In this article we present GeoTxt, a scalable geoparsing system for the recognition and geolocation of place names in unstructured text. GeoTxt offers six named entity recognition (NER) algorithms for place name recognition, and utilizes an enterprise search engine for the indexing, ranking, and retrieval of toponyms, enabling scalable geoparsing for streaming text. GeoTxt offers a flexible application programming interface (API), allowing for customized attribute and/or spatial ranking of retrieved toponyms. We evaluate the system on a corpus of manually geo-annotated tweets.First, we benchmark the performance of the six NERs that GeoTxt provides access to. Second, we assess GeoTxt toponym resolution accuracy incrementally, demonstrating improvements in toponym resolution achieved (or not achieved) by adding specific heuristics and disambiguation methods. Compared to using the GeoNames web service, GeoTxt's toponym resolution demonstrates a 20% accuracy gain. Our results show that places mentioned in the same tweet do not tend to be geographically proximate.

No abstract

InterfaceAutomated Label Verify Label Fig. 1. Our interactive learning framework allows users to train text relevance classifiers in real-time to improve situational awareness.In this example, a real-time tweet regarding a car accident is incorrectly classified as "Irrelevant". Through the SMART 2.0 interface, the user can view its label and correct it to "Relevant", thereby retraining and improving the classifier for incoming streaming data.Abstract-Various domain users are increasingly leveraging real-time social media data to gain rapid situational awareness. However, due to the high noise in the deluge of data, effectively determining semantically relevant information can be difficult, further complicated by the changing definition of relevancy by each end user for different events. The majority of existing methods for short text relevance classification fail to incorporate users' knowledge into the classification process. Existing methods that incorporate interactive user feedback focus on historical datasets. Therefore, classifiers cannot be interactively retrained for specific events or user-dependent needs in real-time. This limits real-time situational awareness, as streaming data that is incorrectly classified cannot be corrected immediately, permitting the possibility for important incoming data to be incorrectly classified as well. We present a novel interactive learning framework to improve the classification process in which the user iteratively corrects the relevancy of tweets in real-time to train the classification model on-the-fly for immediate predictive improvements. We computationally evaluate our classification model adapted to learn at interactive rates. Our results show that our approach outperforms state-of-the-art machine learning models. In addition, we integrate our framework with the extended Social Media Analytics and Reporting Toolkit (SMART) 2.0 system, allowing the use of our interactive learning framework within a visual analytics system tailored for real-time situational awareness. To demonstrate our framework's effectiveness, we provide domain expert feedback from first responders who used the extended SMART 2.0 system.Index Terms-Interactive machine learning, human-computer interaction, social media analytics, emergency/disaster management, situational awareness

Measurements of human interaction through proxies such as social connectedness or movement patterns have proved useful for predictive modeling of COVID-19, which is a challenging task, especially at high spatial resolutions. In this study, we develop a Spatiotemporal autoregressive model to predict county-level new cases of COVID-19 in the coterminous US using spatiotemporal lags of infection rates, human interactions, human mobility, and socioeconomic composition of counties as predictive features. We capture human interactions through 1) Facebook- and 2) cell phone-derived measures of connectivity and human mobility, and use them in two separate models for predicting county-level new cases of COVID-19. We evaluate the model on 14 forecast dates between 2020/10/25 and 2021/01/24 over one- to four-week prediction horizons. Comparing our predictions with a Baseline model developed by the COVID-19 Forecast Hub indicates an average 6.46% improvement in prediction Mean Absolute Errors (MAE) over the two-week prediction horizon up to 20.22% improvement in the four-week prediction horizon, pointing to the strong predictive power of our model in the longer prediction horizons.