Moderating Democratic Discourse

with Large Language Models

Aaditya (Sonny) Bhatia

United States Military Academy, West Point, NY

Advisor: Dr. Gita Sukthankar

University of Central Florida, Orlando, FL

Wednesday, September 18, 2024

Introduction

  • Democracy ➜ policy decisions ➜ discourse
  • Policy decisions are wicked problems 1 that require discourse!
    • Effectiveness determined by solving it; single attempt
    • Measuring impact will shift problem
    • Wisdom of the crowd helps generate and evaluate solutions!
  • Online deliberation works well in small groups
    • Does not scale easily
    • Requires human facilitation

Problem Statement

  • Silos are comfortable ➜ echo chamber effect
  • Lack of diversity ➜ polarization ➜ misinformation
  • Misinformation shapes opinion before correction
  • Facilitation complexity does not scale linearly

How do we use LLMs to improve comment moderation to allow public deliberation at scale?

Background

Polis

  • “Real-time system for gathering, analyzing and understanding” public opinion
  • Developed as an open source platform for public discourse
  • Published several case studies
  • Participants post short messages and vote on others
  • Polis algorithm ensures exposure to diverse opinions
  • \(\vec{comments} \times \vec{votes} =\) opinion matrix
    • fed into statistical models
    • understand where people agree or disagree

Polis live report from Townhall meeting in Bowling Green, Kentucky that took place in 2018.

Human Moderation in Polis

  • Near-real-time feedback (agree, disagree, neutral)
  • Loosely follow Polis moderation guidelines
    • Spam: devoid of relevance to the discussion.
    • Duplicative: restating a previously made point.
    • Complex: articulating multiple ideas or problems.
  • Each organization has its own methodology
  • Process is time consuming and inconsistent

Large Language Models

  • Extremely good at pattern recognition and summarization
  • Generate appealing content
  • Reasoning for small tasks
  • Use for simple decision-making

Research Questions

  • How effective are LLMs for content moderation on Polis?
  • How do various prompting strategies perform against human moderation?
  • Can LLMs effectively augment or entirely replace human moderators?

Methodology

Data

  • Summary Statistics: conversation topic, number of participants, total comments, total votes
  • Comments: author, comment text, moderated, agree votes, disagree votes
  • Votes: voter ID, comment ID, timestamp, vote
  • Participant-Vote Matrix: participant ID, group ID, n-votes, n-agree, n-disagree, comment ID…
  • Stats History: votes, comments, visitors, voters, commenters
Summary of datasets used in the study
Dataset Participants Comments Accepted
american-assembly.bowling-green 2031 896 607
scoop-hivemind.biodiversity 536 314 154
scoop-hivemind.taxes 334 148 91
scoop-hivemind.affordable-housing 381 165 119
scoop-hivemind.freshwater 117 80 51
scoop-hivemind.ubi 234 78 71

Text Generation

Example output produced by guidance. The green highlighted text is generated by the LLM, while rest is programatically inserted into the context. Inference is significantly faster since the model produces fewer tokens. Output format is strictly enforced using stop criteria, regular expressions, and fixed options.

Comment Moderation

  • Retrospective analysis that simulates real-time moderation
  • Goal is to identify spam, irrelevant comments, and those that violate Polis moderation guidelines
  • Framed as a multi-class classification task for spam detection
  • Gold-standard labels available in source dataset

Experimental Configurations

Config Target Classes Examples Deconstruction CoT Technique
1 3 No No N/A
2 3 Yes No N/A
3 3 No No Thought after rejection
4 3 Yes No Thought after rejection
5 7 No No N/A
6 7 No No Thought before decision
7 7 No Yes Thought before decision
8 7 No Yes N/A
9 3 No Yes Thought before decision

Instructions: Three-Class Classification

Discussion Title: Improving Bowling Green / Warren County
Discussion Question: What do you believe should change in Bowling Green/Warren County in order to make it a better place to live, work and spend time?

---
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the given guidelines.

---
Classifications:
- ACCEPT: Comment is coherent, makes a suggestion, or presents a real problem or issue.
- UNSURE: Unclear whether the comment meets the guidelines for ACCEPT.
- REJECT: Comment should definitely be rejected for one of the reasons listed below.

---
Reasons for REJECT:
- SPAM: Comments which are spam and add nothing to the discussion.
- COMPLEX: Comments which state more than one idea. It is difficult to determine the where another person would agree or disagree.

---
Output format:
CLASSIFICATION: One of the following based on given guidelines: ACCEPT, UNSURE, REJECT.
THOUGHT: Express the reasoning for REJECT classification.
Am I certain: Answer with YES or NO. If unsure, state NO.
REASON: One of the following based on given guidelines: SPAM, COMPLEX
EXPLANATION: Provide an explanation for why the comment was classified as REJECT.

Output: Three-Class Classification

Use of Examples

Second-Thought Technique

  • False Positives cause more harm
  • Allow the model to turn a REJECT into UNSURE

Instructions: Seven-Class Classification

Discussion Title: Improving Bowling Green / Warren County
Discussion Question: What do you believe should change in Bowling Green/Warren County in order to make it a better place to live, work and spend time?

---
Classify each comment objectively based on the following guidelines:
- ACCEPT: mentions a real problem related to the discussion.
- ACCEPT: recommends a realistic and actionable solution related to the discussion.
- ACCEPT: makes a sincere suggestion related to the discussion.
- IRRELEVANT: frivolous, irrelevant, unrelated to the discussion.
- IRRELEVANT: does not contribute to the discussion in a meaningful way.
- SPAM: incoherent or lacks seriousness.
- SPAM: provides neither a problem nor a solution.
- UNPROFESSIONAL: the language is informal, colloquial, disrespectful or distasteful.
- SCOPE: cannot be addressed within the scope of original question.
- COMPLEX: introduces multiple ideas, even if they are related to the discussion.
- COMPLEX: discusses distinct problems, making it difficult to determine where another person would agree or disagree.
- UNSURE: may be accepted if it appears somewhat related to the discussion.

---
Output format:
CLASSIFICATION: One of the following based on given guidelines: ACCEPT, UNSURE, SPAM, IRRELEVANT, UNPROFESSIONAL, SCOPE, COMPLEX.
EXPLANATION: Provide an explanation for the classification.

Instructions: Comment Deconstruction

Output format:

PROBLEM: The specific problem mentioned in the comment. If only an action is suggested and no problem is explicitly mentioned, state None.
ACTION: What suggestion or change is proposed. If only a problem is mentioned and no action is suggested, state None.
HOW MANY IDEAS: Number of distinct ideas introduced in the comment.
THOUGHT: Deliberate about how the comment should be classified.
CLASSIFICATION: ACCEPT, UNSURE, SPAM, COMPLEX.
REASON: If comment was not classified as ACCEPT, explain.

Output: Comment Deconstruction and Thought Statements

Results

Comment Moderation

  • Accuracy generally the same
  • Unsure rate increases with complexity of task
  • Deconstruction reduces false positive rate
  • CoT not as effective as deconstruction
  • Examples must be specific to dataset

Configurations

  • 1: Baseline
  • 2: Examples
  • 3: Thought
  • 4: Thought + Examples
  • 5: 7-class Baseline
  • 6: Thought
  • 7: Thought + Deconstruction
  • 8: Deconstruction
  • 9: Deconstruction, 3-class

Conclusion

LLMs in structuring online debates

  • Potential of LLMs for simple tasks
    • Risk of hallucinations and incorrect output
    • Chaining simple tasks for complex reasoning
  • Augmenting vs replacing human moderation processes
  • LLMs’ limitations in processing complex instructions and sentences
    • Complex instructions
    • Relationship modeling based on double and triple negatives
  • Reliability and bias
    • Critical need for ethical and inclusive technology deployment

Future Research Directions

  • Semantic extraction and reasoning during discourse
    • Exploring connections across topics
  • Individual feedback before posting a comment
  • Generalizing techniques to platforms like Kialo, Hacker News

Questions?