Weakly-Supervised Fine-Grained Event Recognition on Social Media Texts for Disaster Management

Authors: Wenlin Yao, Cheng Zhang, Shiva Saravanan, Ruihong Huang, Ali Mostafavi532-539

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The evaluation on two hurricanes, Harvey and Florence, shows that using only 1-2 person-hours of human supervision, the rapidly trained weakly supervised classifiers outperform supervised classifiers trained using more than ten thousand annotated tweets created in over 50 person-hours.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Texas A&M University 2Department of Civil Engineering, Texas A&M University 3Department of Computer Science, Princeton University
Pseudocode No The paper describes algorithms and models (e.g., SLPA, Bi LSTM) in text and with diagrams, but it does not include any formal pseudocode blocks or sections labeled 'Algorithm'.
Open Source Code Yes Our system is available at https://github.com/wenlinyao/AAAI20-EventRecognitionForDisaster.
Open Datasets No The paper describes how tweets were retrieved using the GNIP API and processed to create datasets for Hurricane Harvey and Florence. However, it does not provide public access (e.g., a link, DOI, or explicit statement of public availability) to these specific collected datasets.
Dataset Splits Yes We train and evaluate a supervised classifier (multi-channel) using annotated tweets under the 10-fold cross validation setting. [...] we randomly sample a certain percentage of tweets from nine training folds as training data, ranging from 0.1 to 0.9 in increments of 0.1. [...] For training both our systems and the baseline systems, we used around 65k and 69.8k unlabeled tweets for Harvey and Florence respectively that were posted 12 hours (half a day) preceding the test time period and are therefore strictly separated from the tweets used for evaluation.
Hardware Specification Yes the bootstrapping process stopped after 9 iterations, and each iteration took around 10 minutes using a NVIDIA s Ge Force GTX 1080 GPU.
Software Dependencies No The paper mentions software components and tools like Bi LSTM encoders, pre-trained GloVe, and Adam optimizer, but does not specify their version numbers (e.g., 'TensorFlow 2.x' or 'Python 3.x').
Experiment Setup Yes For all Bi LSTM encoders, we use one hidden-layer of 300 units, pre-trained Glo Ve (Pennington, Socher, and Manning 2014) word embeddings of 300 dimensions, Adam optimizer (Kingma and Ba 2014) with a learning rate of 0.0001. [...] To deal with imbalanced distributions of event categories, we re-scale the prediction loss of each class (proportional to 1 Class Size) so that smaller classes are weighted more heavily in the final loss function. [...] we randomly cover 20% of keywords occurrences in every training epoch [...] the confidence score was initially set at 0.9 and lowered by 0.1 each time when the number of selected tweets is less than 100.