The code for paper "Learning to Predict Concept Ordering for Common Sense Generation"

---

Dataset

The dataset we use are stored in the commongen folder, you could unzip the data in the submission folder. In the folder, the "commongen.SPLIT.src_alpha.txt" and "commongen.SPLIT.tgt.txt" are the original files provided by CommonGen dataset.

Model

The model we use in our experiments are stored in the methods folder. If you want to run the code, you need to create the virtual environment for each model:

    conda create -n MODEL python=3.8
    conda activate MODEL
    pip install -r requirements.txt

,where MODEL is the name you want to do the corrsponding experiment. Then you could run the bash files in each sub-folder. We also write the key hyper-parameters in the bash files. We could provide the trained models if needed.

• We use the evaluation metrics (BLEU/ROUGE/METEOR/CIDEr/SPICE/Coverage) provided by CommonGen project to evaluate the generation quality. For the Kendall's tau for the concept ordering, you could run the code under methods/Probabilistic/util.py