The Intelligent Social Event Observer: Multi-Source Continuous Event Integration, Discovery, and Synopsis Generation with LLMs

✒️ Overview

SEO is a novel framework to integrate evolving events, detect emerging events, and automatically generate event synopsis for social networks in real time.

👇 Dataset

We evaluate our method for event detection on the News14 and WCEP19 datasets. The process files are in the path ./data/processed.

We evaluate our method for synopsis generation on the News14-detail, WCEP19-detail, News14-integrity, and WCEP19-integrity datasets. The datasets are in the path ./data/evaluation.

✨ Quick Start

Step1: Write a configuration file in YAML format

Users can easily configure the parameters of LLMs in a YAML file. The path of the configuration file is SEO/config/config.yaml

openai:
  api_key: # your_openai_API_key
  base_url: 
  temperature: 0.2  
  max_tokens: 2048
llama:
  temperature: 0.2
  max_tokens: 2048
prompts:
  prompt: config/prompts.yaml

Step2: Running the model

python main.py

⏰ Optimized Inference with vLLM

To maximize inference efficiency, we recommend leveraging the vLLM framework for batched and cached inference.

Installation Instructions, requiring CUDA 11.8+:

pip install torch==2.3.0+cu118 -f https://download.pytorch.org/whl/torch_stable.html
pip install xformers==0.0.26.post1 vllm==0.5.1+cu118

📖 Citation

Please cite our repository if you use SEO in your work.