基于Bi-LSTM+Attention公共安全危机识别

Public Safety Crisis Recognition Based on Bi-LSTM+Attention

公共安全危机对社会稳定和人权构成威胁,令人担忧。社交媒体上帖子的可用性使得公共安全危机更容易被探测。然而,手动浏览和分析大量可用帖子耗时且效率低下。鉴于深度学习技术在自然语言处理方面的优势,采用深度学习技术自动识别潜在的公共安全危机成为当前的迫切需求。文中以家庭暴力危机为例,将社交媒体Facebook上有关家庭暴力的英文帖子作为研究对象,通过Facebook GraphAPI获取后进行文本预处理。采用Word2vec方法构建词向量模型,使用Bi-LSTM+self-Attention(SA-BiLSTM)深度学习模型完成了家庭暴力危机识别任务,并与CNN、RNN(recurrent neural network,循环神经网络)、LSTM三个神经网络模型进行了比较。实验结果显示,CNN和LSTM模型表现明显好于RNN,与SA-BiLSTM模型表现相接近;同时,使用self-Attention机制后Bi-LSTM模型综合表现最好,F1值、召回率、准确率均最高,其中召回率和准确率超过90%。该研究成果将为使用深度学习技术自动探测公共安全危机问题提供参考和帮助。

Public safety crisis is a cause of great concern due to the threat toward social stability and human rights.The availability of posts on social media has allowed public safety crisis to be detected more easily.However,it is time consuming and inefficient to manually browse through a massive number of available posts.Therefore,considering the advantages of deep learning technology in natural language processing,adopting deep learning as an approach for automatic identification of public safety crisis is in critical need.We consider domestic violence(DV) crisis as the example,take Facebook English posts about DV as the research object and implement text processing after getting them by GraphAPI.Then we use Word2 Vec to build word vector model and Bi-LSTM+self-Attention(SA-BiLSTM) deep learning model to accomplish DV crisis recognition task,compare it with CNN,RNN and LSTM.The experiment shows that the performances of CNN and LSTM are close to SA-BiLSTM,which is better than RNN.And the performance of SA-BiLSTM is the best with the highest F1,recall and accuracy,both of recall and accuracy values are above 90%.The research results will provide reference and help for the use of deep learning technology to automatically identify public safety crisis issues.