Thus far in this book, we have been using TimeML as an example for annotation and machine learning (ML) tasks. In this chapter, we will discuss the development of TimeML as an annotation task, and guide you through the MAMA cycle, from its first conception to its application to the TimeBank corpus, to the ISO standard that it is today. We hope that by fully working through the MAMA cycle of a task as complex as TimeML, we will be able to give you a clear understanding of some of the decisions, problems, and successes that accompany a full-scale annotation task. Much of the content of this chapter has been discussed in other papers (particularly Pustejovsky et al. 2005 and Pustejovsky et al. 2003), but this is the first time a review of mistakes that were made and problems that were discovered in the development of the model and guidelines will be discussed in detail. In this chapter we’ll go over:
The goal of TimeML
Some of the related research and theories that influenced the project
The MAMA cycle that led to the TimeML specification
The creation of TimeBank
The changes that TimeML underwent to become an ISO standard
Changes that will be applied to TimeML in the future
The ideas for TimeML stem from an ARDA workshop based on a proposal for a workshop to start thinking in terms of community standards for temporal expression and an accompanying corpus (Pustejovsky 2001).
At this point we need to acknowledge that the TimeML annotation task is clearly ...