Multimodal Amharic Hate Speech Detection Using Deep Learning

Abstract

Social media has become a tool of online communication used for discussions, information sharing, and the creation of online content on social networks. Hate speech is one of the detrimental or harmful information spread on social media. Discrimination against individuals results from hate speech. Human rights are also violated by it. As a result, several studies have been attempting to identify hate speech. However, while the majority of studies focus on identifying hate speech in textual data, it is also distributed through films. Amharic is also one of the under-resourced languages that benefits from hate speech detection, so this is another positive. This work seeks to apply a multi-modal, which is a mix of the acoustic and textual elements using a deep learning technique, to address the issue of Amharic hate speech on social media In the beginning, we gathered 1,459 extracted audios from YouTube videos, ranging in length from 30 seconds to 140 minutes. A five-minute audio speech employing silent segmentation techniques follows this. From these, we have a collection of 6497 audios from one to five minutes after the audio segmenting and filtering. Each Audio is annotated by a domain expert for the purpose of performing tests. Then, for our research, we use the Google Speech-to-Text API to transcribe audio speech into text scripts. The features were then extracted, with textual features extracted using word2vec and acoustic features extracted using Mel-Frequency Cepstral Coefficient (MFCC). As a result, this study makes use of four deep learning algorithms: LSTM, BILSTM, GRU, and BIGRU. The results of the multi-modal experiment demonstrate that the multi-modal model with BILSTM outperforms the other experiment for detecting Amharic hate speech with an accuracy of 88.15%. Furthermore, we are working to extend the Amharic hate speech detection taking the video in to account addition to the text and audio.

Topics & Concepts