Media Bias Detecting based on Word Embedding

Abstract

Machine learning research to detect political bias in articles has boomed in recent years. However, there is still no widely accepted and effective word embedding technique for detecting bias. This paper explores the connection between political bias and word embedding models and deduces factors to consider when selecting and developing word embedding techniques. Three classic word embedding models are introduced into experiments to conduct comparisons to achieve this goal. Contextual meaning is observed to lose efficiency in the task. In contrast, frequency is the most relevant feature in predicting media bias. Simultaneously, this paper discovers a unique accuracy distribution generated by Random Forest through experiments. Experiments reveal that it has apparent advantages in accuracy when predicting left-biased articles, which may relate to features undiscovered.