This repository contains the temporal dependency graph dataset as described in this paper:
Annotating Temporal Dependency Graphs via Crowdsourcing
Jiarui Yao, Haoling Qiu, Bonan Min and Nianwen Xue
To appear in EMNLP 2020.

The temporal dependency graph dataset is in tdg_data/. The train.txt contains 400 articles, dev.txt and test.txt both contain 50 articles.

Each article has three parts: file name, sentence list and edge list, seperated by :SNT_LIST and EDGE_LIST.

• The edge list has three parts: edges between timex and timex, edges between event and timex and edges between events and events. For example, in dev.txt, line 21-33 are the timex timex edges, line 34-56 are the event timex edges, and line 57-79 are the event event edges.

• Each edge has 4 fields, separated by tabs:

child    child_type    parent  relation

• child: the child node
• child_type: the child is a timex node or an event node
• parent: the parent node of this child
• relation: the temporal relation between the child and the parent

The child node and parent node are represented by their sentence id, start token id in the sentence and end token id in the sentence. For example, 0_0_3 means this node is in the first sentence of this document, its start token position in this sentence is 0, and its end token position in this sentence is 3. We use -1_-1_-1 to represent the ROOT node.

Please note: some events do not have an event reference time, we add -1_-1_-1 as a mate event reference time.

You can find our annotation guidelines in guidelines_and_example_questions/, which are essentially our Amazon Mechanical Turk annotation interface. Some example questions are also included.

The generic_event.whitelist.wn-fn.variants file contains a list of common event trigger words. We use this list to extract possible events in the 3rd pass annotation.