Haber İçeriklerinden Haber Kaynaklarının Politik Görüşlerinin Tespit Edilmesi
Fen Bilimleri
K22038
Bilimsel Etkinlik Katılım DESTEK Projesi
Dr.Öğr.Üyesi Serdar ÇİFTÇİ
Mühendislik Fakültesi
2022
2022
10-05-2022
09-11-2022
The news companies can shape public opinions with their services. Those services sometimes can be inclined to the company's ideology. Thomas and Kovashka [1] proposed a deep-learning-based model analyzing news images and related texts to determine whether the news comes from right- or left-leaning media sources. They also presented a dataset [2] that they used in their studies. For the SICSS summer school project, I would like to develop a novel deep-learning-based model that combines news images and texts. I will employ an auto-encoder-based architecture for extracting visual features and various embedding methods for texts.
Those multi-modal data will be fused with an attention mechanism that assures a weighted fusion. With the attention mechanism, the deep-learning-based model result would be interpretable, finding the reasons for human biased decisions. For the implementation, I will use the same dataset [2]. All experiments will be conducted in Python programming language using the PyTorch framework. The performance of the proposed model will be evaluated with ablation studies and compared with the state-of-the-art models. The project's outcome could be published as an article or presented at a respected international conference.
1. Thomas, Christopher, and Adriana Kovashka. "Predicting the politics of an image using webly supervised data." Advances in Neural Information Processing Systems 32 (2019).
2. http://www.cs.pitt.edu/~chris/politics
NOT: Katılım Tipi'nde sözlü sunum olmadığından sözlü bildiri seçildi. Proje fikri çalıştayda sunulacak ve çalıştay süresince projenin gerçekleştirilmesi için çalışılacak.