Poetic Taybot

Computational Creativity in Poem Generation

By Trent Eriksen, Christina Irakleous, Nikki Rademaker, and Yanna Smid

Overview

Poetry is an art form that bridges emotion and creativity, but creating it can be challenging. Inspired by Taylor Swift's poetic songwriting, we developed "Poetic Taybot", an artificial poetry generator that uses Taylor Swift's lyrics as its muse.

This project was created for the course Computational Creativity. We combined linguistic algorithms with generative AI to create, present, and perform original poems. The goal was to build a poetry generator using a Markov Chain trained on Taylor Swift's lyrics and visualize the generated poems using AI-generated voice, music, and art.

• Poetry Generation: Taylor Swift's complete discography, including From the Vault tracks, was cleaned and formatted as training data. We implemented a Markov Chain to generate poetic verses, focusing on stylistic elements like alliteration and similes.
• Creative Visualization: Selected poems were narrated through an AI voice synthesizer mimicking Taylor Swift's vocal tone. A MIDI composition inspired by beat poetry and Taylor's themes was created with AI. AI art visualized the mood and tone of each poem.

A few examples of generated poems scan be seen below:

Process

The following steps were completed to create this project:

• Concept Development: We drew inspiration from Taylor Swift's lyrical style and quickly decided to base our generator on her lyrics.
• Data Collection: We collected all lyrics from all existing Taylor Swift albums from a Kaggle dataset (nine albums + three rerecorded albums) and converted them into txt and csv files.
• Data Preprocessing: We cleaned the raw texts to remove noise and unwanted characters.
• Algorithm Implementation: We built a Markov Chain in Python to generate short poems of five lines with no more than 60 characters each, including alliterations.
• Generative Experiments: We tested the algorithm and tuned hyperparameters such as poem length and word limits per line.
• Integration: We selected three poems to visualize, combining audio, visual, and text into a multimedia presentation.

Technical Development

• Data Cleaning: Text was cleaned by converting to lowercase, removing special characters, and standardizing contractions and grammar.
• Algorithm: A Markov Chain algorithm paired words based on likelihood, generating sentences word by word. We added fallback mechanisms and emphasized alliteration.
• Generative Tools: ElevenLabs for voice synthesis, Magenta for MIDI generation processed in Ableton Live 11. AI art generators created visuals for the poems.

Final Poem Presentation

A few selected poems are presented in a video combined with AI-generated voice, AI music, and AI visuals to deliver an immersive artistic experience.