Skip to content

Latest commit

 

History

History
114 lines (90 loc) · 4.52 KB

README.md

File metadata and controls

114 lines (90 loc) · 4.52 KB

Project Overview

This project aims to enhance question generation by introducing a feedback mechanism within the loss function that calculates the similarity between generated and target questions. This feedback-driven approach helps the model produce more accurate and contextually relevant questions.

This project includes the following main steps:

Data Preprocessing

Prepare the SQuAD dataset with fields for 'context,' 'question,' and 'answer.'

Fine-tuning T5 on SQuAD Train a T5 model on the SQuAD dataset, focusing on generating questions from context and answer pairs.

Feedback-Enhanced Loss

Modify the loss function to include feedback by measuring the similarity between generated and target questions, which allows for iterative improvement in question relevance.

The enhanced loss function incorporates feedback to achive more accurate and contextually aligned question generation, is defined as follows:

$Total \ Loss = \alpha * loss + (1 - \alpha) * secondary \ loss \ weight * reward$

where:

  • $loss$: The primary loss from the model's output.
  • $reward$: A feedback term that measures the similarity between the generated question and the target question.
  • $\alpha$: A weighting factor that balances the influence of the primary loss and the feedback.
  • $secondary \ loss \ weight$: A coefficient that adjusts the impact of the feedback term.

Evaluation

The model's performance is evaluated using two automated metrics: METEOR and ROUGE-L, which assess accuracy and alignment between generated and target questions. Below are the results:

Model METEOR ROUGE-L
T5 0.31 0.20
T5 + Feedback-Enhanced Loss 0.33 0.21

These results illustrate the model's improvement in question generation accuracy and relevance after implementing the feedback mechanism.

Ablation Study

An ablation study was conducted to experiment with various parameters and configurations to optimize model performance. This study involved adjusting parameters such as dataset size and feedback weights ($\alpha$ and $secondary \ loss \ weight$ to reach the best-performing model.

Train the Model From Scratch

  1. Create Directories for Data and Model
mkdir data model
  1. Install the dependencies
pip install -r requirements.txt
  1. Run python preprocessing.py to preprocess the data
python preprocessing.py
  1. Run python transformer_model.py to train/eval the model with hyper-parameters found in config.py
python transformer_model.py

Quick Start: Using the Trained Model

This guide will walk you through the steps to quickly get started with using the Question Generation model for generating questions from context and answer.

  1. Install the dependencies
pip install -r requirements.txt
  1. Clone the Repository
git clone https://github.com/ZahraRahimii/Improving-Question-Generation-using-Transformer-Based-Models
cd Improving-Question-Generation-using-Transformer-Based-Models
  1. Load the Model and Tokenizer
from transformers import T5ForConditionalGeneration, T5Tokenizer

# Load the pre-trained model and tokenizer
model = T5ForConditionalGeneration.from_pretrained(os.path.join('model', 't5_trained_model_20'))
tokenizer = T5Tokenizer.from_pretrained(os.path.join('model', 't5_tokenizer_20'))

# Input text to generate questions
context = "The Pacific Ocean is the largest and deepest of Earth's oceanic divisions."
answer = "The Pacific Ocean"
text = "context: " + context + " " + "answer: " + answer

# Tokenize the input
input_ids = t5_tokenizer.encode(text, return_tensors="pt", max_length=400, truncation=True)

# Generate question
output_ids = t5_model.generate(input_ids=input_ids, max_length=40)
generated_question = t5_tokenizer.decode(output_ids[0], skip_special_tokens=True)

print(f"Generated Question: {generated_question}")