chongyangtao / ioi

Tensorflow implementation for IOI model

IOI

This is an implementation of [One Time of Interaction May Not Be Enough: Go Deep with an Interaction-over-Interaction Network for Response Selection in Dialogues, ACL 2019].

Requirements

• Ubuntu 16.04
• Tensorflow 1.4.0
• Python 3.5

Usage

Dowload ubuntu corpus and preprocess the data, run

# download ubuntu corpus and the pre-trained word2vec file 
sh download.sh

# preprocess the data
python data_utils_record.py

All hyper parameters are stored in config.py. To train, run

python main.py --log_root=logs_ubuntu --batch_size=20

To evaluate the model, run

python evaluate.py --log_root=logs_ubuntu --batch_size=20

ioi's Issues

Could you share the dataset?

Thanks for the code guys. Could you also share the exact data that was used to obtain the results reported in the paper (i.e., train, dev, test sets and vocabulary)?

In the mean time, I experimented with the Ubuntu Corpus dataset (ver2) and got following not-so-good results for the development set. Probably did something wrong when preparing the data...

Dev results..........
('pred_scores: ', 500000)
Step: 66000 	| loss: 2.805 	| acc: 0.860 	|  2019-09-15 13:27:36
recall_2_1:  0.904
recall_at_1: 0.657
recall_at_2: 0.800
recall_at_5: 0.949

Also, I think there is a minor bug in the code "main.py" (line 220): Definitions of variables "minimal_val_step" and "minimal_val_loss" are missing.

about the word embedding

Hello, Thanks for sharing the code, I have a small question, is the word embedding pre-trained on all the training data or truncated data (limit context turns and utterance length)?

About multi GPU training

Does it support multi-GPU training?

Cannot download the data.

Hello, I am going to reproduce your paper of IOI, but I have a trouble in downloading the data. Could you please share the ubuntu_data and its word2vec model in other ways that are available for those who in China?