Full text
| |
| Author(s): |
Gustavo Henrique de Rosa
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Bauru. 2022-12-06. |
| Institution: | Universidade Estadual Paulista (Unesp). Faculdade de Ciências. Bauru |
| Defense date: | 2022-11-22 |
| Advisor: | João Paulo Papa |
| Abstract | |
Machine Learning algorithms have been paramount in the latest years, mainly due to their discriminative capacity in Computer Vision and Natural Language Processing tasks. Furthermore, their generative potentials allowed usage in discrete-based (sequences of characters and words) tasks, such as text generation. A specific architecture denoted as Generative Adversarial Networks uses an structure composed of discriminators and generators to establish an equilibrium between artificial data generation and their classification as real data. Several works proposed adversarial-based models to generate text; however, only a few could generate non-repeated text with little semantic significance. Furthermore, a recurring issue regarding Generative Adversarial Networks consists of the difficulty of establishing a training equilibrium and, consequently, generating artificial text that resembles the original ones. Therefore, this thesis enhances the development of text-based adversarial models through similarity functions learned from Siamese Networks, which provides rewards capable of better distinguishing between artificial and original texts. Such models are improved through meta-heuristic optimization, which furnishes specific hyperparameters to the accounted tasks. The experimental results indicate the capacity of the proposed architecture, denoted by Text-Similarity Generative Adversarial Network (TS-GAN), amongst four literature datasets. The TS-GANs obtained state-of-the-art results and, in their post-optimization versions, were able to improve their standard versions (without optimization) in two out of four datasets. (AU) | |
| FAPESP's process: | 19/02205-5 - Adversarial learning in natural language processing |
| Grantee: | Gustavo Henrique de Rosa |
| Support Opportunities: | Scholarships in Brazil - Doctorate |