Witness Detection In Social Network Based On Eventaugmented Attention Mechanism

In order to confirm the truth of the event and restore reality,the government and news media need to obtain direct description and feedback of the event from witnesses,so as to take corresponding measures to report the true content of the situation.With the rapid development of information technology and the widespread popularity of social networks,the traditional witness detection turns from offline to online.Witness detection research on social networks can greatly facilitate relevant departments and news media to contact witnesses and obtain reliable first-hand information.Most of the existing studies only focus on single event categories and perform highly targeted feature extraction based on artificial feature engineering.Most of them use statistical machine learning technology to carry out text classification work.This type of research requires more human work and involves strong manual intervention,which is too targeted and generalized insufficient.Based on the deep learning method and attention mechanism,this thesis designs a witness detection model that can automatically model multiple event category data to achieve the effect of capturing different language description modes under different event categories,thereby reducing human work,improving the performance of witness detection and greatly enhancing the generalization.The specific work of the thesis is as follows:(1)Aiming at the problem of automatic extraction of language description features,this thesis proposes a basic semantic feature extraction model based on recurrent neural network and keyword extraction.The model consists of two parts: tweet-oriented part focuses on sequence modeling and deep feature extraction capability,and extracts basic semantic feature sequences of tweets based on recurrent neural network;event-category-oriented part focuses on extracting the features of weakly sequential data,based on the graph-based ranking model Text Rank to extract basic semantic feature sequences of keywords of event categories.(2)Aiming at the shortcomings of single event category and insufficient generalization,this thesis designs a comprehensive semantic feature extraction model of tweets based on eventaugmented attention mechanism.The model includes two calculations of the attention mechanism: the attention mechanism is used to obtain the weights representing the importance of different keywords for the basic semantic feature sequences of the event category keywords,and then the weighted summation is used to obtain the event category feature vector;then it is embedded in the tweet basic semantic feature sequences as event-augmented embedding,and attention mechanism is used to obtain weights that represent the importance of different words in the tweet under the event category,and then the weighted summation is used to obtain the tweet comprehensive semantic feature vector;finally this thesis uses the feature vector to classify the witnesses and the non-witnesses.The combination of basic semantic feature extraction model and comprehensive semantic feature extraction model constitutes an end-toend witness detection model.The model can uniformly model data of different event categories,covering the content of the existing research and more,and achieves the efficient generalization of witness detection.(3)In this thesis,an efficient crawler software based on Twitter API is designed to obtain social network data,and a Twitter multi-category dataset is constructed in combination with public datasets,Based on this dataset,this thesis designs experiments on various comparative algorithms for existing research,deep learning algorithms and model variants of this thesis,and comprehensively evaluates them from the perspectives of evaluation index results,visualization of attention mechanisms,and preprocessing granularity.The experimental results show that both on the single-category and multi-category data,the model in this thesis reflects the stronger modeling ability,which shows the good generalization,the rationality and feasibility of deep learning methods in solving the witness detection and effectiveness of attention mechanism for processing multi-category data.(4)Based on the above results,this thesis implements a prototype system for witness detection.The system covers all the work of this thesis,and is designed for users with different requirement and identities.Through offline and online parts,it implements multiple functions such as dataset display and visualization of research results,training and fine-tuning models using input dataset,starting crawler and searching keywords to obtain data,loading pre-trained models for witness detection.