Demo

Type something and hit enter! In addition, this model allow you to pick the character you want to talk to.

--> no connection to the server :(

say: to



About

Human Computer Interaction Lab @ EPFL
Dongqi's Home Page

Introduction

This is a tiny LSTM(P) response generator trained with only 35,429 pair of converstaion from the Big Bang Theory. The model is small enough to run beam-inference on a laptop CPU. This demo also demonstrate the use of signal token in Seq2seq model.

More details can be found in the techniqal report: Read techniqal report
Github

Model

name: seq2seq LSTMP
cell_hidden_size: 300
cell_projection_size: 512
num_layer: x1
embedding_size: 20525 x 300

Training

epoch: 18
time: less than 10 minutes on a GTX 1080
per-word cross entropy: 3.72 --> 0.82
optimizer: Adam (learning rate = 0.0005, other parameters use TensorFlow default)

Reading

Data-Driven Response Generation in Social Media
Sequence to Sequence Learning with Neural Networks
A Neural Conversational Model
A Diversity-Promoting Objective Function for Neural Conversation Models