A Pretraining Numerical Reasoning Model for Ordinal Constrained KBQA

This is the code for the paper: A Pretraining Numerical Reasoning Model for Ordinal Constrained Question Answering on Knowledge Base in EMNLP 2021 Findings.

An advanced model to deal with the same problem can be found in the following paper: Injecting Numerical Reasoning Skills into Knowledge Base Question Answering Models in arxiv preprint.

Requirements

• Python 3.8
• Pytorch >= 1.6

Dataset

Download preprocessed datasets from google drive, and unzip it into dataset folder.

Reasoning

To train and evaluate the reasoning model, change to directory ./NumReasoning. You can also download checkpoints from here, and unzip it into checkpoint folder.

Basic Reasoning

To test the original reasoning model.

python main_nsm.py --model_name gnn --data_folder ../CWQ/ --checkpoint_dir ../checkpoint/CWQ_num/ --experiment_name eval_CWQ_gnn_num_50epoch
--entity2id entities_expanded.txt --batch_size 40 --test_batch_size 40 --num_step 4 --entity_dim 50 --word_dim 300 --node_dim 50 --eval_every 1 
--eps 0.95 --num_epoch 50 --use_self_loop --lr 1e-4 --q_type seq --word_emb_file word_emb_300d.npy --reason_kb --encode_type --loss_type kl 
--load_experiment CWQ_nsm-h1.ckpt --is_eval

Num Reasoning

To train the Num reasoning model.

python main_nsm.py --model_name gnn --data_folder ../CWQ/ --checkpoint_dir ../checkpoint/CWQ_num/ --experiment_name CWQ_gnn_num_50epoch
--entity2id entities_expanded.txt --batch_size 40 --test_batch_size 40 --num_step 4 --entity_dim 50 --word_dim 300 --node_dim 50 --eval_every 1  
--eps 0.95 --num_epoch 50 --use_self_loop --lr 1e-4 --q_type seq --word_emb_file word_emb_300d.npy --reason_kb --encode_type --loss_type kl 
--use_num --use_nsm_num --relation_embedding_file ../CWQ/cwq_rel_embedding.npy 
--load_num CWQ_num_model.pth --load_experiment CWQ_nsm-h1.ckpt

To test the Num reasoning model.

python main_nsm.py --model_name gnn --data_folder ../CWQ/ --checkpoint_dir ../checkpoint/CWQ_num/ --experiment_name eval_CWQ_gnn_num_50epoch
--entity2id entities_expanded.txt --batch_size 40 --test_batch_size 40 --num_step 4 --entity_dim 50 --word_dim 300 --node_dim 50 --eval_every 1  
--eps 0.95 --num_epoch 50 --use_self_loop --lr 1e-4 --q_type seq --word_emb_file word_emb_300d.npy --reason_kb --encode_type --loss_type kl 
--use_num --use_nsm_num --relation_embedding_file ../CWQ/cwq_rel_embedding.npy 
--load_num CWQ_num_model.pth --load_experiment CWQ_gnn_num_50epoch-h1.ckpt --is_eval