Comment Moderation Experiment

Polis Comment Moderation using Language Models
Author

Sonny Bhatia

Published

February 29, 2024

We use Mixtral 8x7B Instruct 0.1 4-bit quantized language model to generate reasoning and classify our comments. Guidance framework allows us to control text generation using the Transformers library. All text highlighted in green is generated by the model. Note that Guidance enables us to intertwine prompt and generations, allowing precise control over where and how text is generated.

Import Packages and Setup Environment

Code 
import sys
import os
from pprint import pprint
from tqdm.notebook import tqdm

import polars as pl
from argmap.dataModel import DataModel, Summary, Comments

from dotenv import load_dotenv
load_dotenv()

# this allows categorical data from various sources to be combined and handled gracefully; performance cost is acceptable
pl.enable_string_cache()
Code 
import guidance
from guidance import models, gen, select, instruction, user, assistant, one_or_more, zero_or_more

Define Heatmap Plotting Function

Code 
from sklearn.metrics import confusion_matrix
import seaborn as sns
import matplotlib.pyplot as plt
import numpy as np


def plot_confusion_matrix(true_labels, pred_labels):
    labels_captions = ["Accepted", "Unsure", "Rejected"]
    cm = confusion_matrix(true_labels, pred_labels, labels=[1, 0, -1])
    sns.heatmap(cm, annot=True, fmt="d", cmap="crest",
                xticklabels=labels_captions, yticklabels=labels_captions)
    plt.title('Confusion Matrix Heatmap')
    plt.ylabel('Actual labels')
    plt.xlabel('Predicted labels')
    plt.tight_layout()
    plt.show()
    print(f"Dot Product: {np.dot(true_labels, pred_labels)}")
    print(f"Unsure Count: {pred_labels.size - np.count_nonzero(pred_labels)}")

Define Data Model for Results

Code 
results_schema = {
    'commentId': pl.UInt16,
    'languageModelModerated': pl.Int8,
    'classification': pl.Categorical,
    'thoughts': pl.List(pl.String),
}

Load Dataset

Code 
from IPython.display import display_markdown

# for now, work with one source
DATASET = "american-assembly.bowling-green"
# "scoop-hivemind.biodiversity"
# "scoop-hivemind.freshwater"
# "scoop-hivemind.taxes"
# "scoop-hivemind.ubi"
# "scoop-hivemind.affordable-housing"
# "london.youth.policing"
# "canadian-electoral-reform"
# "brexit-consensus"
# "ssis.land-bank-farmland.2rumnecbeh.2021-08-01"

comments = Comments(DATASET).load_from_csv()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()

summary = Summary(DATASET)

display_markdown(f"""
### Dataset: {DATASET}
### {summary.topic}
### {summary.get('conversation-description')}
### Full Report: [{summary.url}]({summary.url})
""", raw=True)

print("Comments Dataset Overview:")
comments.glimpse()

Dataset: american-assembly.bowling-green

Improving Bowling Green / Warren County

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?

Full Report: https://pol.is/9wtchdmmun

Comments Dataset Overview:
Rows: 896
Columns: 8
$ timestamp    <i64> 1518501496471, 1518499107516, 1518498500733, 1518498456436, 1518498361793, 1518498193970, 1518497788717, 1518497602288, 1518497215716, 1518496966699
$ commentId    <u16> 895, 894, 893, 892, 891, 890, 889, 888, 887, 886
$ authorId     <u16> 1756, 1480, 1751, 1751, 1751, 1751, 1749, 1748, 1747, 1747
$ agrees       <u16> 48, 27, 33, 1, 59, 63, 1, 46, 1, 1
$ disagrees    <u16> 8, 16, 18, 0, 9, 7, 0, 8, 0, 0
$ moderated     <i8> 1, 1, 1, -1, 1, 1, -1, 1, -1, -1
$ commentText  <str> 'Smallhouse Rd (at Campbell Ln) heading into town needs to be widened to have room for three lanes (left turn, straight, right turn).', "Doctors should not overreact to the opioid crisis. People in pain shouldn't suffer because others lack self control and abuse the system.", 'More free meeting space for nonprofits.', 'Skating rink on Gordon ave.', 'After school activity centers for youth with later hours of operation.', 'More affordable home ownership options.', 'We need to build up and support the historic downtown.   There needs to be less apartments and more activities that support the community.', 'Better visibility on roads. Reflectors and better signage', 'Free cookies for everyone! Yay!', 'Legalize marijuana so people will stop talking about legalizing marijuana.'
$ agreeability <f32> 0.7142857313156128, 0.25581395626068115, 0.29411765933036804, 1.0, 0.7352941036224365, 0.800000011920929, 1.0, 0.7037037014961243, 1.0, 1.0

Verify GPU Availability

Code 
import torch

if not torch.cuda.is_available():
    raise Exception("No CUDA device found")


def cuda_print_memory():
    free_memory = sum([torch.cuda.mem_get_info(i)[0]
                      for i in range(torch.cuda.device_count())])
    total_memory = sum([torch.cuda.mem_get_info(i)[1]
                       for i in range(torch.cuda.device_count())])
    allocated_memory = sum([torch.cuda.memory_allocated(i)
                           for i in range(torch.cuda.device_count())])

    print(f"CUDA Memory: {round(free_memory/1024**3,1)} GB free, {round(allocated_memory/1024**3,1)} GB allocated, {round(total_memory/1024**3,1)} GB total")


def cuda_ensure_memory(required_memory_gb):
    required_memory = required_memory_gb * 1024**3
    free_memory = sum([torch.cuda.mem_get_info(i)[0]
                      for i in range(torch.cuda.device_count())])
    if free_memory < required_memory:
        raise Exception(
            f"Insufficient CUDA memory: {round(free_memory/1024**3,1)} GB free, {required_memory_gb} GB required")


print(f"""\
Device: {torch.cuda.get_device_name(0)}
Python: {sys.version}
PyTorch: {torch.__version__}
CUDA: {torch.version.cuda}
CUDNN: {torch.backends.cudnn.version()}
""")

cuda_print_memory()
Device: NVIDIA GeForce GTX 1650
Python: 3.11.8 | packaged by conda-forge | (main, Feb 16 2024, 20:53:32) [GCC 12.3.0]
PyTorch: 2.2.1
CUDA: 12.1
CUDNN: 8902

CUDA Memory: 3.8 GB free, 0.0 GB allocated, 3.8 GB total

Grammar Specification for Text Generation

Code 
@guidance(stateless=True)
def generate_line(lm, name: str, temperature=0, max_tokens=50, list_append=False):
    return lm + gen(name=name, max_tokens=max_tokens, temperature=temperature, list_append=list_append, stop=['\n'])


@guidance(stateless=True)
def generate_phrase(lm, name: str, temperature=0, max_tokens=50, list_append=False):
    return lm + gen(name=name, max_tokens=max_tokens, temperature=temperature, list_append=list_append, stop=['\n', '.'])


@guidance(stateless=True)
def generate_number(lm, name: str, min: int, max: int, list_append=False):
    return lm + select(list(range(min, max+1)), name=name, list_append=list_append)

Initialize Language Model

Code 
torch.cuda.ipc_collect()
torch.cuda.empty_cache()

CUDA_MINIMUM_MEMORY_GB = os.getenv("CUDA_MINIMUM_MEMORY_GB")
MODEL_ID = os.getenv("MODEL_ID")

if MODEL_ID is None:
    print("MODEL_ID environment variable is required.")
    sys.exit(3)

if CUDA_MINIMUM_MEMORY_GB is not None:
    cuda_ensure_memory(int(CUDA_MINIMUM_MEMORY_GB))

# MODEL_ID = 'mistralai/Mistral-7B-Instruct-v0.2' # ~31GB RAM
# MODEL_ID = 'meta-llama/Llama-2-13b-chat-hf' # ~50GB RAM
# MODEL_ID = 'TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ'

print(f"Initializing language model: {MODEL_ID}...")
languageModel = models.TransformersChat(MODEL_ID, device_map="auto")

print("Language model initialized. Memory usage:")
cuda_print_memory()
Initializing language model: TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ...
Language model initialized. Memory usage:
CUDA Memory: 55.6 GB free, 22.8 GB allocated, 79.1 GB total

Moderation Techniques

Polis project recommend the following guidelines for moderation:

  • Spam: comments that have no relevance to the discussion
  • Duplicative: comments that repeat a point that has previously made
  • Comments that state more than one idea

In the following runs, we consider the effect of following variables on the outcome:

  • Classify options presented to the language model
  • Use of examples
  • Chain-of-thought reasoning: adding a “thought” statement before rendering a decision might improve reasoning, as shown in ReAct approach
  • Language models
    • Mistral
    • Mixtral
    • Llama 7b
    • Llama 13b

Later, we combine different techniques to improve moderation.

  1. Duplicate comments can be detected using embeddings instead of language model.
  2. Language model is used to detect spam based on relevance.
  3. Language model is used to detect comments that state more than one idea. It can rewrite them as separate comments that state one idea each.
Code 
@guidance
def guidance_moderation(lm, instructions, classify_options, comments_df, results, examples_accepted=[], examples_rejected=[], reject_reasons=[], two_step_strategy=False, thought=False, decompose_comments=False, explain=False, temperature=0, context_reset=False, progress_bar=None):

    if progress_bar is not None:
        lm.echo = False

    # output the task instructions
    with instruction():
        lm += instructions

    # output some examples to help the assistant understand the task
    i = 1
    for example in examples_accepted:
        with user():
            lm += f"COMMENT EXAMPLE {i}: {example}"
            i += 1

        with assistant():
            lm += classify_options[0]

    for example, reason, explanation in examples_rejected:
        with user():
            lm += f"COMMENT EXAMPLE {i}:\n{example}"
            i += 1

        with assistant():
            lm += f"""\
            {classify_options[2]}
            COMMENT EXAMPLE {i} CLASSIFICATION: {reason}
            EXPLANATION: {explanation}
            """

    # iterate through the comments and ask the assistant to moderate each one
    for commentId, commentText in comments_df.select('commentId', 'commentText').iter_rows():
        if two_step_strategy:
            if context_reset:
                lm + guidance_moderation_comment_two_step(
                    commentId, commentText, classify_options, results, thought, explain, temperature, reject_reasons)
            else:
                lm = lm + guidance_moderation_comment_two_step(
                    commentId, commentText, classify_options, results, thought, explain, temperature, reject_reasons)
        else:
            if context_reset:
                lm + guidance_moderation_comment_one_step(
                    commentId, commentText, classify_options, results, thought, decompose_comments, explain, temperature)
            else:
                lm = lm + guidance_moderation_comment_one_step(
                    commentId, commentText, classify_options, results, thought, decompose_comments, explain, temperature)

        progress_bar.update() if progress_bar is not None else None

    return lm


@guidance
def guidance_moderation_comment_two_step(lm, commentId, commentText, classify_options, results, thought, explain, temperature, reject_reasons=[]):
    with user():
        lm += f"COMMENT {commentId}: {commentText}"

    with assistant():
        lm += "CLASSIFICATION: " + \
            select(classify_options, name="classification") + "\n"
        classification = lm['classification']
        reason = None
        classificationCode = 0
        thoughts = []

        if classification == classify_options[1] and explain:
            lm += "EXPLANATION: " + \
                generate_line("classification_explanation", temperature) + "\n"
            thoughts += [f"EXPLANATION: {lm['classification_explanation']}"]

        # if comment is REJECTED and second thought is enabled, allow model to reconsider
        if classification == classify_options[2] and thought:
            lm += f"""\
            THOUGHT: This comment is {generate_phrase("classification_thought", temperature)}. I should classify it as {generate_phrase("classification_thought_2", temperature)}.
            """
            thoughts += [
                f"THOUGHT: This comment is {lm['classification_thought']}.",
                f"I should classify it as {lm['classification_thought_2']}.",
            ]

            lm += f"REASON: " + \
                select(reject_reasons, name="classification_reason") + "\n"
            reason = lm['classification_reason']

            lm += f"Am I certain about this comment being {reason}? " + select(
                ["YES", "NO"], name="classification_certainty") + "\n"

            thoughts += [
                f"CERTAIN: {lm['classification_certainty']}",
            ]

            if lm['classification_certainty'] == "NO":
                classification = classify_options[1]
                lm += classify_options[1]

        # if comment is still REJECTED after second thought, or if thought is disabled
        if classification == classify_options[2]:
            classificationCode = -1

            # this indicates that second thought was not enabled and we never got the classification of rejected comment
            if reason is None:
                lm += f"CLASSIFICATION: " + \
                    select(reject_reasons, name="classification_reason") + "\n"
                reason = lm['classification_reason']

            if explain:
                lm += "EXPLANATION: " + \
                    generate_line("classification_explanation",
                                  temperature) + "\n"
                thoughts += [f"EXPLANATION: {lm['classification_explanation']}"]

        # if comment is ACCEPTED
        if classification == classify_options[0]:
            classificationCode = 1

        if classification == classify_options[2]:
            classification = reason

        results.addRow({
            'commentId': commentId,
            'languageModelModerated': classificationCode,
            'classification': classification,
            'thoughts': [thoughts]
        })

    return lm


@guidance
def guidance_moderation_comment_one_step(lm, commentId, commentText, classify_options, results, thought, decompose_comments, explain, temperature):
    with user():
        lm += f"COMMENT {commentId}: {commentText}"

    with assistant():
        moderated = None
        thoughts = []

        if decompose_comments:
            lm += "PROBLEM: " + generate_line('problem', temperature) + "\n"
            lm += "ACTION: " + generate_line('suggestion', temperature) + "\n"
            lm += "HOW MANY IDEAS: " + \
                generate_line('complexity', temperature) + "\n"
            thoughts += [
                f"PROBLEM: {lm['problem']}",
                f"ACTION: {lm['suggestion']}",
                f"IDEAS: {lm['complexity']}",
            ]

        if thought:
            lm += "THOUGHT: " + generate_line("thought", temperature) + "\n"
            thoughts += [f"THOUGHT: {lm['thought']}"]

        lm += "CLASSIFICATION: " + \
            select(classify_options, name="classification") + "\n"
        classification = lm['classification']

        if classification == classify_options[0]:
            moderated = 1
        else:
            moderated = 0 if classification == classify_options[1] else -1
            if explain:
                lm += "REASON: " + \
                    generate_line("classification_reason", temperature) + "\n"
                thoughts += [f"REASON: {lm['classification_reason']}"]

        results.addRow({
            'commentId': commentId,
            'languageModelModerated': moderated,
            'classification': classification,
            'thoughts': [thoughts],
        })

    return lm

Two-Step Moderation with Second-Thought Reasoning

First, determine whether the comment would be rejected. Then deduce the reason for rejection. Following is the language model prompt:

Code 
classify_options = ['ACCEPT', 'UNSURE', 'REJECT']
reject_reasons = ['SPAM', 'COMPLEX']

instructions_two_step = f"""\
Discussion Title: {summary.get('topic')}
Discussion Question: {summary.get('conversation-description')}

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

---
Classifications:
- {classify_options[0]}: Comment is coherent, makes a suggestion, or presents a real problem or issue.
- {classify_options[1]}: Unclear whether the comment meets the guidelines for {classify_options[0]}.
- {classify_options[2]}: Comment should definitely be rejected for one of the reasons listed below.

---
Reasons for {classify_options[2]}:
- 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: {", ".join(classify_options)}.
THOUGHT: Express the reasoning for {classify_options[2]} classification.
Am I certain: Answer with YES or NO. If unsure, state NO.
REASON: One of the following based on given guidelines: {", ".join(reject_reasons)}
EXPLANATION: Provide an explanation for why the comment was classified as {classify_options[2]}.
"""

print(instructions_two_step)
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.

Test 1: Baseline Technique

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_two_step,
    'two_step_strategy': True,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'reject_reasons': reject_reasons,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
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.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
CLASSIFICATION: UNSURE
EXPLANATION: This comment is not making a suggestion or presenting a real problem or issue. It is a statement of fact, but it is not clear whether another person would agree or disagree.

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
CLASSIFICATION: UNSURE
EXPLANATION: This comment is not making a suggestion or presenting a real problem or issue. It is a statement of fact, but it is not clear whether another person would agree or disagree.

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
CLASSIFICATION: REJECT
CLASSIFICATION: COMPLEX
EXPLANATION: This comment is stating more than one idea. It is difficult to determine the where another person would agree or disagree.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
CLASSIFICATION: UNSURE
EXPLANATION: This comment is not making a suggestion or presenting a real problem or issue. It is a statement of fact, but it is not clear whether another person would agree or disagree.

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
CLASSIFICATION: REJECT
CLASSIFICATION: COMPLEX
EXPLANATION: This comment is stating more than one idea. It is difficult to determine the where another person would agree or disagree.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
CLASSIFICATION: REJECT
CLASSIFICATION: COMPLEX
EXPLANATION: This comment is stating more than one idea. It is difficult to determine the where another person would agree or disagree.

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
CLASSIFICATION: ACCEPT

Run 1: Baseline Technique

Now we run the same test on all comments and visualize the results

Code 
progress_bar = tqdm(
    total=comments.df.height,
    desc='Comment Moderation',
    unit="comments",
    smoothing=0.1,
)

moderationResults1 = DataModel(DATASET, 'moderation-results-1', DATA_PATH, schema=results_schema)

args = {
    'instructions': instructions_two_step,
    'two_step_strategy': True,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp'),
    'results': moderationResults1,
    'reject_reasons': reject_reasons,
    'context_reset': True,
    'progress_bar': progress_bar,
}

languageModel + guidance_moderation(**args)

moderationResults1.save()
progress_bar.close()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()
pred_labels = moderationResults1.df.sort(
    'commentId').get_column('languageModelModerated').to_numpy()

plot_confusion_matrix(true_labels, pred_labels)

Test 2: Examples provided; second thought not enabled

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

examples_accepted = [
    "This problem requires some serious thought and reform from our elected leaders.",
    "We need better traffic management.",
    "Every individual should have a voice and be able to make a difference.",
]

examples_rejected = [
    (
        "lol why am i even here",
        "SPAM",
        "Incoherent statement that does not contribute to the discussion in any way."
    ),
    (
        "I agree with the previous comment",
        "SPAM",
        "Does not contribute to the discussion. Polis does not support responding to other comments."
    ),
    (
        "We need better policy for public welfare and economic development. Also, we should consider the impact on the environment.",
        "COMPLEX",
        "Introduces multiple unrelated problems. It is difficult to determine where another person would agree or disagree."
    ),
]

args = {
    'instructions': instructions_two_step,
    'two_step_strategy': True,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'examples_accepted': examples_accepted,
    'examples_rejected': examples_rejected,
    'reject_reasons': reject_reasons,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the guidelines.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

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?

CLASSIFICATION: ACCEPT, UNSURE, REJECT.
THOUGHT: If classifying as REJECT, express thoughts about the reasoning behind it.
REASON: SPAM, COMPLEX
EXPLANATION: Provide an explanation that explains why the comment was classified as REJECT.

Classifications:
- ACCEPT: Comment is consistent with the discussion and contributes to the conversation.
- UNSURE: Unsure whether the comment meets the guidelines for rejection.
- REJECT: Comment should definitely be rejected for one of the reasons listed below.

ACCEPT if the comment is coherent, makes a suggestion, or presents a real problem or issue.

REJECT if:
- SPAM: Clearly unconstructive, incoherent, or lacks seriousness.
- COMPLEX: Introduces multiple unconnected ideas using multiple sentences.

user
COMMENT EXAMPLE 1: This problem requires some serious thought and reform from our elected leaders.
assistant
ACCEPT
user
COMMENT EXAMPLE 2: We need better traffic management.
assistant
ACCEPT
user
COMMENT EXAMPLE 3: Every individual should have a voice and be able to make a difference.
assistant
ACCEPT
user
COMMENT EXAMPLE 4:
lol why am i even here
assistant
REJECT
COMMENT EXAMPLE 5 CLASSIFICATION: SPAM
EXPLANATION: Incoherent statement that does not contribute to the discussion in any way.

user
COMMENT EXAMPLE 5:
I agree with the previous comment
assistant
REJECT
COMMENT EXAMPLE 6 CLASSIFICATION: SPAM
EXPLANATION: Does not contribute to the discussion. Polis does not support responding to other comments.

user
COMMENT EXAMPLE 6:
Traffic management has been really bad
assistant
REJECT
COMMENT EXAMPLE 7 CLASSIFICATION: DUPLICATIVE
EXPLANATION: Repeats Comment EXAMPLE 2

user
COMMENT EXAMPLE 7:
We need better policy for public welfare and economic development. Also, we should consider the impact on the environment.
assistant
REJECT
COMMENT EXAMPLE 8 CLASSIFICATION: COMPLEX
EXPLANATION: Introduces multiple unrelated ideas where neither follows the other.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
ACCEPT

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
ACCEPT

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
ACCEPT

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
ACCEPT

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
REJECT
CLASSIFICATION: COMPLEX
EXPLANATION: Introduces multiple unrelated ideas where neither follows the other.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
ACCEPT

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
REJECT
CLASSIFICATION: COMPLEX
EXPLANATION: Introduces multiple unrelated ideas where neither follows the other.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
REJECT
CLASSIFICATION: COMPLEX
EXPLANATION: Introduces multiple unrelated ideas where neither follows the other.

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
ACCEPT

Test 3: Examples not provided; second thought enabled

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_two_step,
    'two_step_strategy': True,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'reject_reasons': reject_reasons,
    'thought': True,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the guidelines.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

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?

CLASSIFICATION: ACCEPT, UNSURE, REJECT.
THOUGHT: If classifying as REJECT, express thoughts about the reasoning behind it.
REASON: SPAM, COMPLEX
EXPLANATION: Provide an explanation that explains why the comment was classified as REJECT.

Classifications:
- ACCEPT: Comment is consistent with the discussion and contributes to the conversation.
- UNSURE: Unsure whether the comment meets the guidelines for rejection.
- REJECT: Comment should definitely be rejected for one of the reasons listed below.

ACCEPT if the comment is coherent, makes a suggestion, or presents a real problem or issue.

REJECT if:
- SPAM: Clearly unconstructive, incoherent, or lacks seriousness.
- COMPLEX: Introduces multiple unconnected ideas using multiple sentences.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
ACCEPT

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
UNSURE

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
ACCEPT

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

user
COMMENT 13: Recreational marijuana should be legal.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
ACCEPT

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
REJECT
THOUGHT: This comment is not disrespectful or distasteful, but it does not mention a problem or issue. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: The comment does not mention a problem or issue, and it does not contribute to the conversation. It is a statement that does not provide any value to the discussion.

Test 4: Examples provided and second thought enabled

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_two_step,
    'two_step_strategy': True,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'examples_accepted': examples_accepted,
    'examples_rejected': examples_rejected,
    'reject_reasons': reject_reasons,
    'thought': True,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the guidelines.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

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?

CLASSIFICATION: ACCEPT, UNSURE, REJECT.
THOUGHT: If classifying as REJECT, express thoughts about the reasoning behind it.
REASON: SPAM, COMPLEX
EXPLANATION: Provide an explanation that explains why the comment was classified as REJECT.

Classifications:
- ACCEPT: Comment is consistent with the discussion and contributes to the conversation.
- UNSURE: Unsure whether the comment meets the guidelines for rejection.
- REJECT: Comment should definitely be rejected for one of the reasons listed below.

ACCEPT if the comment is coherent, makes a suggestion, or presents a real problem or issue.

REJECT if:
- SPAM: Clearly unconstructive, incoherent, or lacks seriousness.
- COMPLEX: Introduces multiple unconnected ideas using multiple sentences.

user
COMMENT EXAMPLE 1: This problem requires some serious thought and reform from our elected leaders.
assistant
ACCEPT
user
COMMENT EXAMPLE 2: We need better traffic management.
assistant
ACCEPT
user
COMMENT EXAMPLE 3: Every individual should have a voice and be able to make a difference.
assistant
ACCEPT
user
COMMENT EXAMPLE 4:
lol why am i even here
assistant
REJECT
COMMENT EXAMPLE 5 CLASSIFICATION: SPAM
EXPLANATION: Incoherent statement that does not contribute to the discussion in any way.

user
COMMENT EXAMPLE 5:
I agree with the previous comment
assistant
REJECT
COMMENT EXAMPLE 6 CLASSIFICATION: SPAM
EXPLANATION: Does not contribute to the discussion. Polis does not support responding to other comments.

user
COMMENT EXAMPLE 6:
Traffic management has been really bad
assistant
REJECT
COMMENT EXAMPLE 7 CLASSIFICATION: DUPLICATIVE
EXPLANATION: Repeats Comment EXAMPLE 2

user
COMMENT EXAMPLE 7:
We need better policy for public welfare and economic development. Also, we should consider the impact on the environment.
assistant
REJECT
COMMENT EXAMPLE 8 CLASSIFICATION: COMPLEX
EXPLANATION: Introduces multiple unrelated ideas where neither follows the other.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
ACCEPT

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
ACCEPT

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
ACCEPT

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
ACCEPT

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
REJECT
THOUGHT: This comment is not productive and does not contribute to the discussion. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: This comment is not productive and does not contribute to the discussion. It is a harmful suggestion that would not help combat the opioid epidemic.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
ACCEPT

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
REJECT
THOUGHT: This comment is not productive and does not contribute to the discussion. I should classify it as REJECT.
CLASSIFICATION: SPAM
Am I certain about this comment being SPAM? YES
EXPLANATION: This comment is not productive and does not contribute to the discussion. It is a harmful suggestion that would not help combat the opioid epidemic.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
ACCEPT

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
ACCEPT

Run 4: Evaluate all comments using second-thought reasoning

Code 
progress_bar = tqdm(
    total=comments.df.height,
    desc='Comment Moderation',
    unit="comments",
    smoothing=0.1,
)

moderationResults4 = DataModel(DATASET, 'moderation-results-4', DATA_PATH, schema=results_schema)

args = {
    'instructions': instructions_two_step,
    'two_step_strategy': True,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp'),
    'results': moderationResults4,
    'examples_accepted': examples_accepted,
    'examples_rejected': examples_rejected,
    'reject_reasons': reject_reasons,
    'thought': True,
    'context_reset': True,
    'progress_bar': progress_bar,
}

languageModel + guidance_moderation(**args)

moderationResults4.save()
progress_bar.close()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()
pred_labels = moderationResults4.df.sort(
    'commentId').get_column('languageModelModerated').to_numpy()

plot_confusion_matrix(true_labels, pred_labels)

Discussion

Initially, the model somewhat follows instructions. Providing examples did not appear to improve the decisions. Enabling thought statement allows the model to reflect on the classification and reconsider the decision. This tends to improve reasoning abilities since the model states clearly what it intends to do before it makes a decision. However, in the following examples, it still appears to make incoherent decisions.

> COMMENT 12: Medical marijuana should be legal.

> REJECT
> THOUGHT: This comment is not constructive and does not contribute to the discussion. I should classify it as SPAM, but I will classify it as REJECT with the reason COMPLEX.
> CLASSIFICATION: COMPLEX
> Am I certain about this comment being COMPLEX? YES
> EXPLANATION: This comment is a statement of fact that does not contribute to the discussion.

In the example above, the model states its intention to classify the comment as SPAM but select COMPLEX instead. The explanation is logically sound but does not match the classification.

To address these issues, we simplify our instructions. Instead of ACCEPT/REJECT followed by a reject reason, we ask the model for a multi-label classification in single step that includes ACCEPTED, UNSURE, and several rejection categories. Since these categories are mutually exclusive, we do not need a separate step to determine the reason for rejection.

One-step Moderation

Categorize as ACCEPTED, UNSURE, SPAM, or COMPLEX

Code 
classify_options = ['ACCEPT', 'UNSURE', 'SPAM', 'COMPLEX']

instructions_one_step = f"""\
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the criteria.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

Discussion Title: {summary.get('topic')}
Discussion Question: {summary.get('conversation-description')}

---
Guidelines:
- ACCEPT: Comment mentions a problem, makes a suggestion, or recommends a solution.
- UNSURE: Unsure whether the comment should be accepted for discussion, especially if the comment appears harmful, unconstructive, or unhelpful.
- SPAM: Clearly incoherent or lacks seriousness.
- COMPLEX: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.

---
Output format:

THOUGHT: Consider the reasoning behind the classification.
CLASSIFICATION: {", ".join(classify_options)}.
REASON: If comment was not accepted, explain.
"""

print(instructions_one_step)
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the criteria.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

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?

---
Guidelines:
- ACCEPT: Comment mentions a problem, makes a suggestion, or recommends a solution.
- UNSURE: Unsure whether the comment should be accepted for discussion, especially if the comment appears harmful, unconstructive, or unhelpful.
- SPAM: Clearly incoherent or lacks seriousness.
- COMPLEX: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.

---
Output format:

THOUGHT: Consider the reasoning behind the classification.
CLASSIFICATION: ACCEPT, UNSURE, SPAM, COMPLEX.
REASON: If comment was not accepted, explain.

Test 5: Classify without thoughts

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_one_step,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'thought': False,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the criteria.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

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?

---
Guidelines:
- ACCEPT: Comment mentions a problem, makes a suggestion, or recommends a solution.
- UNSURE: Unsure whether the comment should be accepted for discussion, especially if the comment appears harmful, unconstructive, or unhelpful.
- SPAM: Clearly incoherent or lacks seriousness.
- COMPLEX: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.

---
Output format:

THOUGHT: Consider the reasoning behind the classification.
CLASSIFICATION: ACCEPT, UNSURE, SPAM, COMPLEX.
REASON: If comment was not accepted, explain.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
CLASSIFICATION: SPAM
REASON: This comment does not mention a problem, make a suggestion, or recommend a solution.

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
CLASSIFICATION: SPAM
REASON: This comment does not mention a problem, make a suggestion, or recommend a solution.

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
CLASSIFICATION: UNSURE
REASON: This comment suggests a solution, but it is unclear whether this solution would be effective or beneficial.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
CLASSIFICATION: SPAM
REASON: This comment does not mention a problem, make a suggestion, or recommend a solution.

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
CLASSIFICATION: SPAM
REASON: This comment does not mention a problem, make a suggestion, or recommend a solution.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
CLASSIFICATION: COMPLEX
REASON: This comment introduces two distinct ideas that should be separated into multiple comments.

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
CLASSIFICATION: ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
CLASSIFICATION: ACCEPT

Run 5: Simple moderation with no thought statements

In this run, we measure the effectiveness of our prompt without any chain-of-thought style reasoning as shown above in test 5. Context reset is enabled. This causes each comment to be removed from the context before the next one is processed, which limits the amount of context memory required and prevents us from going over the language model’s context limit. However, this also means that the model cannot use the context of previous comments to inform its decisions, limiting its ability to identify duplicate comments.

Code 
progress_bar = tqdm(
    total=comments.df.height,
    desc='Comment Moderation',
    unit="comments",
    smoothing=0.1,
)

moderationResults5 = DataModel(DATASET, 'moderation-results-5', DATA_PATH, schema=results_schema)


args = {
    'instructions': instructions_one_step,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp'),
    'results': moderationResults5,
    'thought': False,
    'decompose_comments': False,
    'context_reset': True,
    'progress_bar': progress_bar
}

languageModel + guidance_moderation(**args)

moderationResults5.save()
progress_bar.close()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()
pred_labels = moderationResults5.df.sort(
    'commentId').get_column('languageModelModerated').to_numpy()

plot_confusion_matrix(true_labels, pred_labels)

Test 6: Classify with thoughts

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_one_step,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'thought': True,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
You will be presented with comments posted on Polis discussion platform.
Classify each comment objectively based on whether it meets the criteria.
The comments do not have to provide a solution, but should mention a problem.
Comments should not be disrespectful or distasteful.

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?

---
Guidelines:
- ACCEPT: Comment mentions a problem, makes a suggestion, or recommends a solution.
- UNSURE: Unsure whether the comment should be accepted for discussion, especially if the comment appears harmful, unconstructive, or unhelpful.
- SPAM: Clearly incoherent or lacks seriousness.
- COMPLEX: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.

---
Output format:

THOUGHT: Consider the reasoning behind the classification.
CLASSIFICATION: ACCEPT, UNSURE, SPAM, COMPLEX.
REASON: If comment was not accepted, explain.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
THOUGHT: The comment mentions a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
THOUGHT: The comment mentions a problem and assigns blame.
CLASSIFICATION: ACCEPT

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
THOUGHT: The comment mentions a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
THOUGHT: The comment mentions a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
THOUGHT: The comment does not mention a problem or suggest a solution.
CLASSIFICATION: UNSURE
REASON: The comment appears to be a statement of fact, but it is not clear how it contributes to the discussion.

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
THOUGHT: The comment mentions a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
THOUGHT: The comment does not mention a problem or suggest a solution.
CLASSIFICATION: UNSURE
REASON: The comment appears to be a statement of fact, but it is not clear how it contributes to the discussion.

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
THOUGHT: The comment mentions a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
THOUGHT: The comment mentions a problem and assigns blame.
CLASSIFICATION: ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
THOUGHT: The comment does not mention a problem or suggest a solution.
CLASSIFICATION: UNSURE
REASON: The comment appears to be a statement of fact, but it is not clear how it contributes to the discussion.

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
THOUGHT: The comment does not mention a problem or suggest a solution.
CLASSIFICATION: UNSURE
REASON: The comment appears to be a statement of fact, but it is not clear how it contributes to the discussion.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
THOUGHT: The comment mentions a problem and assigns blame.
CLASSIFICATION: ACCEPT

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
THOUGHT: The comment suggests a solution to a problem.
CLASSIFICATION: ACCEPT

Run 6: Thought statement enabled

Here, we enable the thought statement to match the parameters of test 6.

Code 
progress_bar = tqdm(
    total=comments.df.height,
    desc='Comment Moderation',
    unit="comments",
    smoothing=0.1,
)

moderationResults6 = DataModel(DATASET, 'moderation-results-6', DATA_PATH, schema=results_schema)


args = {
    'instructions': instructions_one_step,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp'),
    'results': moderationResults6,
    'thought': True,
    'decompose_comments': False,
    'context_reset': True,
    'progress_bar': progress_bar
}

languageModel + guidance_moderation(**args)

moderationResults6.save()
progress_bar.close()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()
pred_labels = moderationResults6.df.sort(
    'commentId').get_column('languageModelModerated').to_numpy()

plot_confusion_matrix(true_labels, pred_labels)

Discussion

The above test produce more coherent results. Expressing thoughts before initial classification is more computationally expensive. However, it does improve results significantly. For example,

> COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.

> THOUGHT: The comment mentions a problem and points a finger at the responsible party.
> CLASSIFICATION: ACCEPT

The above comment was accepted with strong reasoning. The same comment was often rejected in previous tests based on lack of accuracy or evidence, which are not valid moderation criteria. The comment meets Polis guidelines for being presented to the community for voting.

> COMMENT 14: I regularly discuss politics with people I disagree with.

> THOUGHT: The comment does not mention a problem or suggest a solution.
> CLASSIFICATION: UNSURE
> REASON: The comment appears to be a general statement that does not contribute to the discussion.

The above comment does not present any useful information, and is a candidate for additional screening. Instead of outright rejecting it, the model indicated unsurity, which allows us to flag the comment and escalate it for huamn review.

Decompose Comments

Decomposing each comment into its requisite components might further help improve the analysis and evaluation abilities of the model. While we do not care about the accuracy or agreeability of the posted comments, we do expect either a problem or a potential solution. Also, we want the comment to present a single idea.

In the following prompt, we ask the model to idenfity the problem the comments address and their proposed solution. If neither is present, the comment is considered spam.

Code 
classify_options = ['ACCEPT', 'UNSURE', 'REJECT']

instructions_decompose = f"""\
Discussion Title: {summary.topic}
Discussion Question: {summary.get('conversation-description')}

---
Classify each comment objectively based on the following guidelines.

Mark the comment as:
- REJECT: frivolous, irrelevant, unrelated to the discussion.
- REJECT: does not contribute to the discussion in a meaningful way.
- REJECT: incoherent or lacks seriousness.
- REJECT: provides neither a problem nor a solution.
- REJECT: the language is informal, colloquial, disrespectful or distasteful.
- REJECT: cannot be addressed within the scope of original question.
- REJECT: introduces multiple ideas, even if they are related to the discussion.
- REJECT: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.
- 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.
- UNSURE: may be accepted if it appears somewhat related to the discussion.

---
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: One of the following based on given guidelines: {", ".join(classify_options)}.
EXPLANATION: Provide an explanation for the classification.
"""

print(instructions_decompose)
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:
Mark the comment as:
- ACCEPT: mentions a real problem related to the discussion.
- ACCEPT: makes a sincere suggestion or recommends a realistic and actionable solution.
- SPAM: Clearly incoherent or lacks seriousness.
- SPAM: provides neither a problem nor a solution.
- SPAM: the problem is frivolous, irrelevant, unrelated to the discussion question, or lacks seriousness.
- SPAM: the language is informal, colloquial, disrespectful or distasteful.
- SPAM: cannot be addressed within the scope of original question.
- COMPLEX: introduces multiple ideas, even if they are related to the discussion.
- COMPLEX: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.
- UNSURE: may or may not be accepted especially if it appears unconstructive or unhelpful.

---
Response 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.

Test 7: Ask Language Model to Decompose Comments

Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_decompose,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'thought': True,
    'decompose_comments': True,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
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:
Mark the comment as:
- ACCEPT: mentions a real problem related to the discussion.
- ACCEPT: makes a sincere suggestion or recommends a realistic and actionable solution.
- SPAM: Clearly incoherent or lacks seriousness.
- SPAM: provides neither a problem nor a solution.
- SPAM: the problem is frivolous, irrelevant, unrelated to the discussion question, or lacks seriousness.
- SPAM: the language is informal, colloquial, disrespectful or distasteful.
- SPAM: cannot be addressed within the scope of original question.
- COMPLEX: introduces multiple ideas, even if they are related to the discussion.
- COMPLEX: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.
- UNSURE: may or may not be accepted especially if it appears unconstructive or unhelpful.

---
Response 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.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
PROBLEM IDENTIFIED: The pension fund crisis in the local schools.
SUGGESTED ACTION: Local schools should learn to live with budget cuts.
HOW MANY IDEAS: 1
THOUGHT: This comment mentions a real problem and suggests a solution, albeit a vague one.
CLASSIFICATION: ACCEPT

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
PROBLEM IDENTIFIED: Tuition at the university has risen significantly, and the university management is being held responsible.
SUGGESTED ACTION: None.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem but does not suggest a solution.
CLASSIFICATION: UNSURE
REASON: This comment is unclear about what change it suggests.

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
PROBLEM IDENTIFIED: Officials blocking constituents on social media.
SUGGESTED ACTION: Officials should not block constituents on social media except under rare conditions.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
PROBLEM IDENTIFIED: City residents who pay city taxes but live in the county cannot vote in city elections.
SUGGESTED ACTION: City residents who pay city taxes but live in the county should be allowed to vote in city elections.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
PROBLEM IDENTIFIED: None.
SUGGESTED ACTION: None.
HOW MANY IDEAS: 1
THOUGHT: This comment is a statement of fact and does not introduce a new problem or solution.
CLASSIFICATION: SPAM
REASON: This comment is not relevant to the discussion question.

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
PROBLEM IDENTIFIED: Local law enforcement may lack sufficient training in community relations.
SUGGESTED ACTION: Local law enforcement should receive additional training in community relations.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
PROBLEM IDENTIFIED: None.
SUGGESTED ACTION: None.
HOW MANY IDEAS: 1
THOUGHT: This comment is a statement of fact and does not introduce a new problem or solution.
CLASSIFICATION: SPAM
REASON: This comment is not relevant to the discussion question.

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
PROBLEM IDENTIFIED: Opinion pieces without a specific author can make it difficult to determine media bias.
SUGGESTED ACTION: Opinion pieces should have a specific author to make it easier to determine media bias.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
PROBLEM IDENTIFIED: Institutionalized racism in the media's coverage of drug-related deaths.
SUGGESTED ACTION: The media should provide equal coverage of drug-related deaths in all communities.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
PROBLEM IDENTIFIED: The opioid epidemic.
SUGGESTED ACTION: Harsher sentences for drug users.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution, but the solution is controversial and may not be effective in addressing the problem.
CLASSIFICATION: UNSURE
REASON: This comment suggests a solution that is not actionable and may not be effective in addressing the problem.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
PROBLEM IDENTIFIED: The opioid epidemic.
SUGGESTED ACTION: View drug addiction as a health problem rather than a criminal justice problem.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
PROBLEM IDENTIFIED: Drug possession arrests leading to imprisonment rather than rehabilitation.
SUGGESTED ACTION: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
PROBLEM IDENTIFIED: Medical marijuana is not legal.
SUGGESTED ACTION: Medical marijuana should be legal.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
PROBLEM IDENTIFIED: Recreational marijuana is not legal.
SUGGESTED ACTION: Recreational marijuana should be legal.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
PROBLEM IDENTIFIED: None.
SUGGESTED ACTION: None.
HOW MANY IDEAS: 1
THOUGHT: This comment is a statement of fact and does not introduce a new problem or solution.
CLASSIFICATION: SPAM
REASON: This comment is not relevant to the discussion question.

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
PROBLEM IDENTIFIED: Local community colleges may not offer courses that interest those seeking an education if they do not aid in job placement.
SUGGESTED ACTION: Local community colleges should offer courses that interest those seeking an education, whether or not the courses aid in job placement.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
PROBLEM IDENTIFIED: None.
SUGGESTED ACTION: None.
HOW MANY IDEAS: 1
THOUGHT: This comment is a statement of opinion and does not introduce a new problem or solution.
CLASSIFICATION: SPAM
REASON: This comment is not relevant to the discussion question.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
PROBLEM IDENTIFIED: Private schools receiving a share of Federal or State education dollars may hurt public schools.
SUGGESTED ACTION: Private schools should not receive a share of Federal or State education dollars.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
PROBLEM IDENTIFIED: None.
SUGGESTED ACTION: Implement beautification projects and other quality-of-life enhancements like parks and libraries.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a potential solution but does not explicitly state a problem.
CLASSIFICATION: UNSURE
REASON: This comment suggests a solution without identifying a problem.

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
PROBLEM IDENTIFIED: Nonviolent inmates convicted of drug possession are not being set free.
SUGGESTED ACTION: More nonviolent inmates convicted of drug possession should be set free.
HOW MANY IDEAS: 1
THOUGHT: This comment identifies a problem and suggests a solution.
CLASSIFICATION: ACCEPT

The above results might be improved through context reset, which bounds the context memory and allows the model to focus on the current comment. We will test this hypothesis in the following test.

Run 7: Decompose Comments

In this run, we decompose each comment into its requisite components to improve the analysis and evaluation abilities of the model. This run matches the parameters of test 7.

Code 
progress_bar = tqdm(
    total=comments.df.height,
    desc='Comment Moderation',
    unit="comments",
    smoothing=0.1,
)

moderationResults7 = DataModel(DATASET, 'moderation-results-7', DATA_PATH, schema=results_schema)

args = {
    'instructions': instructions_decompose,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp'),
    'results': moderationResults7,
    'thought': True,
    'decompose_comments': True,
    'context_reset': True,
    'progress_bar': progress_bar
}

languageModel + guidance_moderation(**args)

moderationResults7.save()
progress_bar.close()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()
pred_labels = moderationResults7.df.sort(
    'commentId').get_column('languageModelModerated').to_numpy()

plot_confusion_matrix(true_labels, pred_labels)

The above run took about 6 seconds per comment on H100 GPU, which is about 90 minutes for the entire dataset of ~900 comments. This is computationally expensive but appears to yield significantly better results.

Test 8: Further refining the prompt

Code 
classify_options = ['ACCEPT', 'UNSURE', 'SPAM',
                    'IRRELEVANT', 'UNPROFESSIONAL', 'SCOPE', 'COMPLEX']

instructions_decompose_2 = f"""\
Discussion Title: {summary.topic}
Discussion Question: {summary.get('conversation-description')}

---
Classify each comment objectively based on the following guidelines:
- IRRELEVANT: frivolous, irrelevant, unrelated to the discussion.
- IRRELEVANT: does not contribute to the discussion in a meaningful way.
- INCOHERENT: 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: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.
- 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.
- UNSURE: may be accepted if it appears somewhat related to the discussion.

---
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: One of the following based on given guidelines: {", ".join(classify_options)}.
EXPLANATION: Provide an explanation for the classification.
"""

print(instructions_decompose_2)
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:
Mark the comment as:
- IRRELEVANT: frivolous, irrelevant, unrelated to the discussion.
- IRRELEVANT: does not contribute to the discussion in a meaningful way.
- INCOHERENT: 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: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.
- 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.
- UNSURE: may be accepted if it appears somewhat related to the discussion.

---
Response 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, IRRELEVANT, UNPROFESSIONAL, SCOPE, COMPLEX.
REASON: If comment was not classified as ACCEPT, explain.
Code 
moderationResults = DataModel(DATASET, schema=results_schema)

args = {
    'instructions': instructions_decompose_2,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp').head(20),
    'results': moderationResults,
    'thought': True,
    'decompose_comments': True,
    'explain': True,
}

languageModel + guidance_moderation(**args)
instruction
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:
Mark the comment as:
- IRRELEVANT: frivolous, irrelevant, unrelated to the discussion.
- IRRELEVANT: does not contribute to the discussion in a meaningful way.
- INCOHERENT: 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: Introduces distinct unconnected ideas using complex sentences that should be separated into multiple comments.
- 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.
- UNSURE: may be accepted if it appears somewhat related to the discussion.

---
Response 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, IRRELEVANT, UNPROFESSIONAL, SCOPE, COMPLEX.
REASON: If comment was not classified as ACCEPT, explain.

user
COMMENT 0: Local schools should help make up the pension fund crisis by learning to live with budget cuts.
assistant
PROBLEM: The pension fund crisis in the state.
ACTION: Local schools should learn to live with budget cuts.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a solution to a problem, but it is not a realistic or actionable solution.
CLASSIFICATION: UNSURE
REASON: The comment is not realistic or actionable because it is not clear how budget cuts for local schools would help solve the pension fund crisis.

user
COMMENT 1: University management bears more responsibility than federal and state education funding cuts for tuition rising so much over the years.
assistant
PROBLEM: Tuition has risen significantly over the years.
ACTION: University management should take more responsibility for the rise in tuition.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, but it is not a realistic or actionable solution.
CLASSIFICATION: UNSURE
REASON: The comment is not realistic or actionable because it is not clear how university management can be held more responsible for the rise in tuition.

user
COMMENT 2: Officials that use social media accounts in a professional capacity should not be allowed to block constituents except under rare conditions like being threatened.
assistant
PROBLEM: Officials blocking constituents on social media.
ACTION: Officials should not be allowed to block constituents on social media except under rare conditions.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 3: I should be able to vote in city elections if I pay city taxes, even though I may live in the county.
assistant
PROBLEM: City residents who pay city taxes but live in the county are not able to vote in city elections.
ACTION: City residents who pay city taxes but live in the county should be able to vote in city elections.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 4: Most local officials run for office to serve the public good.
assistant
PROBLEM: None
ACTION: None
HOW MANY IDEAS: 1
THOUGHT: This comment does not contribute to the discussion in a meaningful way.
CLASSIFICATION: IRRELEVANT
REASON: The comment does not mention a problem or a solution.

user
COMMENT 5: Local law enforcement would benefit from additional training in community relations.
assistant
PROBLEM: Local law enforcement may not have adequate community relations training.
ACTION: Local law enforcement should receive additional training in community relations.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 6: Most journalists strive to be factual and unbiased, whether or not they succeed in all their writing,
assistant
PROBLEM: None
ACTION: None
HOW MANY IDEAS: 1
THOUGHT: This comment does not contribute to the discussion in a meaningful way.
CLASSIFICATION: IRRELEVANT
REASON: The comment does not mention a problem or a solution.

user
COMMENT 7: Opinion pieces without a specific author make it more confusing to determine whether a media source is biased or not.
assistant
PROBLEM: It can be difficult to determine whether a media source is biased or not when opinion pieces do not have a specific author.
ACTION: Opinion pieces should have a specific author to make it easier to determine whether a media source is biased or not.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 8: The fact that there is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids is evidence of institutionalized racism.
assistant
PROBLEM: There is not as much coverage about African American deaths due to cocaine as there is about deaths in the white community from opioids.
ACTION: There should be more coverage about African American deaths due to cocaine.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 9: Harsher sentences for drug users would help combat the opioid epidemic.
assistant
PROBLEM: The opioid epidemic.
ACTION: Harsher sentences for drug users.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a solution to a problem, but it is not a realistic or actionable solution.
CLASSIFICATION: UNSURE
REASON: The comment is not realistic or actionable because harsher sentences for drug users have not been shown to combat drug epidemics.

user
COMMENT 10: In order to better combat the opioid epidemic, it is time to view drug addiction as a health problem rather than a criminal justice problem.
assistant
PROBLEM: The opioid epidemic.
ACTION: View drug addiction as a health problem rather than a criminal justice problem.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 11: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
assistant
PROBLEM: Drug possession arrests leading to imprisonment.
ACTION: Law enforcement should send drug possession arrestees for rehabilitation services rather than imprisonment.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 12: Medical marijuana should be legal.
assistant
PROBLEM: Medical marijuana is not legal.
ACTION: Medical marijuana should be legal.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 13: Recreational marijuana should be legal.
assistant
PROBLEM: Recreational marijuana is not legal.
ACTION: Recreational marijuana should be legal.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 14: I regularly discuss politics with people I disagree with.
assistant
PROBLEM: None
ACTION: None
HOW MANY IDEAS: 1
THOUGHT: This comment does not contribute to the discussion in a meaningful way.
CLASSIFICATION: IRRELEVANT
REASON: The comment does not mention a problem or a solution.

user
COMMENT 15: WKU and local community colleges should continue to offer courses that interest those seeking an education, whether or not the courses aid in job placement.
assistant
PROBLEM: Courses that interest those seeking an education are not offered by WKU and local community colleges if they do not aid in job placement.
ACTION: WKU and local community colleges should offer courses that interest those seeking an education, whether or not the courses aid in job placement.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 16: The main purpose of higher education is job training.
assistant
PROBLEM: None
ACTION: None
HOW MANY IDEAS: 1
THOUGHT: This comment does not contribute to the discussion in a meaningful way.
CLASSIFICATION: IRRELEVANT
REASON: The comment does not mention a problem or a solution.

user
COMMENT 17: If private schools were to get a share of Federal or State education dollars, it would hurt public schools.
assistant
PROBLEM: Private schools receiving a share of Federal or State education dollars would hurt public schools.
ACTION: Private schools should not receive a share of Federal or State education dollars.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 18: Beautification projects and other quality-of-life enhancements like parks and libraries help attract new residents and industries to our city/county.
assistant
PROBLEM: Bowling Green/Warren County lacks beautification projects and other quality-of-life enhancements like parks and libraries.
ACTION: Bowling Green/Warren County should invest in beautification projects and other quality-of-life enhancements like parks and libraries.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

user
COMMENT 19: More nonviolent inmates convicted of drug possession should be set free.
assistant
PROBLEM: Nonviolent inmates convicted of drug possession are not being set free.
ACTION: More nonviolent inmates convicted of drug possession should be set free.
HOW MANY IDEAS: 1
THOUGHT: This comment suggests a problem and a solution, and it is a realistic and actionable solution.
CLASSIFICATION: ACCEPT

Run 8: Refine Decomposition Prompt

Code 
progress_bar = tqdm(
    total=comments.df.height,
    desc="Comment Moderation",
    unit="comments",
    smoothing=0.1,
)

moderationResults8 = DataModel(DATASET, 'moderation-results-8', DATA_PATH, schema=results_schema)

args = {
    'instructions': instructions_decompose_2,
    'classify_options': classify_options,
    'comments_df': comments.df.sort('timestamp'),
    'results': moderationResults8,
    'thought': True,
    'decompose_comments': True,
    'context_reset': True,
    'progress_bar': progress_bar
}

languageModel + guidance_moderation(**args)

moderationResults8.save()
progress_bar.close()

true_labels = comments.df.sort('commentId').get_column('moderated').to_numpy()
pred_labels = moderationResults8.df.sort(
    'commentId').get_column('languageModelModerated').to_numpy()

plot_confusion_matrix(true_labels, pred_labels)

Further Analysis

Notes: Prompt Refinements

REJECT: - does not provide a solution or mention a problem - informal and colloquial - relevant to posted question - The comment introduces multiple ideas, but they are all related to improving the area. - if no problem identified, say so - if no solution identified, say so - Cannot be addressed within the scope of original question.

Comment: Bowling Green needs a homeless shelter. (Moderated: -1) - PROBLEM: Bowling Green lacks a homeless shelter. - ACTION: Establish a homeless shelter in Bowling Green. - THOUGHT: The comment identifies a problem and proposes a solution.

No problem identified - only a solution

Comment: The driving conditions on the 31-W bypass are very poor and need to be changed. (Moderated: -1) - PROBLEM: The driving conditions on the 31-W bypass are very poor. - ACTION: Change the driving conditions on the 31-W bypass. - THOUGHT: The comment mentions a problem and suggests a solution, so it should be accepted.

No solution identified - only a problem

Comment: Bowling Green should become a Metro. (Moderated: -1) - PROBLEM: Bowling Green does not have a metro system. - ACTION: Implement a metro system in Bowling Green.

Comment: Know who all your neighbors are (Moderated: -1) - PROBLEM: Lack of familiarity with neighbors. - ACTION: Encourage getting to know neighbors. - IDEAS: 1 - THOUGHT: The comment mentions a problem and suggests a solution.

Cannot be addressed within the scope of original question.

Future Work

We should consider the effect of temperature on spam classification. For the above experiment, we kept temperature = 0 to limit randomness and allow reproducibility of results. However, a higher temperature might allow the model to explore more options and improve its reasoning abilities, especially when employing the chain-of-thought reasoning approach.