Software Requirement Engineering using Machine Learning Techniques

Project Overview

This repository demonstrates the use of Retrieval-Augmented Generation (RAG) and Prompt Engineering to classify software requirements into functional (F) and non-functional requirements (NFRs). These techniques leverage Large Language Models (LLMs) to enhance classification accuracy by combining retrieval and context-aware generation.

---

Features

Prompt Engineering

Prompt Engineering uses optimized task-specific prompts with few-shot learning to guide the LLMs in understanding and classifying requirements effectively.

Highlights:

• Leverages representative examples for accurate predictions.
• Processes LLM responses with a parser for consistent output.

---

Retrieval-Augmented Generation (RAG)

RAG improves classification by retrieving contextually relevant examples from a semantic vector database and integrating them into the prompt.

Workflow:

1. Retrieve: Use embeddings to find semantically similar examples from a database (e.g., Pinecone).
2. Generate: Append retrieved examples to the prompt and process through an LLM.
3. Parse: Ensure output aligns with predefined categories.

---

Results

Method	Model	F1	Accuracy
Prompt Engineering	GPT-3.5-Turbo	96.03	74.74
RAG	GPT-3.5-Turbo (RAG)	96.63	79.79

---

Getting Started

Prerequisites

• Python 3.x
• Jupyter Notebook
• Libraries: transformers, pinecone, openai, etc. (see requirements.txt)

Dataset

The classification is performed using the PROMISE and PROMISE_exp datasets:

1. PROMISE Dataset:

   • Contains 625 natural language software requirements.
   • Includes:
     • 255 functional requirements (F).
     • 370 non-functional requirements (NFRs) distributed across categories like security, usability, performance, etc.

2. PROMISE_exp Dataset:

   • An extended version of PROMISE with 969 requirements.
   • Includes:
     • 444 functional requirements (F).
     • 525 non-functional requirements (NFRs) distributed across additional categories like maintainability, scalability, and fault tolerance.

Dataset Split:

• Training: 80%
• Testing: 20%

Non-Functional Categories in PROMISE_exp:

Category	Count
Availability (A)	31
Fault Tolerance (FT)	18
Legal (L)	15
Look & Feel (LF)	49
Maintainability (MN)	24
Operability (O)	77
Performance (PE)	67
Scalability (SC)	22
Security (SE)	125
Usability (US)	85
Portability (PO)	12
Total	969

The dataset provides a comprehensive basis for training and testing, ensuring diverse and real-world software requirement scenarios are covered.

Model Architectures

---

Usage

You can .

---

License

This project is licensed under the MIT License. See the LICENSE file for details.