As you can see from the preceding chapter, modeling events, times, and their temporal relationships in an annotation is a large and complicated task. In this chapter we will discuss the TARSQI Toolkit, as well as other systems that were created to generate TimeML as part of the TempEval-2 challenge held in 2010. In this chapter, we will:
Discuss how a complicated annotation can be broken down into different components for easier processing
Provide an in-depth discussion of the first attempt to create a system for creating TimeML
Show examples of how that system has been improved over the years
Explain the approaches taken by other examples of systems designed to create TimeML
Discuss the differences between rule-based and machine learning (ML) systems for complex annotation tasks
Provide examples of ways that the TARSQI Toolkit could be expanded in the future
Overall, in this chapter we won’t be going into detail about how each aspect of TimeML was automated; rather, we will provide a breakdown of how the task was approached, and give a sense of some of the different options available for tackling a complicated annotation.
The TARSQI Toolkit is not the creation of a single person, and we would like to acknowledge all of the people who have contributed to its creation and improvement (in alphabetical order): Alex Baron, Russell Entrikin, Catherine Havasi, Jerry Hobbs, Seo-Hyun Im, Seok Bae Jang, Bob Knippen, Inderjeet Mani, Jessica ...