagement with the best version of the opposing case — could be computationally approximated by showing users the arguments that cross-group participants rate as \"well-reasoned.\"\n\nFishkin's [\"Deliberation Can Save Democracy\" (2021)](https://www.persuasion.community/p/deliberation-can-save-democracy) articulates the underlying theory: if people think their voice actually matters, they will engage seriously with information, ask tough questions, and think carefully. The bridging argument interface creates exactly this condition — it shows users that their reasoning evaluation matters, not just their positional vote.\n\n### 4.2 Deep Canvassing: Reasoning vs. Assertion\n\nBroockman and Kalla's deep canvassing research ([Science, 2016](https://www.science.org/doi/10.1126/science.aad9713)) provides compelling evidence about what makes cross-group persuasion work — and it is not presenting facts or making assertions. A single ~10-minute door-to-door conversation using deep canvassing techniques produced durable reductions in transphobia, with effects larger than the average change in American homophobia attitudes from 1998 to 2012.\n\nThe mechanism is not argumentation in the traditional sense. Instead: ask questions, listen sincerely, ask more questions. Invite the person to remember when they were treated unfairly. Non-judgmentally exchange narratives. The key counterintuitive finding: **presenting facts and data does not work; narrative engagement with shared experiential premises does**.\n\nThis creates a design tension for Deliberus. The formal argumentation model assumes that reasoning quality matters and that people respond to logical structure. Deep canvassing suggests that the engagement mechanism is relational and narrative, not logical. The resolution may be:\n- **For initial position-taking**: the argument graph and reasoning-quality signals work for people already in intellectual engagement mode\n- **For attitude change at group boundaries**: the facilitated narrative exchange that deep canvassing uses may be necessary\n- **Bridging arguments as conversation starters, not conversation enders**: an argument identified as \"well-reasoned by both groups\" becomes an invitation to explore why — which is precisely the kind of open question that deep canvassing uses to open productive dialogue\n\nCrucially, Broockman and Kalla found that canvassers need to be \"genuinely open-minded\" — not persuaders deploying rhetorical techniques, but listeners actively seeking to understand. This maps onto Deliberus's role: the platform doesn't *argue* for bridging positions; it *reveals* where reasoning resonates across groups, and lets participants follow that thread.\n\n### 4.3 Cross-Pressured Positions and Political Science\n\nPolitical scientists have long studied \"cross-cutting\" cleavages — political divisions that cut across rather than reinforce each other, creating \"cross-pressured\" voters who hold conservative views on some dimensions and liberal views on others. The classical finding (Berelson, Lazarsfeld, McPhee, 1954; updated by Mutz, 2002) is that cross-pressured voters are *less certain* about their positions, *less participatory*, but more *open to persuasion*. High cross-pressure correlates with delayed vote decisions and ticket-splitting.\n\nThe implication for Deliberus: participants who hold internally cross-cutting positions (agree with Group A on some things and Group B on others) are precisely the participants most likely to find bridging arguments compelling. They already have evidence in their own belief system that the two groups' frameworks can coexist. A good platform design would identify these cross-pressured participants and use them as a \"bridging cohort\" — their judgment of argument quality is particularly diagnostic for bridging potential.\n\n### 4.4 The Negotiation Framework: Interests vs. Positions\n\nFisher and Ury's principled negotiation framework (*Getting to Yes*, 1981) makes a distinction that directly illuminates bridging arguments: the difference between **positions** (what people say they want) and **interests** (the underlying needs and concerns driving the positions).\n\nClassic negotiation example: two people want the same orange. Position A: \"I want the orange.\" Position B: \"I want the orange.\" Compromise: cut it in half. But if you ask *why* — Person A wants to eat the fruit, Person B wants the peel for baking — they can both get 100% of what they actually need.\n\nIn deliberation, this maps precisely onto the bridging argument concept:\n- **Positions** = the conclusions participants vote on (\"we should/shouldn't implement carbon taxes\")\n- **Interests** = the underlying concerns and values driving those positions (economic efficiency, climate protection, liberty, equity)\n\nArguments built from *interest language* rather than *position language* have higher bridging potential because they address the underlying concerns that different groups share. An argument like \"Carbon pricing creates the right economic incentives to direct innovation toward emissions reduction\" is positioned in interest-space (economic incentives, innovation) rather than position-space (tax yes/no). Groups with different positions can both engage with the interest-level argument.\n\nThe PAKT framework operationalizes this by explicitly annotating argument \"values\" — the interest-level concerns that each argument addresses. Deliberus could use value annotations as a bridging proxy: arguments annotated with values that appear in both groups' vocabularies are bridging candidates, regardless of whether the groups agree on the conclusion.\n\n---\n\n## 5. Shared Premises as Bridges: The Value Divergence Pattern\n\n### 5.1 The Structural Pattern\n\nEven in deep political disagreement, empirical analysis consistently finds substantial shared premises. The disagreement often lives not in the factual beliefs or even the fundamental values, but in *how much weight* different values receive.\n\nMoral Foundations Theory (Haidt & Graham, 2009; [Graham et al., JPSP 2009](https://fbaum.unc.edu/teaching/articles/JPSP-2009-Moral-Foundations.pdf)) provides the clearest empirical characterization. The six moral foundations (Care/Harm, Fairness/Cheating, Loyalty/Betrayal, Authority/Subversion, Sanctity/Degradation, Liberty/Oppression) are universal — every human uses all six. What differs across political groups is the *weight* assigned to each foundation in moral reasoning.\n\nLiberals consistently weight Care/Harm and Fairness most heavily. Conservatives weight all six foundations more equally, including Loyalty, Authority, and Sanctity that liberals largely downweight. A significant 2022 study ([American Political Science Review](https://www.cambridge.org/core/journals/american-political-science-review/article/abs/liberals-and-conservatives-rely-on-very-similar-sets-of-foundations-when-comparing-moral-violations)) found that when controlling for comparison type, liberals and conservatives rely on very similar foundations — the apparent divergence in Haidt's original work was partly a methodological artifact.\n\nThis has a profound implication for Deliberus: **the disagreement structure is often not \"different values\" but \"different value weights\"**. Making this visible transforms the deliberation.\n\n### 5.2 The Value-Premise Divergence Pattern\n\nConsider a policy debate about immigration enforcement. The argument structure, when fully decomposed, might look like this:\n\n*Premises shared by both sides*:\n- P1: \"Orderly migration systems are preferable to chaotic ones\" (both agree)\n- P2: \"People deserve humane treatment regardless of legal status\" (both agree, though may weight differently)\n- P3: \"Enforcement resources are limited\" (both agree)\n\n*Diverging value weights*:\n- Group A weights P1 heavily → supports strong enforcement as orderly\n- Group B weights P2 heavily → opposes enforcement that separates families\n\n*Different empirical assessments*:\n- Group A believes enforcement deters illegal migration (empirical, potentially verifiable)\n- Group B believes enforcement fails to deter while creating humanitarian costs (empirical, potentially verifiable)\n\n*Different institutional trust*:\n- Group A trusts government enforcement agencies to act humanely\n- Group B does not trust enforcement agencies based on documented abuses\n\nWhen this structure is made explicit, the deliberation becomes more tractable. The shared premises (P1-P3) are not contested and can be established as common ground. The divergence lives in three separable places: value weights, empirical beliefs (which evidence could adjudicate), and institutional trust (which is a different kind of claim). A Deliberus argument graph makes this structure explicit; the implicit debate conflates all three and produces intractable gridlock.\n\nThis is Deliberus's most distinctive contribution to the deliberation landscape: **not just finding what people agree on, but mapping *where exactly* their reasoning diverges** — so that deliberation can be targeted at the actual crux rather than thrashing around at the surface level of positions.\n\n### 5.3 The Habermas Machine vs. the Deliberus Approach\n\nThe Habermas Machine ([Tessler et al., Science 2024](https://www.science.org/doi/10.1126/science.adq2852)) finds common ground by optimizing for statements that participants approve of. This is a consensus approach: it finds the formulations people can agree on and presents them as the group's position.\n\nDeliberus's approach is different. Where the Habermas Machine seeks consensus, Deliberus seeks **clarity about the structure of disagreement**. The Habermas Machine says \"here is what we agree on.\" Deliberus says \"here is *precisely where* we disagree, and here is what our reasoning has in common even across that disagreement.\"\n\nThese are not just philosophical preferences — they produce different outcomes. Consensus optimization (Habermas Machine) tends toward vague statements that everyone can endorse because they are ambiguous enough to carry multiple meanings. Structured disagreement mapping (Deliberus) tends toward precise identification of the contested crux: the specific empirical claim, value weight, or institutional trust assessment where the reasoning diverges.\n\nFor policy-making, vague consensus is often worse than explicit disagreement. A policy built on \"we all care about safety\" that papers over a genuine disagreement about how to achieve safety will fail in implementation, because the disagreement reasserts itself when the details must be specified. A policy built on \"here is the specific tradeoff between enforcement effectiveness and humanitarian costs, and here is how we weigh it\" has a chance of being implementable because the disagreement has been located and addressed, not suppressed.\n\n### 5.4 Deliberative Polling Confirms: Precision Enables Movement\n\nFishkin's Deliberative Polling results are consistent with the \"precise disagreement\" thesis: when participants engage with the *best version* of opposing arguments, delivered by expert representatives in Q&A, substantial opinion movement results. The mechanism is not that people are persuaded by rhetoric — it is that they discover their disagreement is more tractable than they thought. They find that some of their empirical beliefs were wrong, or that the opposing group shares more of their values than they assumed, or that their position relied on assumptions that don't hold up to scrutiny.\n\nThis is exactly what bridging argument analysis enables at scale: instead of expert Q&A sessions for small groups, a computational system can identify, for millions of participants, which arguments cross-group participants find well-reasoned — pointing exactly toward the arguments most likely to produce Fishkin-style precision in understanding the actual disagreement structure.\n\n---\n\n## 6. Technical Architecture for Bridging Arguments\n\n### 6.1 The QBAF Backbone\n\nQuantitative Bipolar Argumentation Frameworks (QBAFs) provide the most suitable formal backbone for bridging argument computation. In a QBAF, arguments have:\n- An *intrinsic strength* (base weight, not accounting for attack/support relations)\n- *Attack relations* from other arguments (reduce strength)\n- *Support relations* from other arguments (increase strength)\n- A *final strength* computed from the above via gradual semantics\n\nFor Deliberus, the QBAF can be extended with a *per-group strength vector* instead of a single intrinsic strength. Each argument node stores not one number but a vector over opinion groups:\n\n```\nArgument = {\n  intrinsic_strength: Float (default prior)\n  group_conclusion_votes: Map[Group → AgreementScore]     # \"do you agree with this claim?\"\n  group_reasoning_votes: Map[Group → CompellingnessScore] # \"is this argument well-reasoned?\"\n  support_relations: [Argument]\n  attack_relations: [Argument]\n}\n```\n\nA *bridging argument score* for argument A across groups G₁ and G₂ is then:\n\n```\nbridging(A, G₁, G₂) = min(reasoning_score(A, G₁), reasoning_score(A, G₂))\n```\n\nAggregated across all group pairs:\n\n```\nbridging(A) = harmonic_mean over all (Gᵢ, Gⱼ) pairs of: min(reasoning_score(A, Gᵢ), reasoning_score(A, Gⱼ))\n```\n\nUsing harmonic mean rather than arithmetic mean penalizes arguments that are compelling to some groups but not others — a true bridging argument must be compelling *across* groups, not merely averaging well.\n\nThe aggregative semantics work in QBAFs ([arxiv:2603.06067](https://arxiv.org/abs/2603.06067)) extends this naturally: attack and support relations can themselves have group-specific weights, modeling that an objection considered decisive by one group may be irrelevant to another. The full QBAF propagation then computes group-specific final strengths for each argument, and bridging scores are computed over these final strengths rather than intrinsic strengths.\n\n### 6.2 Cross-Referencing Opinion Matrix with Argument Graph\n\nThe full architecture combines two data structures:\n\n**Opinion matrix** (Polis-style):\n- Rows: participants\n- Columns: high-level positions/conclusions\n- Values: agree/disagree/pass\n- Output: cluster membership for each participant (opinion groups G₁, G₂, ..., Gₖ)\n\n**Argument graph** (QBAF/PAKT-style):\n- Nodes: claims, evidence, premises\n- Edges: support/attack relations with strength weights\n- Per-node annotations: values, frames, sources\n- Per-node vote vectors: {conclusion_vote, reasoning_vote} × {participant}\n\nCross-referencing these structures:\n1. Use opinion matrix clustering to assign each participant to a group\n2. Aggregate per-participant votes in the argument graph into per-group vectors\n3. Compute bridging scores for each argument node using per-group reasoning votes\n4. The *bridging argument report* — analogous to Polis's bridging statement report — surfaces the top-ranked bridging arguments with their supporting evidence for why they bridge\n\n### 6.3 The \"Reasoning Pathway\" Feature\n\nA bridging argument is not just a node — it is a *path* through the argument graph. For Deliberus, a particularly valuable output is identifying *bridging reasoning pathways*: sequences of argument steps where each step is well-reasoned according to cross-group judgment, leading from a shared premise to one of the contested conclusions.\n\nThis can be computed as a graph path search: find paths through the argument graph where all edges (each inferential step) have high cross-group compellingness scores. These paths represent the most \"epistemically honest\" routes from shared ground to contested conclusions — the arguments most likely to produce genuine engagement rather than rejection.\n\nWorked example in a climate debate:\n- Shared premise: \"Greenhouse gas emissions cause warming\" (high cross-group acceptance, well-documented)\n- Bridging inference: \"Economic instruments are more efficient than command-and-control regulation\" (high cross-group reasoning acceptance from free-market frame)\n- Contested conclusion fork: Path A leads to \"Therefore, carbon pricing\" (supported by Group B); Path B leads to \"Therefore, carbon liability law\" (preferred by Group A)\n\nThe bridging pathway analysis reveals: the two groups share premises and reasoning methods up to the fork, where they diverge on *mechanism*. The disagreement is not about climate science or even about whether economic instruments are good — it is specifically about *which* economic instrument to use. That is a much more tractable and productive disagreement to have.\n\n### 6.4 LLM-Assisted Argument Mining for Bridging Detection\n\nPractical implementation requires extracting argument structure from natural language contributions. The BIRD framework ([Feng et al., ICLR 2025](https://arxiv.org/abs/2404.12494)) is directly applicable: decompose LLM processing into Abduction (identify relevant factors), Entailment (check which are supported by evidence), and Deduction (compute calibrated strength). For bridging argument mining:\n\n1. **Abduction**: LLM identifies the premises and value assumptions underlying a participant's contribution\n2. **Value annotation**: LLM tags identified premises with Moral Foundation or value-type labels (from PAKT-style ontology)\n3. **Cross-group matching**: System queries: \"Which arguments from Group B share value annotations V with this argument from Group A?\"\n4. **Reasoning quality prediction**: LLM (or fine-tuned model) predicts cross-group reasoning quality scores using techniques from Habernal & Gurevych (2016) extended with group-specific training signal\n\nEl Baff et al.'s (EMNLP 2024) finding that LLMs can improve argument effectiveness across ideologies suggests these models already have latent representations of what makes arguments resonate with different ideological audiences. Deliberately extracting and making these representations visible — rather than using them to silently rewrite arguments — is the Deliberus approach.\n\n### 6.5 The \"Conditional Compellingness\" Formulation\n\nThe most computationally interesting formulation of bridging arguments is as a *conditional* query:\n\n> \"If you believe premises P₁ and P₂ [which we know you accept], is argument A compelling?\"\n\nThis is distinct from asking \"do you find argument A compelling?\" unconditionally — it isolates the inferential step from the premises rather than the complete package including the premises. Two participants can both say \"yes, if P₁ and P₂ then A follows\" while disagreeing on whether P₁ and P₂ are true.\n\nThis is a form of *conditional compellingness*: the argument is compelling conditional on the premises. Surfacing that participants agree on conditional compellingness, while disagreeing on the premises themselves, precisely locates where the deliberation needs to focus — on the premises, not on the inference pattern.\n\nImplementation: when a participant evaluates an argument, the interface presents the premises explicitly (\"assuming A and B, does C follow?\") rather than presenting the complete argument as a single unit. This is cognitively more demanding than a single vote, but it is the information needed to detect bridging at the inferential level rather than the positional level.\n\n---\n\n## 7. The Deliberus Synthesis\n\n### 7.1 What No Existing Platform Does\n\nA survey of the current landscape confirms that no production system combines these capabilities:\n\n| Platform | Opinion Clustering | Argument Structure | Bridging Detection | Cross-Group Reasoning |\n|----------|-------------------|-------------------|-------------------|----------------------|\n| Polis | ✅ (PCA + K-means) | ❌ | ✅ (conclusions only) | ❌ |\n| Kialo | ❌ | ✅ (binary pro/con) | ❌ | ❌ |\n| Community Notes | ✅ (bridging algo) | ❌ | ✅ (fact claims) | ❌ |\n| Habermas Machine | ❌ | ❌ | ✅ (consensus) | ❌ |\n| PAKT | ❌ | ✅ (values/frames) | ❌ | ❌ |\n| **Deliberus** | ✅ | ✅ | ✅ | **✅ (novel)** |\n\nThe \"Cross-Group Reasoning\" column — measuring which arguments are found well-reasoned *across* opinion groups rather than just where conclusions converge — is the theoretical gap Deliberus fills.\n\n### 7.2 The Research Foundation\n\nThe bridging argument concept has genuine theoretical novelty but rests on solid foundations:\n\n**From political science**: Deliberative Polling demonstrates that structured exposure to the best opposing arguments produces substantial, durable opinion movement. Bridging arguments are the computational implementation of this mechanism at scale.\n\n**From argumentation theory**: QBAFs and the PAKT framework provide the formal structures needed for per-group argument strength and value-annotated argument nodes. The mathematics of bridging in these frameworks is well-developed.\n\n**From NLP**: Argument convincingness prediction (Habernal & Gurevych) and cross-ideological effectiveness improvement (El Baff et al.) demonstrate that the necessary NLP capabilities exist. The gap is in applying them with *group-aware* signal rather than population-aggregate signal.\n\n**From negotiation theory**: Fisher and Ury's interest/position distinction maps directly onto the premise/conclusion distinction in argument structure, and provides a well-tested practical framework for why interest-level bridging is more productive than position-level agreement.\n\n**From cognitive psychology**: Moral Foundations Theory explains why different groups find different argument types compelling — they are weighting different values. Making these weightings explicit transforms apparent value conflicts into tractable disagreements about priorities.\n\n### 7.3 The Core Product Claim\n\nDeliberus's bridging argument feature produces something neither Polis nor any argumentation platform produces: a computational answer to the question **\"What reasoning do we share, and exactly where does it diverge?\"**\n\nThe answer takes the form of:\n- **Bridging argument cards**: Arguments rated as well-reasoned by participants across opinion groups, with their cross-group compellingness scores shown\n- **Shared premise maps**: The premises that appear in arguments across groups, making visible what everyone already agrees on\n- **Divergence points**: Precisely where the argument graph branches across groups — the specific empirical claim, value weight, or institutional trust judgment where reasoning paths separate\n- **Bridging pathway highlights**: The sequences of reasoning steps that most groups find sound, leading up to the divergence point\n\nThis is different from the Habermas Machine's \"statement we can all endorse\" and from Polis's \"claim we all agree with.\" It is a *map of the reasoning space* that shows not just where consensus exists but where it does not — and why.\n\n### 7.4 Design Implications (Without Deciding Architecture)\n\nSeveral design implications follow from this analysis, without committing to specific architectural choices:\n\n**Dual-axis voting is load-bearing**: The entire bridging argument detection system depends on separately capturing reasoning quality and conclusion agreement. This must be a first-class data structure, not an afterthought. The interface design challenge is making this feel natural rather than burdensome.\n\n**Clustering precedes bridging analysis**: Users must first be clustered by opinion patterns (Polis-style) before bridging scores can be computed across clusters. This means some version of the opinion matrix is needed even in a primarily argumentation-focused platform.\n\n**Progressive depth matters more here**: The dual-axis vote is more cognitively demanding than a simple agree/disagree. The platform should offer the simple version (conclusion vote only) as default, with the reasoning quality vote as an optional deeper engagement for users willing to invest more effort.\n\n**AI-assisted structuring is necessary**: Extracting argument premises, annotating values, and computing bridging scores requires computational support. Users cannot be expected to manually annotate every argument with value labels. LLM-assisted extraction (Claimify-style, extended) is the mechanism.\n\n**The bridging argument report is a standalone product**: Even without the full deliberation platform, a system that takes a contested topic, collects arguments from participants across opinion groups, and produces a bridging argument report would be a significant contribution to governance, policy-making, and public deliberation. This could be Deliberus's MVP.\n\n**The \"I disagree but find this compelling\" signal is gold**: Designing the UX to make it easy and rewarding to signal \"I disagree with this conclusion but find the reasoning sound\" — and making clear to users that this is a valuable contribution — is critical to getting the cross-group compellingness data. Users need to understand that their honest intellectual engagement, even with arguments they oppose, is exactly what makes the system work.\n\n---\n\n## References and Sources\n\n### Foundational Bridging Research\n- Blair, C., de Raaij, J., Procaccia, A.D., et al. (2025). [The Structure of Bridging](https://www.cs.toronto.edu/~nisarg/papers/bridging.pdf). Harvard / University of Toronto.\n- Ovadya, A. & Thorburn, L. (2023). [Bridging Systems: Open Problems for Countering Destructive Divisiveness](https://knightcolumbia.org/content/bridging-systems). Knight First Amendment Institute.\n- Small, C.T., Bjorkegren, M., et al. (2021). [Polis: Scaling Deliberation by Mapping High Dimensional Opinion Spaces](https://www.e-revistes.uji.es/index.php/recerca/article/view/5516/6558). *Recerca* 26(2).\n\n### Argument Convincingness and Cross-Ideological Effectiveness\n- Habernal, I. & Gurevych, I. (2016). [Which argument is more convincing? Analyzing and predicting convincingness of Web arguments using bidirectional LSTM](https://aclanthology.org/P16-1150/). ACL 2016.\n- Habernal, I. & Gurevych, I. (2016). [What makes a convincing argument? Empirical analysis and detecting attributes of convincingness in Web argumentation](https://aclanthology.org/D16-1129/). EMNLP 2016.\n- Durmus, E. & Cardie, C. (2018). [Exploring the Role of Prior Beliefs for Argument Persuasion](https://aclanthology.org/N18-1094/). NAACL 2018.\n- El Baff, R., Al Khatib, K., Alshomary, M., Konen, K., Stein, B., & Wachsmuth, H. (2024). [Improving Argument Effectiveness Across Ideologies using Instruction-tuned Large Language Models](https://aclanthology.org/2024.findings-emnlp.265/). EMNLP 2024 Findings.\n\n### PAKT and Deliberation Knowledge Graphs\n- Plenz, M., Heinisch, P., Frank, A., & Cimiano, P. (2024). [PAKT: Perspectivized Argumentation Knowledge Graph and Tool for Deliberation Analysis](https://arxiv.org/abs/2404.10570). arXiv:2404.10570. [GitHub](https://github.com/Heidelberg-NLP/PAKT)\n- [A Deliberation Knowledge Graph: Bridging Institutional and Civic Democratic Discourse](https://link.springer.com/chapter/10.1007/978-3-032-02225-7_9). EGOVIS 2025.\n\n### QBAF and Formal Argumentation\n- [Aggregative Semantics for Quantitative Bipolar Argumentation Frameworks](https://arxiv.org/abs/2603.06067). arXiv:2603.06067 (March 2026).\n- Bikarke, A. (2024). [Social Argumentation Systems](https://discovery.ucl.ac.uk/10208188/1/SocialArgumentation.pdf). UCL.\n\n### Deliberative Democracy and Cross-Group Evidence\n- Tessler, M.H., Bakker, M.A., et al. (2024). [AI can help humans find common ground in democratic deliberation](https://www.science.org/doi/10.1126/science.adq2852). *Science* 386.\n- Fishkin, J. (2021). [Deliberation Can Save Democracy](https://www.persuasion.community/p/deliberation-can-save-democracy). *Persuasion*.\n- [Is Deliberation an Antidote to Extreme Partisan Polarization? (America in One Room)](https://preprints.apsanet.org/engage/api-gateway/apsa/assets/orp/resource/item/5f3da4bf10cdbe00192da9dd/original/is-deliberation-an-antidote-to-extreme-partisan-polarization-reflections-on-america-in-one-room.pdf). APSA Preprint.\n- Broockman, D. & Kalla, J. (2016). [Durably reducing transphobia: A field experiment on door-to-door canvassing](https://www.science.org/doi/10.1126/science.aad9713). *Science* 352(6282).\n\n### Community Notes Bridging\n- [From Birdwatch to Community Notes, from Twitter to X: Four years of community-based content moderation](https://arxiv.org/html/2510.09585v2). arXiv:2510.09585.\n\n### Moral Foundations and Value Divergence\n- Graham, J., Haidt, J., & Nosek, B.A. (2009). [Liberals and conservatives rely on different sets of moral foundations](https://fbaum.unc.edu/teaching/articles/JPSP-2009-Moral-Foundations.pdf). *Journal of Personality and Social Psychology* 96(5).\n- [Liberals and Conservatives Rely on Very Similar Sets of Foundations When Comparing Moral Violations](https://www.cambridge.org/core/journals/american-political-science-review/article/abs/liberals-and-conservatives-rely-on-very-similar-sets-of-foundations-when-comparing-moral-violations/97840A41FF7B09B910F20B97A0A901E6). *American Political Science Review* (2022).\n\n### Negotiation and Interest-Based Reasoning\n- Fisher, R. & Ury, W. (1981). *Getting to Yes: Negotiating Agreement Without Giving In*. Houghton Mifflin. [Summary](https://www.beyondintractability.org/bksum/fisher-getting).\n- [Principled Negotiation: Focus on Interests to Create Value](https://www.pon.harvard.edu/daily/negotiation-skills-daily/principled-negotiation-focus-interests-create-value/). Harvard Program on Negotiation.\n\n### LLM Reasoning and Argument Quality\n- Feng, Y., et al. (2025). [BIRD: A Trustworthy Bayesian Inference Framework for Large Language Models](https://arxiv.org/abs/2404.12494). ICLR 2025 (Oral).\n- [Human/AI Collective Intelligence for Deliberative Democracy: A Human-Centred Design Approach](https://arxiv.org/html/2603.16260). arXiv:2603.16260 (March 2026).\n\n---\n\n*Research compiled March 2026 for the Deliberus project. This document explores a theoretically novel intersection — no decisions are made here about Deliberus's architecture or ontology. The analysis is intended to inform design deliberation, not to predetermine it.*\n"}