Meena chatbot

Here’s my attempt at recreating Meena, a state of the art chatbot developed by Google Research and described in the paper Towards a Human-like Open-Domain Chatbot.

For this implementation I used the tensor2tensor deep learning library, using an evolved transformer model as described in the paper.

The training set used is the OpenSubtitles corpus in the Italian language. Many other languages are available here.


Similarly to the work done in the paper, this model consists of 1 encoder block and 12 decoder blocks for a total of 108M parameters. The optimizer used is Adafactor with the same training rate schedule as described in the paper.


Here are the results after training the model on 40M sentences of the OpenSubtitles dataset in the italian language. The learning rate starts at 0.01 and remains constant for 10k steps then decay with the inverse square root of the number of steps.

Learning rate schedule

Here’s the plot of the evaluation loss during training. Evaluation loss plot

The final perplexity score is 10.4 which is very close to the perplexity score achieved by Google’s meena chatbot 10.2.

The paper shows a correlation between perplexity score and the Sensibleness and Specificity Average which is correlated with the “human likeness” of the chatbot. Our perplexity score shows that our bot is better than other chatbots such as Cleverbot and DialoGPT: Perplexity SSA correlation

The dataset used however does not represent well normal conversations between humans. However Opensubtitles provide very large datasets in many languages.

Run pretrained model

Simply run notebook meena_chatbot_inference.ipynb.

Otherwise download the following model and extract it. Set proper MODEL_DIR and CHECKPOINT_NAME in and run

Pretrained model checkpoint

Train a new model

For training simply run the ipython notebook on Google Colab, the model will be saved on Google Drive. At the end of the execution you can interact with the chatbot.

Export the model

The model can be exported by copying the following files in a folder:

  • hparams.json
  • The trained model checkpoint
  • The vocabulary .subwords file

and run after setting the proper model directory.

Serving provides a simple HTTP API for serving the chatbot.