Implementing char-RNN from Scratch in PyTorch, and Generating Fake Book Titles

This week, I implemented a character-level recurrent neural network (or char-rnn for short) in PyTorch, and used it to generate fake book titles. The code, training data, and pre-trained models can be found on my GitHub repo here.

 
Heart in the Dark
Me the Bean
Be the Life
Yours
 

Model Overview

 
Diagram of the char-rnn network architecture.  Source .

Diagram of the char-rnn network architecture. Source.

 

The char-rnn language model is a recurrent neural network that makes predictions on the character level. In contrast, many language models operate on the word level.

Making character-level predictions can be a bit more chaotic, but might be better for making up fake words (e.g. Harry Potter spells, band names, fake slang, fake cities, fantasy terms, etc.). Word-level language models might have an advantage for generating longer pieces of text, like summaries or fiction, as they don’t need to figure out how to spell, in a sense.

There do exist character-word hybrid approaches. For example, the GPT-2 model uses byte pair encoding, an approach that interpolates between the word-level for common sequences and the character-level for rare sequences.

This particular char-rnn implementation is set up to handle multiple categories of text. In this use case, it is able to make predictions for different book genres, e.g. Romance, Fantasy, Young Adult, etc.

Training Data

The training data used for this model is a modified version of this Goodreads data scrape of 20K book titles. I transformed the CSV file into separate text files for the top 30 genres. The resulting split dataset can be found in my repo here.

GPU training time with this model took about 20 minutes on an NVIDIA GeForce GTX 1080 Ti. Generating samples only takes a few seconds.

Results

The following results are a selected sampling of outputs. Note that I’m mainly including examples that consist of real words, with a few exceptions.

Romance

Heart in the Dark
Years of the Dark
You the Book
The Stove to the Story

Fantasy

Growing the Dark
Book of the Dark
Red Sande

Fiction

In the Bead Store
Jen the Bead
King the Bean

Historical

A to the Bean
Other and Story

Science Fiction

Darke Sers
Voringe
In the Beantire

Mystery

Bed Singe
Kiss of the Dark
Red Story

Classics

A Mander of the Suckers
Gorden the Story of Merica

Childrens

Dark Book of the Story of the Sures of the Surating
Late
Story of the Bean

Paranormal

A Store of the Store
Red Store
Stariss and Storiss
Wind Store

New Adult

Live Me Life
Growing Me
In the Bean
Me the Bean

Poetry

Yours
Me

Erotica

Volle the Story of Men
King of the Dark
Dork of the Dark
Work of the Dark
Bed Storys of the Dark
Your Mind

Biography

Be the Life
On Anger and Of Mand Anger

Comically, there are many book titles that revolve around beans, beads, stores, and darkness. While I did notice some subtle differences between genres, it doesn’t appear to be particularly drastic overall.