DualGeoSolver

Official implementation of the IJCAI-24 paper Learning to Solve Geometry Problems via Simulating Human Dual-Reasoning Process.

Datasets & Checkpoints

All the datasets (GeoQA and GeoQA+) and checkpoints can be found in here.

Unzip the tarball by running:

tar -zxvf GeoQA-Data.tar.gz

The GeoQA dataset is in GeoQA-Pro folder and GeoQA+ dataset is in GeoQA+ folder. The GeoQA+ dataset is slightly different from the original one, because we have fix some annotation errors in original GeoQA+. For example, there are some program sequences which have only one operand following a binary operator in the original GeoQA+ dataset.

Our constructed knowledge base is presented in geometry-knowledges-chinese.json and geometry-knowledges-english.json for Chinese version and English version respectively (in directory of each dataset). While training, we only utilize the Chinese version. The GPT-3.5-Turbo annotated geometry knowledge for each reasoning step is presented in geometry-knowledge-data-map.json.

Run

Preparation

Firstly, download ViTMAE checkpoint from here and put it in your working directory.

Secondly, download Chinese-RoBERTa checkpoint from huggingface and put it in chinese-roberta-wwm-ext directory.

Now, your working directory should looks like this:

|--config
|
|--chinese-roberta-wwm-ext
|
|--vit-64-base-160epoch
|
|--GeoQA-Data
|  |--GeoQA-Pro
|  |--GeoQA+
|
|--ManualProgram
|
|--requirements.txt
|
|--train.py
...

Training

You can adjust hyper parameters in config/solver.json. After you finish this, run:

allennlp train config/solver.json --include-package train -s /path/to/save

Evaluating

Run:

allennlp evaluate /path/to/checkpoint_dir GeoQA-Data/GeoQA-Pro/pro_test.pk --include-package evaluate --cuda-device 0

Refactor codes are coming soon ...