Spaces:
Sleeping
Sleeping
Prototypes
- Text-generation with rhyme and rhythm
- Ans Farooq
- 2390370f
- Jake Lever
Intro
- This file contains information and reflections about each prototype of the project.
29/10/2021 - First Prototype - L4_Project_first.ipynb
Current state/Improvements made
The first prototype combines a causal language model and a masked language model, GPT-2 and RoBERTa, to generate the starts of sentences and fill in the rest until the end rhyming word.
For this prototype, the starting word and end rhyming word of each line was pre-determined and hard-coded. This was temporary as I was focused on getting GPT-2 and RoBERTa working and generating some coherent lines of text.
Future improvements
- Use a Python library to generate rhyming words
- Use user input for the topic of the limerick
- Feed summary of topic from wikipedia to GPT-2 before it generates the start of each line
16/11/2021 - Second Prototype - L4_Project_second.ipynb
Current state
- The second prototype uses GPT-2 and RoBERTA, but instead of hard-coded starting and end rhyming words, it uses the pronouncing Python library to find rhyming words and the wikipedia library to feed a summary of the topic to GPT-2 before generation.
Future improvements
- Word frequencies/counts for filtering rhyming library words
- Generate the first line a few times until a decent end word, e.g noun, hand-built filters, or word list
- Have a look into improving the rhyme finding, better libraries? Phonetics? Could just filter out words with multiple pronunciations.