Skip to content

Latest commit

 

History

History
137 lines (110 loc) · 2.87 KB

README.md

File metadata and controls

137 lines (110 loc) · 2.87 KB
Raw

Chitchat Bot

This project use Tensorflow to make a simple chitchat bot in Chinese. The chatbot use RNN model, and is trained on Chinese corpus.

Dataset

Chinese conversation corpus from this repo.

  1. Download dataset

wget https://raw.githubusercontent.com/rustch3n/dgk_lost_conv/master/dgk_shooter_min.conv.zip

  1. Unzip the zip file, we get a dgk_shooter_min.conv file, move it to data folder. The format of conv file is:
.conv 格式:
//M 表示话语，E 表示分割。
E
M 话语 a
M 话语 b
M 话语 c
M 话语 d
E
M 话语 a
M 话语 b
M 话语 c
M 话语 d

For example

E
M 畹/华/吾/侄/
M 你/接/到/这/封/信/的/时/候/
M 不/知/道/大/伯/还/在/不/在/人/世/了/
E
M 咱/们/梅/家/从/你/爷/爷/起/
M 就/一/直/小/心/翼/翼/地/唱/戏/
M 侍/奉/宫/廷/侍/奉/百/姓/
M 从/来/不/曾/遭/此/大/祸/
M 太/后/的/万/寿/节/谁/敢/不/穿/红/
M 就/你/胆/儿/大/
M 唉/这/我/舅/母/出/殡/

Pre-process

  1. Remove / , and group into question and answer

python pre_process.py

  1. Build vocabulary, and convert texts into vecotrs according to their index

python vectorizer.py

The generated files: train_encode.vec and train_decode.vec are used for training, whose vocabulary is train_encode_vocabulary and train_decode_vocabulary, respectively.

Train

We use seq2seq model on the training data. It will about 24 hours to get the current performance.

python train.py

After about 589,500 steps, we get the following result:

587000 0.000305144 2.56671503615
0 12.543833172375583
1 23.283587395237618
2 20.160063524352733
3 14.362551288543015
587500
587500 0.000305144 2.59014926028
0 8.500202516716408
1 16.222189910741516
2 19.48035661490604
3 16.742266002404577
588000
588000 0.000295989 2.58287223005
0 10.108112100344743
1 15.532271967787727
2 21.85774124267814
3 16.732469326616826
588500
588500 0.000295989 2.60331172109
0 10.217336668387876
1 15.90760132489462
2 20.418236969921555
3 17.01274752262256
589000
589000 0.00028711 2.57389337921
0 8.111649969765645
1 17.354193061727415
2 18.568189964633955
3 15.888793235232294
589500
589500 0.00028711 2.57721918631

The training loss is 2.57, and the PPL is about 15.

Decoding

Use the trained model to similate chat bot:

python decoding.py

Testing result

me > 你好
AI > 你好
me > 你吃饭了吗
AI > 吃了
me > 见到你真好
AI > 我也是
me > 我爱你
AI > 我也爱你
me > 你结婚了吗
AI > 我没有
me > 你是女生还是男生
AI > 是
me > 女生还是男生
AI > 我们是
me > 你的回答好糟糕
AI > 你是说，我们不是在谈论
me > 今天天气怎么样
AI > 我们会再见到你的
me > 你在哪里
AI > 我在哪里
me > 谢谢你了
AI > 别客气
me > 再见
AI > 再见