README

Hi! Please see readme.log for example output.

you may found some hard-code API_KEY in codes but don't worry the api is fully managed by myself and have a restriction.

Feel free to use it!

work.pdf contains brief intro of out works'architecture and the results of the experiments.

File Function

This script simulates an AI-driven multi-agent dialogue system using multiple custom ConversableAgents and third-party libraries (e.g., OpenAI's GPT model) to collaboratively process user input for analysis and debate on topics related to the cryptocurrency market.

Main Modules and Features

1. Dependency Importing

   • Use a variety of external dependency libraries (e.g. autogen-agentchat, autogen-ext[openai,web-surfer], playwright, etc.).
   • Load the .env file to set the API key and model URL configuration.

2. **ConversableAgent Definitions

   • Defines multiple ConversableAgents with a clear division of responsibilities (news selection, debate support, summary logging, etc.):
     • newsSelector: Filters and selects news related to the cryptocurrency market for the last seven days, blocking advertisements and irrelevant information.
     • debateAdviser: analyses user comments and stores them sorted by pro and con.
     • assistantRecorder: Records and summarizes the main points of a two-agent dialogue.
     • debateParticipantA/B: Provide clear, relevant arguments as the pro and con sides of the debate.
     • judge: Judge the strengths and weaknesses of both sides of the debate and choose a winner.
     • host: moderates the flow of the dialogue, forcing the discussion to focus on the given topic. 3.

3. Dialogue Functions: Implement the `host.initiate

   • implements the host.initiate_chats method, which assigns tasks to agents to simulate debates and analysis around cryptocurrency topics.
   • Each dialogue task has a custom input message, maximum number of rounds, summary strategy, etc.

4. Data processing

   • Read the results.json file, parse and construct user social media comments from the cryptocurrency market as input context.
   • Inject the parsed content into the user's initial message to provide dialogue context.

5. Asynchronous main functions

   • Run the above multi-agent dialogue logic asynchronously using asyncio.
   • Generate dialogue result output at the end of the logic.

Core Technical Points

• Multi-Agent Collaboration: Define multiple ConversableAgents, each with its own role, to form a complex multi-agent dialogue and collaboration process.
• API Integration: Provide language generation and information retrieval capabilities for agents by configuring OpenAI GPT-4 models and tools from Google Search.
• Dynamic Context Management: Loads dynamic contexts for analysis by reading external JSON files and environment variables

The file main.py implements a multi-intelligence based dialogue and analysis script focused on the cryptocurrency domain. It simulates an efficient information extraction, debate and summary process in the domain through a reasonable division of labour and context injection.

main.py implements a multi-intelligence dialogue system that collaboratively generates summaries and decisions through analysis, filtering, and debate around topics related to the cryptocurrency market.

Module/Function Points	Description
Dependency Import	Introduces external libraries (e.g. autogen-ext and OpenAI API) and tools (e.g. dotenv and playwright) for functionality extensions.
ConversableAgent Definition	Define multiple agents (e.g., news filtering, comment analysis, debate participation, etc.) to implement different business logic responsibilities.
Conversation Functionality	Use host to moderate agent conversations, simulating the process of debating and analyzing cryptocurrency market topics.
	Data Processing
Output Generation	Aggregate the results of the dialogue to generate a final summary and key points of the market-related debate.
Asynchronous main function	Uses asyncio to drive the entire dialogue logic asynchronously, efficiently scheduling agent tasks for collaborative processing.

Recommendation: Add exception handling and logging to each module to improve robustness.

Translated with DeepL.com (free version)

TF-IDF Score Formula

$$text{TF-IDF Score} = \sum_{i = 1}^{N} \text{TF-IDF}(w_{i})$$

Definition of Variables

$( N )$ = Total number of words in the document (tweet).
$( w_{i} ) = The ( i )-th word in the document. $ $( {TF-IDF}(w_{i})) = \text The TF-IDF score of word ( w_{i} )$, calculated as:

$\text{TF-IDF}(w_{i}) = \text{TF}(w_{i}) \times \text{IDF}(w_{i})$

where:

• Term Frequency (TF): The frequency of word $( w_{i} ) $in the document:

$text{TF}(w_{i}) = \frac{\text{Number of times } w_{i} \text{ appears in the document}}{\text{Total number of words in the document}}$

• Inverse Document Frequency (IDF): A measure of how unique the word is across the entire corpus:

$text{IDF}(w_{i}) = \log \left(\frac{\text{Total number of documents}}{\text{Number of documents containing } w_{i}} + 1\right)$

Interpretation

• Higher TF-IDF scores indicate that a word is important in the document (tweet) but not common across all documents.
• Words that appear frequently in one document but rarely in others will have higher importance scores.

Normalize Engagement Metrics

Each engagement metric is normalized using Min-Max Scaling with 95th Percentile Clipping:

1. Normalized Reply Count

$\text{Norm Reply} = \frac{\min(\text{tweet reply count}, Q_{95}^{\text{reply}}) - \min(\text{tweet reply count})}{Q_{95}^{\text{reply}} - \min(\text{tweet reply count}) + 1e^{-8}}$

2. Normalized Quote Count

$text{Norm Quote} = \frac{\min(\text{tweet quote count}, Q_{95}^{\text{quote}}) - \min(\text{tweet quote count})}{Q_{95}^{\text{quote}} - \min(\text{tweet quote count}) + 1e^{-8}}$

3. Normalized Conflict Ratio

$text{Conflict Ratio} = \frac{\text{tweet reply count} + \text{tweet quote count}}{\text{tweet like count} + \text{tweet retweet count} + 1}$

$text{Norm Ratio} = \frac{\min(\text{Conflict Ratio}, Q_{95}^{\text{ratio}}) - \min(\text{Conflict Ratio})}{Q_{95}^{\text{ratio}} - \min(\text{Conflict Ratio}) + 1e^{-8}}$

Explanation of Variables

$( Q_{95}^{\text{reply}}, Q_{95}^{\text{quote}}, Q_{95}^{\text{ratio}} ) $= 95th percentile values for each metric.

• Clipping ensures outliers don’t dominate the scores. $( 1e^{-8} )$ prevents division by zero.

Interpretation

• Each metric is scaled between 0 and 1.
• Values are adjusted to reduce the impact of outliers.
• Higher values indicate greater engagement relative to the dataset.

Compute Engagement Score

The Engagement Score is calculated as the sum of various engagement metrics:

$E = \text{tweet reply count} + \text{tweet retweet count} + \text{tweet like count} + \text{tweet quote count} + \text{tweet bookmark count}$

Explanation of Variables

( E ) = Engagement Score (total interactions on a tweet). $( \text{tweet reply count} ) = Number of replies to the tweet.$ $( \text{tweet retweet count} ) = Number of retweets.$ $( \text{tweet like count} ) = Number of likes.$ $( \text{tweet quote count} ) = Number of quote tweets.$ $( \text{tweet bookmark count} ) = Number of times the tweet was bookmarked.$

Interpretation

• Higher engagement scores indicate higher user interaction with the tweet.
• The score considers all forms of engagement, not just likes or retweets.
• Tweets with high engagement scores are more likely to be influential or viral.

Compute Raw Importance Score

The raw importance score is calculated as the product of the TF-IDF Score and the Engagement Score:

$\text{Importance Score} = \text{TF-IDF Score} \times \text{Engagement Score}$

Explanation of Variables

$( \text{Importance Score} ) = The raw measure of a tweet’s importance.$ $( \text{TF-IDF Score} ) $= The importance of words in the tweet based on Term Frequency-Inverse Document Frequency (TF-IDF):

$text{TF-IDF Score} = \sum_{i=1}^{N} \text{TF-IDF}(w_i)$

where: $( N ) = Total number of words in the tweet.$ $( w_i ) = The ( i )-th word in the tweet.$ $( \text{TF-IDF}(w_i) ) = The TF-IDF value of word ( w_i ).$

$( \text{Engagement Score} ) = The total interaction with a tweet:$

$\text E = {tweet reply count} + \text{tweet retweet count} + \text{tweet like count} + \text{tweet quote count} + \text{tweet bookmark count} $

Interpretation

• The TF-IDF Score captures textual importance based on how unique and relevant the words are.
• The Engagement Score captures user interaction with the tweet.
• Multiplication ensures that only highly engaged tweets with important textual content receive higher importance scores.
• This raw score will later be normalized to a 0-100 scale for better interpretability.