This is an implementation for An Enhanced Knowledge Injection Model for Commonsense Generation (Paper) in COLING 2020.

pip install --user fairseq==0.9.0

Please refer to the github repo for data donwloading and some baselines.

We also provide our retrieved and preprocessed data in the BaiduDisk with code 36iv. In detail, all1v4 is retrieved from in-domain dataset, and wikiv4 is retrieved from out-of-domain dataset (wiki). You should download the data from BaiduDisk and put them in this main directory.

Train the model.

bash src/train2.sh

Evaluate the model.

python ./src/generate2.py --task translation_segment2
    --search_tag '' 
    --arch bart9a_large
    --model_problem wiki
    --problem wiki
    --version 1 
    --gpu 0 
    --bsz 12
    --lenpen 0.0 
    --beam 5 
    --maxlen 24 
    --minlen 3 
    --ngram 3