Grammatikfehlerkorrektur mit Deep Reinforcement Learning

Raj Kumar RANA

Master's thesis

Abstract:

Reinforcement Learning (RL) method was used to fine-tune a sequence-to-label model for the Grammar Error Correction (GEC) task. Required components of RL like the environment, reward function and policy gradient algorithm were implemented. An action-search algorithm was implemented to mitigate the training instability due to the high dimension of the state and action spaces. The action-search algorithm used the world model to guide the policy with action selection. The results of our models trained with Supervised Learning (SL) and RL were compared to other GEC systems on three GEC benchmarks.

Kľúčové slová

Deep RL, Grammar Error Correction, NLP

Jazyk práce: Czech

Datum vytvoření / odevzdání či podání práce: 6. 2. 2023

Obhajoba závěrečné práce

• Vedúci: prof. Dr. Andreas Fischer

UNIVERSITY OF SOUTH BOHEMIA IN ČESKÉ BUDĚJOVICE

Faculty of Science

Master programme / odbor:
Artificial Intelligence and Data Science / Artificial Intelligence and Data Science