Exploring Word Representations on Time Expression Recognition

Time expression extraction has attracted longstanding interest over time, due to its great importance in many downstream tasks of Natural Language Processing (NLP) and Information Retrieval (IR). Although current approaches, either rule-based or learning-based, can achieve impressive performance in major datasets, they usually rely heavily on handcrafted rules or task-specific pre-tagging features. Recent advances in pretrained word representations motivate us to explore semisupervised approaches for this task. We first show that simple neural architectures built on top of pre-trained word representations perform competitively and efficiently on time expression recognition. Then we further propose a neural model of jointly learning to identify and type time expression.