EFE Reasoner

This repository contains the implementation of the Explicit Feature Extraction (EFE) Reasoner, a model designed to improve reasoning about numerical magnitudes in math word problems. The method extends the Deductive-Reasoner architecture by explicitly providing token embeddings for numbers, enabling the model to capture ordering relationships between quantities.

The approach is described in the paper "Explicit Feature Extraction(EFE) Reasoner: A model for Understanding the Relationship between Numbers by Size".

Features

• Adds explicit number features to the Deductive-Reasoner framework
• Uses pre-trained language models such as RoBERTa to embed problem text
• Demonstrated accuracy gains on the SVAMP dataset

Setup

Install the required Python packages:

pip install -r requirements.txt

Example Usage

Run training with the default configuration:

python main.py \
  --wandb 0 \
  --data_path data/processed/svamp \
  --bert_model roberta-base

Hyperparameters can be adjusted in main.py via the get_*_args functions.

Data

The repository expects preprocessed SVAMP-style datasets in the data/ directory. Prepare the dataset separately before running the code.

Code Structure

• datasets/ – dataset loading utilities
• model/ – EFE Reasoner implementation (wrapper_model.py etc.)
• main.py – training and evaluation script

License

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