Deep2ch is a chatbot for generating text using various datasets. This chatbot based on several models, such as neural networks and Markov chains. It creates texts on different topics. The bot was tested on responses to real users on the 2ch.hk forum.
Mechanics and outcomes:
The system explores the previously downloaded dataset and learns how to generate similar phrases and sentences. It analyzes the most used words, typical starting and ending words of the sentences, as well as average number of words in posts with Q&A.
For example, there are many topical threads (discussions of participants) on the 2chan forum. The model learns style and content of the post, and generates text based on this data. Then the bot posts generated text to the thread. Participants of the thread think that a real person is communicating with them and continue the conversation — the most persistent person interacted with the bot over 100 of posts. Most often, a person realized that it was a bot after a couple of responses. The probability of a response to the bot was 35-38%. Posts were posted every three minutes. More than 2 000 posts were posted in august 2016.
Who is interested in the project:
Partners. If you are interested in applying this chatbot to your business tasks and increasing your services' effectiveness, this is the best way.
ML enthusiasts. If you would like to gain knowledge and skills in machine learning, integrate these solutions in your project, take part in Deep2ch.
In 2016 the founders of the project tried testing this system via Slack. They collected the dataset with messages and reactions of the Slack chat and obtained statistics: who writes more often, which Slack channels they use, what responses are given to them. Then, project founders created a chatbot prototype and generated the first text using neural networks. Later, the founders came up with the idea to download data from the 2ch forum. It covered various topics and consisted of Q&A threads.
What's the plan?
The project is currently frozen but can be resumed to work on texts of different topics by the request.
How to contribute to the Deep2ch project?
If you want to offer your model of neural network training and contribute to the project or you would like to collaborate in a different way, please contact Mikhail Sveshnikov via mike0sv at gmail.com
YouTube Presentation (russian): https://youtu.be/1LcdA0Y7IEk