TrajTrace: Tracing Moving Objects over Social Media

Abstract

Online social media has lots of moving object information. Extracting movement information is a challenging work. It suffers from spatio-temporal information extraction, vague time, and vague location. Previous information extraction methods merely focus on the trajectory extraction. In this demonstration, we develop a web-based application, TrajTrace, to track moving objects. TrajTrace extracts triple <object, movementState, location> from social media text employing our proposed span-level joint entity and relation extraction model, OMLer. OMLer casts joint extraction as a token pair multi-categories classification task. It predicts the triple list corresponding to the input sequence. We employ BERT to encode the input sentence word by word. The self-attention mechanism and BiLSTM are applied to learn sequence features. Then, an order-first time matching algorithm is designed to solve the lacking temporal information problem in the extracted triples. Utilizing the proposed TF-IDF based clustering algorithm, we make the vague time accurate. The vague geographic location is converted to accurate latitude and longitude using the Bezier geodetic coordinate conversion algorithm. Toward aircraft and ships, besides the keyword search and the trajectories visualization, TrajTrace provides the historical activity area search of a specific object and the spatio-temporal distribution of moving objects at a given time or location.