The project is compelted in Google AI Winter Camp. There are another repo1 and repo2 from my teamates.
The motivation of our project is that when we communicate with others, our conversations are not only based on the logic and context, but also our personality. People tend to chat with whom has similar characters. So we wonder if we can create a chatting bot with personality that can provide more enjoyable user experiences.
Our project consists of two parts. Firstly, analyze one’s personality basing on current and historic dialog using MBTI personality classification indicators. Secondly, concatenate the personality analysis above with current dialog embedding as chatting generation model’s input. Then seq-2-seq model is used to generate chatting response with personality.
The overview of our model is:
In the discrimination model, we use the Kaggle MBTI dataset which includes one’s last 50 posts (on-line behaviors) and personality label from which we extracted 30 posts randomly. We fed biLSTM model (self-attention mechanism based) with Elmo pre-trained model’s embedding.
In the generation model, we concatenate the dialog embedding with personality analysis from the discrimination model as the model’s input, and use seq-to-seq model to generate the response.
The model outputs both personality scores and dialogue response.
