Overview / Usage

In recent years, there has been an increasing interest in open-ended language generation thanks to the rise of large transformer-based language models trained on millions of webpages, such as OpenAI's famous GPT2 model.

Onceupon.space is the Lego for storybooks. It allows users to build their own characters of their stories from the ground up and with the magical help of NLP, express their stories through colorful illustrations.

Onceupon.space is trained on Aesop fables, making text generation relevant and creative. Users may also use the text generation feature that allows users to get relevant story suggestions from AI - all based on their story and their characters.

Write your story today at onceupon.space

​

​

Methodology / Approach

At its heart, onceupon.space has

1. text to sketch - An NLP engine that analyses the user's sentence to understand what characters and their qualities. Qualities of characters include metrics like how many? where? what size? what color etc. The NLP engine is deployed on a scalable serverless container.
2. Text generation engine - The text generation engine is a fine-tuned GPT2 model deployed on a scalable serverless container.

Our system works with the following strategy, initially, the first step selects the best potential set of character from the given text input from which we can possibly generate the questions. The next step(second step) is to find the descriptive attributes around the subject (attributes like how many characters and the size). In the third step, we will convert the user given character to its correct grammatical singular form and check if that particular character is in the dataset. If it is present, the coordinates for the dataset are downloaded and are displayed on th canvas.

For text generation, the system sums up the story written so far and feeds it to the GPT2 model deployed on the serverless container with other attributes like output sequence length, temperature, samples to be regenerated, and size of the output samples.

We also use Intel's Devcloud for testing our data models.

Technologies Used

Intel Optimized Python Library, NLTK, Stanford Parser, Keras, Tensorflow.

This experiment was made with Intel Python. The dataset used for text to sketch was from 'Google quick Draw!' can be found here.

Other links

• Product Hunt
• Onceupon.space

Collaborators