Texty is a Slack integration of a deep neural conversation model inspired by /giphy .

In the backend Texty is based on a hierarchical encoder-decoder recurrent neural network trained on a large dataset of movies dialogue (http://arxiv.org/pdf/1507.04808v2.pdf). Without any prior knowledge of language Texty can learn to generate sentences one word at a time conditioned on the context of the conversation. These models generally have issues with generating interesting utterances and often just give you generic things like "I don't know", "yes", "I'm sorry". We modified the model sampling to generate more interesting/zany responses (inspired by recent research http://arxiv.org/abs/1510.03055).

Generating more interesting responses can lead to degradation in the language model but the Slack integration allows to get online feedback using the reactions to help the bot improved and generate more amusing responses.

Share this project: