Project done during "Statistical Natural Language Processing" course.
Working team
-
Lupancu Viorica-Camelia (MLC)
-
Mocanu Ada-Astrid (MSAI)
-
Plătică Alexandru-Gabriel (MLC)
-
Știrbu Alexandru-Ilie (MSAI)
Team coordinator: Trandabăț Diana
-
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.
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.
-
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