Türkçe His Analizinde Optimizasyon ve Öneğitimli Modellerin Kullanımı

Abstract

Emotion analysis means to automatic determination of the emotions in any digital content. It has become a necessity to distill and evaluate with emotion analysis the big data that emerged as a result of the dizzying technological developments in the internet and digital communication and the social changes caused by these. However, a large amount of labeled data with proportional class distribution is needed to perform emotion analysis with machine learning methods. On the other hand, expressing feelings only in words without gestures adds a great uncertainty to the problem.

Herein, it is aimed to perform high performance Turkish emotion analysis with machine learning techniques. For this purpose, the performance of emotion analysis is increased by proposing solutions to unbalanced class distribution which increases uncertainty, and data scarcity. In the study, inner class performances are used to measure the success of all classes in datasets showing imbalanced distribution. While maintaining the general accuracy performances achieved by the innovation brought to the deep learning decision mechanism with the optimum emotion vectors method developed to reduce uncertainty, the inner class emotion analysis performances were increased.

In addition, for the first time in Turkish, pre-trained language models were adapted to emotion analysis, and a pre-trained emotion model, which does not require large amounts of labeled data, was developed. The results obtained from the developed method are presented comparatively with different data sets and learning methods. As a result of the experiments performed on the Turkish emotion data sets with the pre-trained emotion model, the highest performance in the field was obtained.

The proposed study will enable high-level artificial intelligence tasks such as more accurate public opinion polling, customer relationship management, brand management, detection of cyberbullying, determination of the tendency in elections, determining partisan comments in Turkish resources. It will also be a resource for researchers in this field, and can be used in Turkish text mining tasks.