Emotional Chatting Machine

This reponsitory includes two models

• a basic seq2seq model with attention and beamsearch
• ECM model

Thanks to https://github.com/AaronYALai/Seq2seqAttn_ECM, I learn a lot from this reponsitory. However, I can't get a satisfactory results on the chinese corpus when I use AaronYALai's model. So, I implement a new ECM model based on the tensorflow seq2seq API.

Requirements

• python 2.7
• tensorflow == 1.4

Sample data

sample data here is only for showing the data format, not for training.

• category: target sentence emotion category
• choice: target sentence emotional word annotation
• source: source sentence
• target: target sentence

Just tell me how it works

Set up work space

Create a new folder by following the parameters "workspace" in the yaml configuration file. for example:

./works/example/

Prepare your data and configuration file

you can check the sample data folder for the data format.Then you need to put them under path like

./works/example/data/

Train the model

for training the basic model:

python train.py

for training ECM model:

python train_ECM.py

Infer

• "infer_ECM.py" will first create a calculation graph of infer model then load the trained parameters, and finally perform the inference, which is not suitable for deployment.

python infer_ECM.py

• "save_infer_model.py" will first create a calculation graph, load the training parameters, and then save the infer model as a model file. At this time, you can use different languages of tensorflow API(C++/Java) to load the infer model.

python save_infer_model.py
python infer_ECM_deploy.py

Model performance

The following result is based on the dataset I crawled from BaiDu Tieba, including 3 emotion types:

• no emotion: 321052
• pos emotion: 137086
• neg emotion: 240233

Parameters:

embeddings:
   embed_size: 300
   vocab_size: 40000
encoder:
   bidirectional: True
   cell_type: LSTM
   num_layers: 2
   num_units: 512
decoder:
   attn_num_units: 512
   cell_type: LSTM
   num_layers: 2
   num_units: 512
   state_pass: True
   infer_max_iter: 25
   emo_cat_emb_size: 256
   emo_internal_memory_units: 256
   num_emotion: 3

Training perplexity

Extra dataset

NTCIR Short Text Conversation Task(STC-3) chinese Emotional Conversation Generation (CECG) Subtask http://coai.cs.tsinghua.edu.cn/hml/challenge/dataset_description/ You can also find the dataset in this project.