Download
This repository contains a new generative model of chatbot based on seq2seq modeling. The trained model available here used a small dataset composed of ~8K pairs of context (the last two utterances of the dialogue up to the current point) and respective response. The data were collected from dialogues of English courses online. This trained model can be fine-tuned using a closed-domain dataset to real-world applications. The canonical seq2seq model became popular in neural machine translation, a task that has different prior probability distributions for the words belonging to the input and output sequences since the input and output utterances are written in different languages. The architecture presented here assumes the same prior distributions for input and output words. Therefore, it shares an embedding layer (Glove pre-trained word embedding) between the encoding and decoding processes through the adoption of a new model.
Features
- The algorithm iterates by including the predicted token into the incomplete answer
- Chat with the pre-trained model
- Chat with the new model trained by our new GAN-based training algorithm
- Our model can be applied to other NLP tasks
- Contains a new generative model of chatbot based on seq2seq modeling
- This trained model can be fine-tuned using a closed domain dataset to real-world applications
Project Samples
