PaperlayerView exact revision on arXiv

Source-linked AI summary

Recommending the Most Encompassing Opposing and Endorsing Arguments in Debates

Marius C. Silaghi, Roussi Roussev

arXiv:1411.5416v1cs.AI

TL;DR

The paper asks how to recommend a suitable justification to a new voter and how to assemble a compact set covering known arguments. It models available vote–justification associations and relations with weighted bipartite graphs, proposing a framework for recommending relevant opinions while treating natural-language arguments as opaque.

  • Problem

    The paper addresses how to recommend a justification of a given type to a new voter and how to select a compact list that subsumes known arguments for or against an issue.

    Source: §2 · p. 4

  • Method

    It uses vote–justification associations and threading structure, modeling relations between supporting and opposing justifications in a weighted bipartite graph.

    Source: §1 · p. 2, §2 · p. 3, §3 · p. 6

  • Results

    The paper proposes a logical framework for reasoning about debate arguments and recommending relevant opinions to readers navigating the debate graph.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The framework uses artificial intelligence to help people find relevant arguments, while deliberation and reevaluation of views remain with people.

    Source: §4 · p. 7, §5 · p. 9

    Natural-language arguments are not expected to be robustly parsed, and further research is needed to efficiently exploit the Bayesian Network directly for inference.

    Source: §2 · p. 3, §4 · p. 9

Abstract

from arXiv · show

Arguments are essential objects in DirectDemocracyP2P, where they can occur both in association with signatures for petitions, or in association with other debated decisions, such as bug sorting by importance. The arguments of a signer on a given issue are grouped into one single justification, are classified by the type of signature (e.g., supporting or opposing), and can be subject to various types of threading. Given the available inputs, the two addressed problems are: (i) how to recommend the best justification, of a given type, to a new voter, (ii) how to recommend a compact list of justifications subsuming the majority of known arguments for (or against) an issue. We investigate solutions based on weighted bipartite graphs.

1 Introduction

DirectDemocracyP2P groups a signer's arguments into a justification associated with a petition signature, and allows support, opposition, and abstention. The paper motivates recommending relevant justifications while treating their natural-language arguments as opaque and using relations between justifications to focus on the meaning of the associated vote.

1 Introduction

  • The system extends petition signatures with justifications from supporters and opposers, creating a need to help voters find relevant justifications among potentially large numbers of submissions.Each author is authenticated and can reference one justification for a petition.Source: §1 · p. 1
  • Support, opposition, and abstention can provide controlled, authenticated feedback to petition supporters and observers.The paper notes that these manifestations may indicate opposition strength, though less statistically reliably than polls with controlled subject distribution.Source: §1 · pp. 1–2
  • A justification is a natural-language explanation containing arguments, intended to support reciprocal understanding and encourage activism.The paper also presents justifications as a possible way to bring democracy to media and mitigate critiques of manipulation and censorship.Source: §1 · p. 2
  • Justifications are classified by their accompanying signature; the proposed solutions initially distinguish supporting from opposing justifications.Abstention is also described as a signature type in the broader classification.Source: §1 · p. 2
  • The framework treats arguments as opaque and uses explicit relations such as threading, contradicting, and including to study the semantics of the associated vote.This basic case is presented as a focus on conclusion semantics, with more complex argument forms left as a possible extension.Source: §1 · p. 2

2 Background and Problem

The paper frames recommendation as helping constituents find existing justifications or components that cover arguments relevant to their own votes. Because natural-language arguments are opaque to computation, it proposes using vote associations and user-specified relations between justifications.

2 Background and Problem

  • Prior argumentation work often studies rule-governed committees seeking logically justified decisions, whereas this forum setting assumes participants have different values.The paper says that generating a unique justification proving lack of corruption is not the main issue in its setting.Source: §2 · p. 3
  • The computational system does not expect to robustly parse natural-language arguments, so it plans to exploit vote–justification associations and threading structure.The assumed setting has no supreme individual authority; at least in grassroots organizations, authority rests with the constituency.Source: §2 · p. 3
  • The first problem is finding an existing justification that subsumes the arguments a new voter considers relevant for a pro or against vote.If no such justification exists from that voter's perspective, the voter can create a new one.Source: §2 · p. 4
  • Candidate discovery relies on a hierarchy built from relations between justifications, not from their opaque content; users specify what relations such as subsumes and refutes mean.The paper treats these relations as arcs in a bipartite graph and proposes finding the most complete subsuming candidates.Source: §2 · p. 4
  • The second problem is selecting a subset of justifications that subsumes arguments for a vote, helping both new-argument authors and undecided voters form an opinion.Candidate criteria include support, recency, and completeness with respect to existing relations, or combinations of those criteria.Source: §2 · pp. 4–5

3 Concepts

The concepts formalize justifications, signatories, and relations between opposing and supporting justifications. The resulting problems seek bounded sets of justifications that answer voters, with variants incorporating refutation and subsumption relations.

3 Concepts

  • A justification answers a voter when it is associated with that voter's signature or was created as claiming to refute the justification selected by that voter.The framework introduces basic relation-based models for the algorithms that follow.Source: §3 · p. 5
  • The Subsuming Justification Problem (SJP) uses opposing and supporting justification sets, their signatures, refutation relations, and a bound K.Its objective is to find up to K justifications of each type that answer the maximum number of signatories.Source: §3 · p. 5
  • The Weighted Subsuming Justification Problem (WSJP) weights a claimed refutation relation by the number of signatories of the left-hand justification who explicitly state that it refutes the right-hand one.It seeks bounded supporting and opposing sets answering the maximum number of signatories.Source: §3 · p. 6
  • The Components Subsuming Justification Problem (CSJP) adds same-type claimed subsumption relations to the justification and refutation framework.Its bounded sets aim to answer the maximum number of signatories; the subsumption claim is explicitly supplied and may not be automatically verifiable.Source: §3 · pp. 6–7

4 Algorithms

The algorithms search refutation relations to find justifications that counter the largest number of opposite-type justifications, using depth-weighted votes. A Bayesian network is also proposed to model the problem and generate test cases.

4 Algorithms

  • A limited-depth search applies refutation transitivity to find justifications that counter the largest number of opposite-type justifications.Figure 1 presents the counter-argument algorithm; the search starts from a given justification and extends through claimed refutations.Source: §4 · p. 7, Figure 1 · p. 8
  • Votes combine justification signatures and refutation links, with γ < 1 discounting contributions from deeper levels.The two directions of refutation are weighted by α and β in votes(k).Source: §4 · p. 7
  • Higher search levels offer more insight but propagate assumption errors more strongly; the authors anticipate useful levels in {2, 3, 4}.The best parameters are to be identified by simulation.Source: §4 · p. 8
  • Bipartite graphs can represent supporting and opposing justifications, while transitive closure over subsumption relations can produce a compact set covering relevant justifications.Framework combinations can allow individual votes on static relations, and timestamps add a recency relation that extends the concept of answers.Source: §4 · p. 8

Generative models

  • The Bayesian-network proposal models each justification's arguments as a subset of possible arguments and its refutation as a less-recent justification, with distributions based on argument counts.Argument-set nodes are described as independent and uniformly distributed over the power set; refutation nodes depend on more-recent justifications.Source: §4 · pp. 8–9
  • The network's high variable cardinality leaves efficient direct inference for justifications containing the largest argument subsets as further research, while sampling can supply artificial ground truth for evaluation.The proposed evaluation compares recognition by the preceding mini-max algorithm against sampled test cases, insofar as the network models the world correctly.Source: §4 · p. 9

4 Algorithms

  • Deliberation remains with people, while AI helps them find relevant arguments rather than reasoning on their behalf.Participants reevaluate their fears after seeing arguments, given their different foundational values.Source: §4 · p. 9

5 Conclusion

The paper proposes an abstract argumentation framework for reasoning about debate arguments and recommending relevant opinions. It is designed for decentralized submissions by involved peers and for recommendations in DirectDemocracyP2P.net.

5 Conclusion

  • The paper proposes a logical framework, described as a type of abstract argumentation framework, for reasoning about debate arguments.It is intended to recommend relevant opinions to readers navigating the debate graph.Source: §5 · p. 9
  • The framework is designed to recommend relevant opinions to readers navigating the debate graph in DirectDemocracyP2P.net.Source: §5 · p. 9
  • The framework targets debates where peers submit atomic arguments in a decentralized fashion within attack/support relations.The paper identifies DirectDemocracyP2P.net as the system for which the framework is designed to provide recommendations.Source: §5 · p. 9

Related papers

Loading 1411.5416v1…