University of Twente Student Theses
Transformers on a Diet: Semi-Supervised Generative Adversarial Networks for Reducing Transformers' Hunger for Data in Aspect-Based Sentiment Analysis
Koelewijn, Dirk (2022) Transformers on a Diet: Semi-Supervised Generative Adversarial Networks for Reducing Transformers' Hunger for Data in Aspect-Based Sentiment Analysis.
|
PDF
6MB |
| Abstract: | The vast amount of reviews and opinions being shared online for practically all available goods and services has an enormous potential value. Although the current state-of-the-art Aspect-Based Sentiment Analysis (ABSA) methods show impressive results in extracting valuable opinions, these Transformer-based models require large high-quality annotated training datasets. Datasets that are not always available and which are very costly to create. To reduce the hunger of these models for annotated data, we for the first time apply Generative Adversarial Networks (GANs) to ABSA. We investigate using both regular and Wasserstein semi-supervised GANs to generate artificial word embeddings, with varying amounts of unlabelled data and varying generator complexity. We show that adding such a GAN can significantly improve performance, even without using unlabelled data. Furthermore, we identify how much unlabelled data works best and show that generators with more hidden layers perform better. Altogether, we show that our method allows for reducing annotated data by 50\% while still achieving similar performance. |
| Item Type: | Essay (Master) |
| Clients: | Accenture Netherlands, Amsterdam, Netherlands |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Computer Science MSc (60300) |
| Link to this item: | https://purl.utwente.nl/essays/93794 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page