COLAB LINIK : https://colab.research.google.com/drive/1fPAEiVf6ldzq3sneEI2Mcpm_Bz1bD-7U?usp=sharing

PDF LINK: https://arxiv.org/pdf/2109.00122

DATASET LINK: https://www.kaggle.com/datasets/visalakshiiyer/question-answering-financial-data/data

---

Built a modular neuro-symbolic architecture with BERT retriever, LSTM generator, and symbolic executor to answer complex financial questions over tables. Followed a structured ML lifecycle for data preprocessing, modular training, integration, and iterative evaluation. This enables precise reasoning and scalable development.

🧠 FinQA: Program Generation for Numerical Reasoning over Financial Text

FinQA (Financial Question Answering) is a system that performs numerical reasoning over financial documents such as earnings reports. It generates programs (step-by-step symbolic operations) to answer complex numerical questions that require arithmetic, logic, and contextual understanding.

This repository implements a complete Retriever + Generator + Executor pipeline inspired by the FinQA paper, with enhancements for modularity and extensibility.

---

📌 Table of Contents

• Project Overview

• Pipeline Architecture

• Components

  • 1. Retriever
  • 2. Generator
  • 3. Executor

• Dataset

• Training & Evaluation

• Results

• Model Enhancements

• References

---

📄 Project Overview

The goal of FinQA is to answer financial questions that require reasoning over:

• Financial tables (structured data)
• Pre-text and post-text paragraphs (unstructured financial text)

Instead of directly predicting the answer, the model generates a reasoning program, executes it over the document, and derives the final result.

---

Machine Learning Lifecycle

🧠 Problem → 📊 Data → ⚙️ Preprocessing → 🧱 Model Design → 🎯 Training → 🔗 Integration → 📈 Evaluation → 🔄 Iteration → 🚀 Deployment

🏗️ Pipeline Architecture

 ┌────────────┐     ┌───────────────┐     ┌────────────┐     ┌────────────┐
 │ Financial  │     │    Retriever  │     │  Generator │     │  Executor  │
 │ Document   │───▶ │ (BERT-based)  │───▶ │ (LSTM + BERT)│───▶│ Math Ops   │───▶Final Answer
 └────────────┘     └───────────────┘     └────────────┘     └────────────┘
                                                    

1. Retriever: Selects relevant sentences/tables from the document.
2. Generator: Generates a step-by-step reasoning program.
3. Executor: Executes the program over the table/text values to get the final answer.

---

🧩 Components

1. Retriever

• Model: BERT/RoBERTa-based encoder trained for sentence-level relevance classification.
• Inputs: Full document (pre-text, table, post-text) and question.
• Outputs: Binary labels for each sentence/table cell indicating its relevance.
• Loss Function: Binary Cross-Entropy

Enhancements:

• Option to use cosine similarity + neural classifier
• Handles table row/column vectorization

---

2. Generator

• Encoder: BERT-based contextual encoder

• Decoder: LSTM-based program generator

• Vocabulary:

  • Reserved operations: add, subtract, multiply, divide, greater, etc.
  • DSL tokens: intermediate variables #0, #1, ...
  • Table values and numeric constants from context

• Step Memory: Tracks previously generated operations (#0, #1) for multi-step reasoning

• Training:

  • Teacher forcing with program supervision
  • Cross-entropy loss over DSL tokens

• Output: List of DSL operations (program)

Example:

Program: [select(table_value_1), select(table_value_2), add(#0, #1)]

---

3. Executor

• Symbolic interpreter for the DSL

• Executes operations like add, subtract, select, greater, etc.

• Supports:

  • Numeric precision
  • Intermediate variable storage
  • Type checking for safe math operations

---

📊 Dataset

• Based on the official FinQA dataset

• JSON structure per sample:

  {
    "pre_text": "...",
    "post_text": "...",
    "table": [["Header1", "Header2", ...], ["Row1", "Row2", ...], ...],
    "qa": "What is the revenue growth?",
    "program": ["select(...)", "divide(...)", "subtract(...)"],
    "gold_inds": [sentence indices used by retriever]
  }

• Custom preprocessing steps include:

  • Tokenization
  • Sentence splitting
  • Table value indexing
  • Operation mapping

---

---

📈 Results

Component	Accuracy	Notes
Retriever	~84% (sentence-level F1)	BERT-based
Generator	~68% (program match accuracy)	LSTM decoder
Executor	100% (symbolic interpreter)	Executes correct programs

---

🔧 Model Enhancements

• ✅ Custom tokenizer and vocabulary expansion
• ✅ Integration of table schema as headers
• ✅ Attention over both table and text
• ✅ Robust step memory handling
• ✅ Support for program templates
• ✅ DSL support for numerical + logical ops

---

📚 References

• FinQA: Financial Question Answering via Program Generation
• Official FinQA GitHub Repo

---

MODEL

infer