Social media is increasingly used by people during large-scale natural disasters to request emergency help. Previous work has had success in applying machine-learning classifiers to detect tweets in coarse-grained categories, such as disaster type and relevance. However, there is a dearth of work that focuses on detecting tweets containing requests for help that are actionable by first responders. Using over 5 million tweets posted during 2017's Hurricane Harvey in Houston, U.S., we show that though such requests are uncommon, their often life-or-death nature justifies the development of tweet classifiers to detect them. We find that the best-performing classifiers are a convolutional neural network (CNN) trained on word embeddings, support vector machine (SVM) trained on average word embeddings, and multilayer perceptron (MLP) trained on a combination of unigrams and part-of-speech (POS) tags. These models achieve F1 scores of over 0.86, confirming their efficacy in detecting urgent tweets. We highlight the utility of average word embeddings for training non-neural models, and that such features produce results competitive with more traditional n-gram and POS features.

人们在大规模自然灾害期间越来越多地使用社交媒体来请求紧急帮助。先前的工作已经成功应用了机器学习分类器来检测粗粒度类别（例如灾难类型和相关性）中的推文。但是，缺乏工作重点在于检测包含第一响应者可以采取行动的帮助请求的推文。使用2017年在美国休斯敦的哈维飓风期间发布的超过500万条推文，我们表明，尽管此类请求并不常见，但它们的生死攸关的性质证明了开发推文分类器进行检测的合理性。我们发现效果最好的分类器是经过词嵌入训练的卷积神经网络（CNN），经平均词嵌入训练的支持向量机（SVM），多层感知器（MLP）接受了字母组合词和词性（POS）标签的组合训练。这些模型的F1分数超过0.86，证实了它们在检测紧急推文中的功效。我们强调了平均单词嵌入在训练非神经模型中的效用，并且这种特征产生的结果与传统的n-gram和POS特征相比具有竞争力。