📄 Template-Based Financial Report Generation in Agentic and Decomposed Information Retrieval

Official Implementation of Template-Based Financial Report Generation in Agentic and Decomposed Information Retrieval
To appear in the Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2025)

📄 Paper

📎 Template-Based Financial Report Generation in Agentic and Decomposed Information Retrieval

👤 Authors

Yong-En Tian

Yu-Chien Tang

Kuang-Da Wang

---

🧾 Overview

This repository provides two generation pipelines (AgenticIR and DecomposedIR) for financial and climate report generation.
It supports task decomposition, retrieval-based generation, and multiple evaluation strategies.

📦 Requirements

Set up the Python environment using Conda:

conda create -n myenv python=3.12.2
conda activate myenv
pip install -r requirements.txt

👉 Note: This project requires a .env file for API keys. Create a file named .env in the root directory and include:

OPENAI_API_KEY=your_openai_api_key_here

📁 Project Structure

.
├── Finance/                      # Financial data (to be released)
├── SumIPCC/                      # Climate report generation
├── crawler.py                    # Script to crawl financial data
├── requirements.txt              # Dependencies
├── .env
└── README.md                     # This file         

📊 Finance

📚 Contents

• AgenticIR/ – Agent-based generation

• DecomposedIR/ – Report generation via decomposed queries

• data/ – Earnings call transcripts, balance-sheet, cash-flow and income statement

🔧 Features

1. Agentic Report Generation:

   • Uses agents to generate financial reports.
   • Supports task decomposition and retrieval-based generation.

2. Decomposed Report Generation:

   • Breaks down report generation into smaller subquery for improved accuracy.
   • Embedding-based retrieval of relevant data chunks.

3. Evaluation Tools:

   • Supports DecompEval and G-Eval metrics for reference-free evaluating the quality of generated summaries and readability of generated summaries.

🚀 Usage Instructions

🧠 AgenticIR

1. Generate Reports

# Arguments:
# -r, self-reflection
# -c, company (="INTC", "GFS", "SSNLF", "TSM", "UMC")
python Finance/AgenticIR/agentic_main.py -c INTC [-r]

2. Evaluate Reports

# Arguments:
# -c, company (="INTC", "GFS", "SSNLF", "TSM", "UMC")
python Finance/AgenticIR/Eval/eval.py -c INTC

3. View Evaluation Results

Open notebook:
Finance/AgenticIR/Eval/result_eval.ipynb

🧩 DecomposedIR

1. Generate Reports

# Arguments:
# -r, self-reflection
# -c, company (="INTC", "GFS", "SSNLF", "TSM", "UMC")
python Finance/DecomposedIR/decomposed_main.py -c INTC [-r]

2. Evaluate Reports

python Finance/DecomposedIR/Eval/eval.py

3. View Evaluation Results

Open notebook:
Finance/DecomposedIR/Eval/result_eval.ipynb

🌍 SumIPCC (Climate Report Generation)

📚 Contents

• AgenticIR/ – Agent-based generation

• DecomposedIR/ – Report generation via decomposed queries

• data/ – Original and processed SumIPCC data

🔧 Features

1. Agentic Report Generation:

   • Uses agents to generate climate reports.
   • Supports task decomposition and retrieval-based generation.

2. Decomposed Report Generation:

   • Breaks down report generation into smaller subquery for improved accuracy.
   • Embedding-based retrieval of relevant data chunks.

3. Evaluation Tools:

   • Supports ROUGE and BERTScore metrics for evaluating the quality of generated summaries.
   • Compares generated summaries with ground truth data.

4. Data Processing:

   • Processes datasets to create structured templates and ground truth for evaluation.

🚀 Usage Instructions

🧠 AgenticIR

1. Generate Reports

# Arguments:
# -r, self-reflection
python SumIPCC/AgenticIR/agentic_main.py [-r]

2. Evaluate Reports

python SumIPCC/AgenticIR/Eval/eval.py

3. View Evaluation Results

Open notebook:
SumIPCC/AgenticIR/Eval/result_eval.ipynb

🧩 DecomposedIR

1. Generate Reports

# Arguments:
# -r, self-reflection
python SumIPCC/DecomposedIR/decomposed_main.py [-r]

2. Evaluate Reports

python SumIPCC/DecomposedIR/Eval/eval.py

3. View Evaluation Results

Open notebook:
SumIPCC/DecomposedIR/Eval/result_eval.ipynb

📂 Data Format

Original and processed SumIPCC data are organized as follows:

SumIPCC/data/
├── all_data/                       # Raw SumIPCC data
└── process_data/
    └── Report_1/                   # Example
        ├── AgenticIR_result/       # Agent-based outputs
        ├── DecomposedIR_result/    # Decomposed outputs
        ├── full_paragraphs.txt     # Original report
        ├── groundtruth.py          # Ground truth summaries
        ├── question_decomposed.py  # Query decompositions
        └── report_template.py      # Report template structure

📝 License

This project is licensed under the MIT License.

📄 Citation

Please cite our work if you find it helpful:

@inproceedings{tian2025template,
  title={Template-based financial report generation in agentic and decomposed information retrieval},
  author={Tian, Yong-En and Tang, Yu-Chien and Wang, Kuang-Da and Yen, An-Zi and Peng, Wen-Chih},
  booktitle={Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages={2706--2710},
  year={2025}
}