Personal-Bullying Detection Based on Multi-Attention and Cognitive Feature

摘要

The rapid growth of social media in recent years has fed into some anti-social behavior such as kinds of cyberbullying. Previous researches only apply a single network model to complete detection. In this paper, aim to personal-bullying of Chinese social media, we propose a novel network framework with Multi Interactive-Attention and Language-environment Cognitive (MIALC) for personal-bullying detection: (1) we apply three attention features to capture multi-level and deep semantic information without using any external parsing result. Among them, the stroke attention feature can mine internal structural information of Chinese word. Meanwhile, (2) the ParagraphVector aims at extracting language-environment cognitive information from social media text, since the language-environment factors have restrictive effects on the expression of personal-bullying. The experimental results show that our proposed MIALC framework is effective.