GenROPO: Generating Romanian poems

Project done during "Statistical Natural Language Processing" course.

Working team

1. Lupancu Viorica-Camelia (MLC)

2. Mocanu Ada-Astrid (MSAI)

3. Plătică Alexandru-Gabriel (MLC)

4. Știrbu Alexandru-Ilie (MSAI)

Team coordinator: Trandabăț Diana

• Project presentation

• Project description

State-of-the-art

• https://arxiv.org/abs/1909.09534 - Creative GANs for generating poems, lyrics, and metaphors - a GAN approach for generating poems in English

• https://aclanthology.org/P17-1016.pdf - Automatically Generating Rhythmic Verse with Neural Networks - a softer method based on a LSTM (English) and applied poetic directions

• https://arxiv.org/pdf/2102.04114.pdf - GENERATE AND REVISE: REINFORCEMENT LEARNING IN NEURAL POETRY - a method based on a bi-LSTM, including several experiments and postprocessing techniques (English)

• https://aclanthology.org/D14-1074.pdf - Chinese Poetry Generation with Recurrent Neural Networks - poem generator experiments on convolutional and recurrent networks, tested using BLEU score and human volunteers.

Conclusions

The common point of all these papers consists in Neural Network type approaches, more precisely a Deep Learning one.

We will try to use a GAN architecture in which the Generator is a Transformer and the Discriminator is a simple network.

We consider each word a token meant to be generated by the Generator along with other tokens in order to create a strophe/stanza.

Our goal is to generate plausible poems in Romanian language.

Work per module

• Collecting data Module: Lupancu Camelia

• Encoding data Module: Platica Alexandru

• Processing data (Tools and Generator): Mocanu Astrid

• Processing data (Discriminator and GAN): Stirbu Alexandru

• Postprocessing & Testing: Stirbu Alexndru & Mocanu Astrid

• Documentation & Presentation: Mocanu Astrid & Lupancu Camelia

Arhitecture