Time expression recognition and normalization: a survey

Time information plays an important role in the areas of data mining, information retrieval, and natural language processing. Among the linguistic tasks related to time expressions, time expression recognition and normalization (TERN) is fundamental for other downstream tasks. Researchers from these areas have devoted considerable effort in the last two decades to define the problem of time expression analysis, design the standards for time expression annotation, build annotated corpora for time expressions, and develop methods to identify time expressions from free text. While there are some surveys concerned with the development of time information extraction, retrieval, and reasoning, to the best of our knowledge, there is no survey focusing on the TERN development. We fill in this blank. In this survey, we review previous researches, aiming to draw an overview of the development of time expression analysis and discuss the role that time expressions play in different areas. We focus on the task of recognizing and normalizing time expressions from free text and investigate three kinds of methods that researchers develop for TERN, namely rule-based methods, traditional machine-learning methods, and deep-learning methods. We will also discuss some factors about TERN development, including TIMEX type factor, language factor, and domain and textual factors. After that, we list some useful datasets and softwares for both tasks of TER and TEN as well as TERN and finally outline some potential directions of future research. We hope that this survey can help those researchers who are interested in TERN quickly gain a comprehensive understanding of the development of TERN and its potential research directions.