Abstract
The thesis proposes several deep sequence models to solve sequence prediction problems in real-world contexts. Our approach shows promising results that outperform the state-of-the-art in investigated domains. Furthermore, we release several large-scale domain-specific corpora and make them open-source for the respective communities. In itinerary recommendation, existing methods focus on the single-person perspective but do not address real-world problems resulting from natural crowd behavior, such as the Selfish Routing problem. We propose the Strategic and Crowd-Aware Itinerary Recommendation (SCAIR) algorithm, which formulates the strategy of route recommendation as Markov chains and enables our simulations to be carried out in poly-time and outperform realistic baselines using a real-world dataset. In career modeling, companies are exploring artificial intelligence to improve human resource management. We present a corpus that contains over 475,073 job titles based on 192,295 user profiles from a major professional networking site. Furthermore, we propose a contextual job title vector representation that improves deep language models to outperform humans and baselines in our experiments. In crisis signal detection, classifying crisis events is crucial to build awareness and inform relevant parties for proactive actions. We propose CrisisBERT, an end-toend transformer-based model for crisis detection and crisis recognition, which shows promising results across accuracy and F1 scores. In the recent COVID-19 pandemic, existing disease-related corpora are too small and noisy to support large-scale analysis. We present EPIC30M, a large-scale epidemic corpus with more than 30 million micro-blog posts containing three general diseases and six global epidemic outbreaks.