Deep and Contextually Engineered Features for Metaphor Detection

The primary objective of this research is to develop an automated system capable of identifying metaphorical expressions in brief texts. This entails investigating the most effective features for this task, achieved through a combination of deep learning architecture and meticulously crafted contextual features. These methodologies will be thoroughly examined in this paper. Initial observations revealed that certain feature sets demonstrated strong individual performance, while others were less effective. However, through integration with the former, even the initially weaker sets proved highly beneficial. Subsequently, these combined feature sets were subjected to classification using various established machine learning algorithms. To ensure a comprehensive evaluation, all five algorithms were employed for comparison purposes. Notably, Support Vector Machine (SVM) emerged as the optimal algorithm for this task. Overall, the experimental outcomes exhibited favorable results across all evaluated metrics. Additionally, a comparative analysis of the results, particularly in terms of F1-measure, against existing works within the same domain is presented in this paper.