SPOCK: Subjective Projection for Optimizing Deep Neural Classification Models Based on Contextual Representation From Large Knowledgebase
Violence has long been studied in the social sciences, particularly in peace studies. Violence has become more common as a result of the recent advancement of social media, and its form has shifted toward inciting violence and other passive attacks. As the rate of violence related to social media continues to rise, so does the need for understanding and computationally detecting various types of violence. Over the last few years, Natural Language Processing (NLP) models aimed at detecting hate and offensive speech more specifically related to religion, race, gender, or sexual orientation. However, these NLP models are mostly domain specific and tend to underperform when we use violent and peace data sets due to a lack of vocabulary and domain-specific knowledge. In this study, we propose a novel network on top of BERT embedding to understand the domain-specific knowledge context. We have targeted the violent and peace data set and optimized the embedding by implementing custom loss functions for better space projection of violent and peaceful words. We tried understanding how violent and peace words can be projected into a new dimension using BERT base word embedding. For future work, we will train this model with over 10k annotated and fine-tune the pre-existing BERT model. Besides this, improving the model's interpretability and accuracy will be the major focus of any future work.
Computer science|Web Studies|Information science|Criminology
Machlovi, Naseem, "SPOCK: Subjective Projection for Optimizing Deep Neural Classification Models Based on Contextual Representation From Large Knowledgebase" (2023). ETD Collection for Fordham University. AAI30522009.