Using LLMs to Structure and Visualize Policy Discourse

Aaditya (Sonny) Bhatia

United States Military Academy, West Point, NY

Advisor: Dr. Gita Sukthankar

University of Central Florida, Orlando, FL

Sunday, August 4, 2024

Context

Policy decisions pose a complex, wicked problem ^{1 2}

• Effectiveness determined by solving it; single attempt
• Measuring impact will shift problem
• Public discourse helps shape solutions; crucial for policy-making

Determining public opinion

• Surveys and polls -> Social Media -> Discussion Platforms
• People willing to express freely
• Digital platforms provide a wealth of data
• Unstructured, vast, and complex

First, let’s understand the problem space.

What are wicked problems and why do they matter?

• A problem that is difficult or impossible to perfectly solve because of incomplete, contradictory, and changing requirements that are often difficult to recognize.
• An example would be building a highway to connect two ends of a city.
• You can go through the city or around it
• and each option has its own unique implications.
• There is no way to find out which solution is better without trying it out
• and the solution will change the problem itself.

Most policy decisions are wicked problems

• public discourse helps shape solutions to wicked problems
• Historically, public opinion was determined through surveys and polls.
• Social media provided researchers with a new avenue but limited by each platform’s constraints.
• Specific discussion platforms provide a wealth of data, but it is often unstructured, vast, and complex.

1. characterized by no definitive formulation, no stopping rule, no specific optimization criteria, and no right/wrong solutions

2. (E. J. Conklin 2006)

Problem Statement

How can LLMs enable us to

• ingest massive streams of unstructured information
• incorporate diverse perspectives, and
• distill them into actionable insights, that
• align with public opinion?

What are the inherent risks associated with the deployment of LLMs?

How can LLMs enable us to

• ingest massive streams of unstructured information
• incorporate diverse perspectives, and
• distill them into actionable insights, that
• align with public opinion?

Related Works

gIBIS

• Networked decision support system (J. Conklin and Begeman 1988)
• Structured conversation using
  • Issues
  • Positions
  • Arguments
• Helped identify underlying assumptions
• Promoted divergent and convergent thinking
• Limited by structure, learning curve, scalability
• Provided the basis for several other tools

gIBIS application supports several types of nodes and links for structured thinking. (J. Conklin and Begeman 1988)

• Issue Based Information System, or IBIS
• is a discussion facilitation technique that relies on a shared display
• great for clarifying wicked problems
• it’s the most commonly used approach for argumentation
• and provides the basis for several argument mapping tools and discussion platforms
• gIBIS = graphical IBIS, a UI developed as a networked decision support system
• Allowed multiple users to hold structured discussions and establish a common operation picture
• Three types on nodes (point at picture)
  • Issues represent questions or problems
  • Positions - potential answers to those issues
  • Arguments - that support or refute those answers
• evolved into powerful method for research thinking and design deliberation
• Helps identify underlying assumptions
• Promotes divergent thinking by encouraging exploration
• Enables convergent thinking by allowing the group to reach consensus

Polis

• “Real-time system for gathering, analyzing and understanding” public opinion (Small 2021)
• 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.

• Polis is a real-time system for analyzing public opinion
• It’s been developed as an open-source platform and used by several governments and organizations across the world
• When an organization hosts a discussion
  • they post a discussion topic and question for the public
  • participants respond to the discussion using short comments - 140 characters, like a tweet
  • these comments are expected to either introduce an issue or suggest a solution
  • cannot reply
  • vote on other comments to indicate how they feel
• Polis’s underlying mechanism ensures that people are exposed to diverse opinions
• On the right, you can see a screenshot from a live report from a town hall meeting in Bowling Green, Kentucky
• (describe it maybe)
• This voting activity results in an opinion matrix
  • this is fed into statistical models
  • specifically understand where people tend to agree or disagree
• We use Polis because it provides a rich dataset of public opinion that we can structure and analyze

D-Agree Crowd-Scale Discussion

• Automated agent to facilitate online discussion (Ito, Hadfi, and Suzuki 2022)
• IBIS-based discussion representation
• Extracts and analyzes discussion structures from online discussions
• Posts facilitation messages to incentivize participants and grow IBIS tree
• Best results when agent augmented human facilitators (Hadfi and Ito 2022)
• Results
  • Use of the agent produced more ideas for any given issue
  • Agent had 1.4 times more replies and 6 times shorter response intervals
  • Increased user satisfaction and sense of accomplishment

• In this project, the researchers developed an automated agent that facilitates chat-style discussions
• All the other approaches so far relied on humans to structure the discussion
• In this one, people are not expected to follow a structure
• This technique uses an automated agent that constructs an IBIS tree of the discussion
• Then, based on which parts of the tree need more attention, it injects facilitation messages into the conversation
• this leads to more divergent thinking and a better quality of ideas
• In the study, the authors determined that the participants felt more accomplished when the agent was present

Methodology

Here is our overall pipeline.

• We will look at each of these steps in detail in the following part.

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

Polis Data

This is the data that obtained from Polis.

• It has five tables.
• keep track of each comment posted and vote cast during an event
• Participants-Vote: a sparse matrix of participants and their votes. It also includes clustering assignments (group-id) based on voting patterns, providing a way to analyze different opinion groups

Datasets

• We use these six datasets in our study, each with a different topic of discussion.
• We selected these based on the consistency of the comment moderation to ensure that the accepted comments were high quality and relevant to the discussion.

Source: github.com/compdemocracy/openData

Embeddings

• Calculated at comment level using Sentence Transformers library
• Models considered
  • intfloat/e5-mistral-7b-instruct
  • WhereIsAI/UAE-Large-V1
  • OpenAI/text-embedding-ada-002
  • OpenAI/text-embedding-3-large

• Language Model Selection Criteria
  • Open weights
  • Clustering performance on HuggingFace MTEB
  • Memory footprint

• Once we get the data, we calculate its embeddings.
• Embeddings are numerical vectors that represent the semantical meaning of a word or sentence.
• Transformer embeddings are calculated based on the context in which it appears.
• These are used by language models to infer the next word in a sentence.
• We calculated embeddings at the comment level, which was done by averaging the embeddings of all the words in a comment.
• A comment could be 2-3 sentences, limited to 140 characters, and should contain only one idea.
• This ensures that the embeddings capture the meaning of the entire comment.

Models

• We considered several models, including a Mistral-based model from Microsoft Research, UAE-Large, and OpenAI’s models.
• We selected these models based on their performance on the HuggingFace MTEB leaderboard, memory footprint, and availability of open weights.
• Ultimately, we selected UAE-Large for our study due to its superior performance with an extremely low memory footprint.

Reference: osanseviero.github.io/hackerllama/blog/posts/sentence_embeddings

Text Generation

• Models Considered
  • Llama 2 13B Chat
  • Mistral 7B Instruct v0.2
  • Mixtral 8x7B Instruct 4-bit Quantized
• Guidance
  • Python-based framework developed by Microsoft Research
  • Constrain generation using regular expressions, context-free grammars
  • Interleave control and generation seamlessly

lm += f"""\
The following is a character profile for an RPG game in JSON format.
```json
{{
    "id": "{id}",
    "description": "{description}",
    "name": "{gen('name', stop='"')}",
    "age": {gen('age', regex='[0-9]+', stop=',')},
    "armor": "{select(options=['leather', 'chainmail', 'plate'], name='armor')}",
    "weapon": "{select(options=valid_weapons, name='weapon')}",
    "class": "{gen('class', stop='"')}",
    "mantra": "{gen('mantra', stop='"')}",
    "strength": {gen('strength', regex='[0-9]+', stop=',')},
    "items": ["{gen('item', list_append=True, stop='"')}", "{gen('item', list_append=True, stop='"')}", "{gen('item', list_append=True, stop='"')}"]
}}```"""

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.

Text Generation Models

• We considered the three models listed here
• and ended up using a quantized version of Mixtral
• Mistral performed quite well but Mixtral showed better reasoning capabilities

Guidance

• To control the language model’s output, we used a framework called Guidance
• Traditionally, you would prompt the language model, receive and parse the response, and hope that it is in the correct format
• Guidance lets you specify exactly what you want using regular expressions, context-free grammars, or just a simple list of options
• You can also pause generation, insert text into the context, and then ask the model to generate more
• Also, since the model has to come up with fewer tokens, the process is much much faster
• In the example below, we ask the model to generate a structured JSON object
• The output is on the right.
• The highlighted text in the output is generated by the language model, while the rest is outputted from the string.
• (next) For the name, we tell the model to stop generating when it reaches a double quote
• (next) For the age, we only allow numbers and stop when it tries to produce a comma
• (next) For armor, we provide a list of options and ask the model to select one
• This way, the output format is perfectly enforced in a very efficient way

Topic Modeling

• Here is the topic modeling pipeline
• I used the BERTopic framework
• first, I generated embeddings for each comment
• then I used UMAP to reduce the dimensionality
• and a DBSCAN algorithm to identify the clusters or topics
• then, we extracted the keywords to identify the most important terms in each cluster
• and finally, we generated a topic representation based on these terms

Pipeline implemented using BERTopic framework (Grootendorst 2022)

Topic Outlier Assignment

Each bar represents a topic cluster, with y-axis representing statement count; Topic -1 is reserved for outliers that do not initially belong to a cluster.

Upon reassigning outliers, each topic has a larger number of statements; we reassign all topics ensuring that no statement is discarded as noise.

• These charts show the distribution of statements across different topics
• The first chart shows the initial distribution, with each bar representing a topic cluster and the height of those bars indicating the number of comments in each topic.
• The second chart shows the distribution after reassigning outliers, which appears reasonably balanced and indicates an even distribution of comments across topics.

Statement Distribution

This figure shows the distribution of statements projected onto a 2D space using UMAP. Each point represents a statement, and the color represents the topic cluster to which it belongs. The gray points are outliers that do not belong to any cluster.

Statement Distribution after outlier reassignment

• This is the same projection after reassigning the outliers.
• The distribution is more balanced, and every statement is now assigned to a topic.
• Note that assigning these additional statements to each topic inherently changes it, so
  • the topic centers have moved
  • the labels are not the same as previous figure
  • and the 2D projection looks entirely different

Insight Generation

• Actionable insights that urge for specific actions to address issues
• Problems and solutions proposed by participants
• LLM synthesizes insights from comments within each topic
• Advocate for specific insights urging actions to address issues
• Filter these insights to derive actionable items

• An argument is a statement that advocates for a specific position or action.
• the goal was to identify the most important issues and potential solutions
• our key idea is to generate a lot of arguments, score them based on acceptability, and then select the best ones
• I will walk you through the diagram at the bottom
• first, we filtered the comments based on their agreeability
  • agreeability is the proportion of agree votes to total votes
  • this threshold lets us control what proportion of statements are used to form high-quality arguments
• Filtering carries the risk of excluding comments and opinions.
  • We do not want to exclude any valid opinions
  • but also eliminate the ones that are just very unpopular
• We feed these comments and topic description into the language model
  • and ask it to identify areas of improvement
• for each area of improvement
  • we generate a list of problems discussed by the participants
  • and potential solutions proposed by the participants
• we are using this to interpret these problems and solutions
  • we do not want the model to come up with new solutions
  • to minimize hallucinations, we kept our instructions very simple and clear
• once we get a list of actionable insights, we select the top ones from each topic based on their acceptability
  • and use them to generate the argument map

Insight Scoring

• Goal: Quantify acceptance of each generated insight
• Task: Identify comments that support each insight
• Count the individuals that voted positively on supporting comments
• Calculate an “acceptance” factor to indicate the degree of consensus

• For scoring, the goal is to quantify the level of support for each argument
• Since we do not know exactly which comments resulted in a given argument, we use the language model to figure this out
• We consider each argument and comment within a topic, and classify the relationship between them
• Then, we count the individuals that generally agreed with the comments that support a given argument
• We use this to calculate an “acceptance” factor, which indicates the degree of consensus
• In the diagram…

Potential Biases in Insight Generation

• Some comments, especially those posted earlier, may receive more votes than others
  • Use the ratio of agreement votes to total votes
• Certain topics are more popular and have more comments than others
  • Generate a balanced number of insights for each topic
• Certain controversial topics are heavily downvoted
  • Comments: Filter by quantiles instead of fixed thresholds within each topic
  • Insights: Select fixed number of “best insights” from each topic
• Some people vote more than others
  • Count the individuals that support an insight over hard vote count

• When generating insights, we consider these potential biases
• Some comments have more votes than others
  • We normalize this by calculating the agreeability factor
  • which is the ratio of agree votes to total votes
• Some topics are more popular or controversial, and have more comments
  • We normalize this by generating a balanced number of insights for each topic
• Some topics are heavily downvoted for various reasons, so the votes across topics are not comparable
  • For example, a topic in our dataset about legalizing weed was heavily downvoted overall, but still have valid opinions
  • Had we compared them against other topics, those comments would never have surfaced
  • To control for that, we use of quantiles instead of fixed thresholds within each topic.
  • In our case, we use the 0.1 quantile to only trim the worst 10% comments within the topic.
• Some people vote more than others, which can skew our insight support calculation
  • If someone agreed to a comment, they are likely to agree with other supporting positions.
  • We count the individuals that voted on at least one supporting comment instead of the total number of votes

Argument Mapping

• Used Argdown syntax to generate argument maps
• Developed a grammar generator to convert data into Argdown format
• Generated argument maps for each topic to visualize the structure of the debate

===
sourceHighlighter:
    removeFrontMatter: true
webComponent:
    withoutMaximize: true
    height: 500px
===

# Argdown Syntax Example

[Statement]: Argdown is a simple syntax for defining argumentative structures, inspired by Markdown.
  + Writing a list of **pros & cons** in Argdown is as simple as writing a twitter message.
  + But you can also **logically reconstruct** more complex relations.
  + You can export Argdown as a graph and create **argument maps** of whole debates.
  - Not a tool for creating visualizations, but for **structuring arguments**.

<Argument>: Argdown is an excellent tool and should be used by the city of Bowling Green, KY.

[Statement]
  +> <Argument>

• To generate the argument maps, we used the Argdown syntax
• It is a simple syntax inspired by Markdown
• The example below is a simple structure that shows a statement, a couple of supporting positions, and an argument.
• If you click on the map, you will see the argument laid out in a tree-like structure
• The boxes with white backgrounds are statements or positions
• The boxes with solid backgrounds are arguments that are supported by the statements
• The green arrows indicate support while the red arrows indicate opposition

Argument Generation and Scoring

Opioid Epidemic and Healthcare

• In this map, the discussion is about opioid epidemic and healthcare
• The top node is the topic
• The middle row shows the arguments synthesized by the language model
• The bottom row shows the comments that support each argument
• We have the voting data for each comment
  • which we use to calculate support for each argument
• We produced a total of 102 argument for the bowling green dataset
  • and used our filtering approach to select the top arguments and comments from each topic
• we still use all the comments to calculate the support for each argument
  • but only show the top ones in the map

Argument Generation and Scoring

Community Enrichment

Conclusion

• Chaining simple tasks for complex reasoning
• Discovering topics in a large dataset and new generating valuable insights
• Risk of hallucinations and incorrect output
• 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
• Generalizing techniques to platforms like Kialo, Hacker News

• Overall, I found that chaining these simple tasks can help us create complex reasoning pipelines
• We were able to discover topics in a large dataset and generate valuable insights
• There is a risk of hallucinations and incorrect output
• Hallucinations were mitigated by using different chain-of-thought techniques
• and most importantly, by keeping our instructions very simple
• use of guidance to control the generation of text really helped as well
• incorrect information is a risk with any machine learning model
• the best we can do is to pair these models with humans to catch these errors early
• Overall, this helps enable public discourse
• Which can be a powerful tool for enhancing democratic processes
• depending on how it is used
• But we also need to be careful about how we deploy these technologies

In this paper-

• We developed a novel argument generation framework that synthesizes arguments from user-generated comments
• and we introduced a scoring mechanism to quantify the acceptance of each synthesized argument
• A key innovation is the way we produce argument maps.
• These distill complex relationships between topics and insights and make it easier for policy-makers to understand what the public is saying
• Recognized limitations in current LLMs’ ability to fully comprehend the nuanced dynamics of online debates
• Acknowledged challenges in maintaining an unbiased moderation process, highlighting the delicate balance between automated and human moderation

Future research

Questions?

Thank you!

Conklin, E. Jeffrey. 2006. Dialogue Mapping: Building Shared Understanding of Wicked Problems. Chichester, England ; Hoboken, NJ: Wiley.

Conklin, Jeff, and Michael L. Begeman. 1988. “gIBIS: A Hypertext Tool for Exploratory Policy Discussion.” ACM Transactions on Information Systems 6 (4): 303–31. https://doi.org/10.1145/58566.59297.

Grootendorst, Maarten. 2022. “BERTopic: Neural Topic Modeling with a Class-Based TF-IDF Procedure.” arXiv. http://arxiv.org/abs/2203.05794.

Hadfi, Rafik, and Takayuki Ito. 2022. “Augmented Democratic Deliberation: Can Conversational Agents Boost Deliberation in Social Media?” In Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems, 1794–98. AAMAS ’22. Richland, SC: International Foundation for Autonomous Agents; Multiagent Systems.

Ito, Takayuki, Rafik Hadfi, and Shota Suzuki. 2022. “An Agent That Facilitates Crowd Discussion.” Group Decision and Negotiation 31 (3): 621–47. https://doi.org/10.1007/s10726-021-09765-8.

Small, Christopher. 2021. “Polis: Scaling Deliberation by Mapping High Dimensional Opinion Spaces.” RECERCA. Revista de Pensament i Anàlisi, July. https://doi.org/10.6035/recerca.5516.