pascalson / dykgchat

Dynamic Knowledge-Grounded Dialogue Generation. Tensorflow. The data and code of paper "DyKgChat: Benchmarking Dialogue Generation Grounding on Dynamic Knowledge Graphs, EMNLP 2019"

DyKGChat

The project contains the collected data and code of our paper Yi-Lin Tuan, Yun-Nung Chen, Hung-yi Lee. "DyKgChat: Benchmarking Dialogue Generation Grounding on Dynamic Knowledge Graphs", EMNLP 2019.

The extended abstract version is called Dynamic Knowledge-Grounded Dialogue Generation through Walking on the Graph.

• our proposed approach: (Qadpt) Quick Adaptive Dynamic Knoledge-Grounded Neural Converation Model (pronouce: Q-adapt)

Setup

Installation (my environment)

• python3.6
• tensorflow r1.13
• jieba
• nltk3.2.5

Files

• data/: the collected data hgzhz/ and friends/ as well as the trained TransE
• model_ckpts/: the trained models in the paper

Usage

• clone the repository
• run the script run.sh

$bash run.sh <GPU_ID> <method> <model> <data> <exp_name>

• for <GPU_ID>, check your device avalibility by nvidia-smi
• for , choose from train, pred_acc, eval_pred_acc, ifchange
• for , choose from seq2seq, MemNet, TAware, KAware, Qadpt
• for , choose from friends, hgzhz_v1_0(used in our paper), hgzhz(current newest version)
• for <exp_name>, check the directory model_ckpts

More description

• testing method
  • pred_acc: for metrics Generated-KW, BLEU-2, distinct-n
  • eval_pred_acc: for metrics KW-Acc, KW/Generic, perplexity
  • ifchange: for change rates / accurate change rates
• script options
  • the hops_num and change_level are required to be changed in run.sh

dykgchat's Issues

At test stage, how to build the subgraph?

Hi,

Thank for your released data and code.
In this paper, you use the input message and scene info to decide the source-entity and use the response to decide the target-nodes, and use the source and target node to build the subgraph.
However, in the inferening stage, how to get the target nodes to build the subgraph whitout the response info which can decide the target nodes.

Thanks