Category-universal witness discovery with attention mechanism in social network

Abstract

Discovering witnesses of real-world events in social networks based on posts is of great significance for crime investigation, rumor dispelling and emergency management during disasters. However, a unified modeling method for posts of all event categories is scarce. Towards this end, we propose an advanced category-universal witness discovery model (L-CA). By automatically learning the description patterns of witnesses and non-witnesses for different event categories, such as sentiment polarity, language style, etc., our model can recognize witnesses of all event categories with tweets. The model consists of three modules: feature extraction, category-enhanced tweet representation learning and multilayer perceptron classifier. The first module extracts keywords of each category and tweet textual features using TextRank and long short-term memory network respectively, while the second module aims to learn the category-enhanced tweet representation through twice attention calculations, one for category feature learning over keywords and another for tweet representation learning over the concatenation of category features and tweet textual features, and passed into the third module for classification. We construct a dataset with 3205 witnesses and 8455 non-witnesses through Twitter API, covering 14 event categories. A wealth of comparative experiments with advanced methods and model variants are performed on it. The results demonstrate that our model outperforms the state-of-the-art method on average by 9.6%, 2.3% and 2.5% in accuracy on CF_VF, HC and our dataset, respectively. Moreover, the performance of various deep neural architectures in witness discovery is analyzed in this paper.