Preparation
git clone https://github.com/jackyideal/LogiGraph.git
cd LogiGraph
pip install -r requirements.txt
In logical reasoning, capturing explicit logical relationships between nodes is crucial. Traditional GCN may lose some information due to complex feature transformations and non-linear activation.
To address these issues, we propose a lightweight GCN. This network discards feature transformation, non-linear function activation, symmetric normalization term, and self-connection in the traditional GCN. By simplifying the model structure and focusing on the linear aggregation of neighborhood information, the lightweight GCN better preserves and utilizes logical relationships, thereby enhancing model performance in logical reasoning tasks.
Experiments involve the evaluation of our model effectiveness on two datasets, ReClor and LogiQA, which constitute a diverse range of logical reasoning skills. The ReClor dataset is based on different exams, with unbiased instances separated from the test data to evaluate logical reasoning. On the other hand, the LogiQA dataset is derived from the National Civil Servants Examination of China. It has been translated professionally into English.
- LogiGraph: The result of LogiGraph based on the DeBERTa-v2-xlarge encoder.
If you use LogiGraph in your work, please cite our paper.