2020/03/04 by Jitin Krishnan, Krishnan, Jitin, Hemant J. Purohit +3
Computer Science · Social Sciences · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Public Relations and Crisis Communication #Seismology and Earthquake Studies #Social and Information Networks (cs.SI) #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2003.04991
openalex publication_date 2020/03/04 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
During the onset of a disaster event, filtering relevant information from the\nsocial web data is challenging due to its sparse availability and practical\nlimitations in labeling datasets of an ongoing crisis. In this paper, we\nhypothesize that unsupervised domain adaptation through multi-task learning can\nbe a useful framework to leverage data from past crisis events for training\nefficient information filtering models during the sudden onset of a new crisis.\nWe present a novel method to classify relevant tweets during an ongoing crisis\nwithout seeing any new examples, using the publicly available dataset of TREC\nincident streams. Specifically, we construct a customized multi-task\narchitecture with a multi-domain discriminator for crisis analytics: multi-task\ndomain adversarial attention network. This model consists of dedicated\nattention layers for each task to provide model interpretability; critical for\nreal-word applications. As deep networks struggle with sparse datasets, we show\nthat this can be improved by sharing a base layer for multi-task learning and\ndomain adversarial training. Evaluation of domain adaptation for crisis events\nis performed by choosing a target event as the test set and training on the\nrest. Our results show that the multi-task model outperformed its single task\ncounterpart. For the qualitative evaluation of interpretability, we show that\nthe attention layer can be used as a guide to explain the model predictions and\nempower emergency services for exploring accountability of the model, by\nshowcasing the words in a tweet that are deemed important in the classification\nprocess. Finally, we show a practical implication of our work by providing a\nuse-case for the COVID-19 pandemic.\n