An Annotation Framework for Dense Event Ordering


Today’s event ordering research is heavily dependent on the annotated corpora available to it. The corpora influence shared evaluations and drive algorithm development. Partly due to this dependence, most research focuses on partial orderings of a document’s events. For instance, the TempEval competitions and the TimeBank only annotate small portions of the event graph. They focus on the most salient events or on specific types of event pairs (e.g., only events in the same sentence). Deeper temporal reasoners struggle with this sparsity because the entire temporal picture is not represented. This paper proposes a new annotation process with a unique mechanism to force annotators to label connected graphs. It generates 10 times more relations per document than the TimeBank and the corpus is larger than all current corpora. We hope this new annotation framework will encourage research on new models, particularly global models with deep reasoning.



1 Introduction

The TimeBank Corpus [6] helped usher in a wave of event ordering research with data-driven algorithms. It provided for a common dataset of annotations between events and time expressions that allowed the community to compare approaches. Since its creation, other corpora and several competitions have based their tasks on the TimeBank setup. This paper addresses one of the shortcomings of the TimeBank: sparse annotation. We describe a new annotation framework (and corpus) that we believe is required to fulfill the data needs of deeper temporal reasoners.

The TimeBank includes a small subset of all possible relations in its documents. The original annotators were instructed to label relations critical to the document’s understanding. The result is a sparse labeling that leaves much of the document unlabeled. The TempEval contests have largely followed suit and focused on specific types of event pairs. For instance, TempEval [8] only labeled relations between events that syntactically dominated each other. This paper is the first attempt to annotate a document’s entire temporal graph.

A consequence of focusing on all relations is a shift from the traditional temporal relation classification task, where the system is given a pair of events and asked only to label the type of temporal relation, to a temporal relation identification task, where the system must determine for itself which events in the document to pair up. For example, in TempEval-1 and 2 [8, 9], systems were given event pairs in specific syntactic constructions: events and times in the same noun phrase, main events in consecutive sentences, etc. In contrast, we aim for a shift in the community wherein all pairs are considered candidates for temporal ordering, allowing researchers to ask questions such as: how must algorithms adapt to label the complete graph of pairs, and if the more difficult and ambiguous event pairs are included, how must feature-based learners change?

We are not the first to propose these questions, but this paper is the first to directly propose the means by which they can be addressed. The stated goal of TempEval-3 [7] was to focus on relation identification instead of classification, but the training and evaluation data followed the TimeBank approach where only a subset of event pairs were labeled. As a result, many systems focused on classification, with the top system classifying pairs in only three syntactic constructions [2]. We describe the first annotation framework that forces annotators to annotate all pairs11As discussed below, all pairs in a given window size.. With this new process, we created a dense ordering of document events that can properly evaluate both relation identification and relation annotation. \Creffig:dense illustrates one document before and after our new annotations.

Figure 1: A TimeBank annotated document is on the left, and this paper’s annotation is on the right. Solid arrows indicate a before relation and dotted arrows indicate an included_in relation.

2 Previous Annotation Work

The majority of corpora and competitions for event ordering contain sparse annotations. Annotators for the original TimeBank [6] only annotated relations judged to be salient by the annotator. Subsequent TempEval competitions [8, 9, 7] mostly relied on the TimeBank, but also aimed to improve coverage by annotating relations between all events and times in the same sentence. However, event tokens that were mentioned fewer than 20 times were excluded and only one TempEval task considered relations between events in different sentences. In practical terms, the resulting evaluations remained sparse.

A major dilemma underlying these sparse tasks is that the unlabeled event/time pairs are ambiguous. Each unlabeled pair holds two possibilities:

  1. 1.

    The annotator looked at the pair of events and decided that no temporal relation exists.

  2. 2.

    The annotator did not look at the pair of events, so a relation may or may not exist.

The training of temporal reasoners is hampered by this ambiguity. In the worst case scenario, evaluation results are untrustworthy. To combat this, our annotation adopts the vague relation, introduced by TempEval 2007, and our approach forces annotators to use it. This is the only work that includes such a requirement.

Finally, this paper is not the first to look into more dense annotations. Some research has explored annotation schemes that encourage annotators to connect all events to the timeline. Bramsen et al. [3] annotated multi-sentence segments of text to build directed acyclic graphs, but the work didn’t focus on events. Kolomiyets et al. [5] annotated “temporal dependency structures”, though they only focused on relations between pairs of events. Most relevant to this paper is . They have the densest annotation, but “the annotator was not required to annotate all pairs of event mentions, but as many as possible”.This paper takes a different tack to annotation by requiring annotators to label every possible pair of events/times in a given window. Thus this work is the first annotation effort that can guarantee its event/time graph to be strongly connected.

We compare the size and density of our corpus to current corpora in \Creftab:relation-ratios. Our corpus is the densest annotation, and contains the largest number of temporal relations to date.

Events Times Rels R
TimeBank 7935 1414 6418 0.7
Bramsen 2006 627 615 1.0
TempEval-07 6832 1249 5790 0.7
TempEval-10 5688 2117 4907 0.6
TempEval-13 11145 2078 11098 0.8
Kolomiyets-12 1233 1139 0.9
Do 201222Do 2012 reports 6264 relations, but this includes both the relations and their inverses (personal communication). We thus halve the count for accurate comparison to other corpora. 324 232 3132 5.6
This work 1729 289 12715 6.3
Table 1: Events, times, temporal relations and the ratio of relations to events + times (R) in various corpora.

3 A Framework for Dense Annotation

Frameworks for annotating text typically have two independent facets: (1) the practical means of how to label the text, and (2) the higher-level rules about when something should be labeled. The first is often accomplished through a markup language. TimeML is used for most temporal corpora, and this paper is no different. The second facet is the focus of this paper: when should an annotator label an ordering relation?

Our proposal thus starts with documents that have been already been annotated with events, time expressions, and document creation times (DCT). The following example sentence serves as our motivating example: {quoting} Police confirmed Friday that the body found along a highway in this municipality 15 miles south of San Juan belonged to Jorge Hernandez. This sentence is represented by a 4 node graph (3 events and 1 time). Its complete graph requires 6 edges to connect all nodes, but as already discussed, few of our current datasets contain complete graphs. This particular sentence is from the TimeBank, and 3 of the 6 edges are labeled.

The impact of these annotation decisions (i.e., when to annotate a relation) can be significant. In this example, a learner must somehow deal with the 3 unlabeled edges. One option is to assume that they are vague or ambiguous. However, all 6 edges have clear well-defined ordering relations: {quoting} belonged before confirmed
before found
before confirmed
before Friday
is included in Friday
is included in Friday33The previous sentence (not shown here) reveals that found occurred on Friday.

Learning algorithms handle these unlabeled edges by making incorrect assumptions, or by ignoring large parts of the temporal graph. Several models with rich temporal reasoners have been published, but since they require more connected graphs, improvement over pairwise classifiers have been minimal [4, 10]. This paper thus proposes an annotation process that builds denser graphs with formal properties that learners can rely on, such as locally complete subgraphs.

3.1 Ensuring Dense Graphs

While the ideal goal is to create a complete graph, the time it would take to hand-label n(n-1)/2 edges is prohibitive. We approximate completeness by creating locally complete graphs over neighboring sentences. The resulting event graph for a document is strongly connected, but not complete. Specifically, the following edge types are included:

  1. 1.

    Event-Event, Event-Time, and Time-Time pairs in the same sentence

  2. 2.

    Event-Event, Event-Time, and Time-Time pairs between the current and next sentence

  3. 3.

    Event-DCT pairs for every event in the text (DCT is the document creation time).

  4. 4.

    Time-DCT pairs for every time expression in the text.

Our process requires annotators to annotate all of the above edge types. This is achieved through an annotation tool that prohibits the skipping of relations. We describe the target relation set and this tool next.

3.1.1 Temporal Relations

The TimeBank corpus uses 14 relations based on the Allen interval relations. The TempEval contests have used both a small set of relations (TempEval-1) and the larger set of 14 relations (TempEval-3). Published work has mirrored this trend, and different groups focus on different aspects of the semantics.

We chose a middle ground between coarse and fine-grained distinctions for annotation, settling on 6 relations: before, after, includes, is included, simultaneous, and vague.

The main reason for not using a more fine-grained set is because we annotate pairs that are far more ambiguous than those considered in previous efforts. Decisions between relations like before and immediately before can complicate an already difficult task. The added benefit of a corpus (or working system) that makes fine-grained distinctions is also not clear. We lean toward higher annotator agreement with relations that have greater separation between their semantics44For instance, a relation like starts is a special case of includes if events are viewed as open intervals, and immediately before is a special case of before. We avoid this overlap and only use includes and before.

3.1.2 Enforcing Annotation

Imposing the above rules on annotators requires automated assistance. We built a new tool that reads TimeML formatted text, and computes the set of required edges. Annotators are prompted to assign a label for each edge, and skipping edges is prohibited. The tool is unique in that it includes a transitive reasoner that infers relations based on the annotator’s latest annotations. For example, if event e1 is included in t1, and t1 before e2, the tool automatically labels e1 before e2. The transitivity inference is run after each input label, and the human annotator cannot override the inferences. This prohibits the annotator from entering edges that break transitivity. As a result, several properties are ensured through this process: the graph (1) is a strongly connected graph, (2) is consistent with no contradictions, and (3) has all required edges labeled. These 3 properties are new to all current ordering corpora. The annotation tool will be publicly available.

3.2 Annotation Guidelines

Since the annotation tool frees the annotators from the decision of when to label an edge, the focus is now what to label each edge. This section describes the annotation guidelines for dense annotation.

The 80% confidence rule: Since the annotators are forced to label all required edges, the decision to label an edge as vague instead of a defined temporal relation is critical. We adopted an 80% rule that instructed annotators to choose a specific non-vague relation if they are 80% confident that it was the writer’s intent that a reader infer that relation. By not requiring 100% confidence, we allow for alternative interpretations that conflict with the chosen edge label as long as that alternative is sufficiently unlikely. In practice, annotators had different interpretations of what constitutes 80% certainty, and this generated much discussion. We mitigated these disagreements with the following rule.
Majority annotator agreement: An edge’s final label is the relation that received a majority of annotator votes, otherwise it is marked vague. If a document has 2 annotators, both have to agree on the relation or it is labeled vague. A document with 3 annotators requires 2 to agree, and 4 annotators require 3 to agree. This agreement rule acts as a check to our 80% confidence rule, backing off to vague when decisions are uncertain (arguably, this is the definition of vague).

In addition to the above, we created other guidelines to encourage consistent labelings. Several of these are inspired by Bethard and Martin [1].
Modal and conditional events: interpreted with a possible worlds analysis. The core event was treated as having occurred, whether or not the text implied that it had occurred. For example, {quoting} They [EVENT expect] him to [EVENT cut] costs throughout the organization. This event pair is ordered (expect before cut) since the expectation occurs before the cutting (in the possible world where the cutting occurs). Negated events and hypotheticals are treated similarly. One assumes the event does occur, and all other events are ordered accordingly. Negated states like “is not anticipating” are interpreted as though the anticipation occurs, and surrounding events are ordered with regard to its presumed temporal span.
Aspectual Events: annotated as is included in their event arguments. For instance, events that describe the manner in which another event is performed are considered encompassed by the broader event. Consider the following example: {quoting} The move may [EVENT help] [EVENT prevent] Martin Ackerman from making a run at the computer-services concern. This event pair is assigned the relation (help is included in prevent) because the help event is not meaningful on its own. It describes the proportion of the preventing accounted for by the move. In TimeBank, the intentional action class is used instead of the aspectual class in this case, but we still consider it covered by this guideline.
Events that attribute a property to a person or event are interpreted to end when the entity ends. For instance, ‘the talk is nonsense’ evokes a nonsense event with an end point that coincides with the end of the talk.
Time Expressions: the words now and today were given “long now” interpretations if the words could be replaced with nowadays and not change the meaning of their sentences. The time’s duration starts sometime in the past and includes the DCT. If nowadays is not suitable, then the now was included in the DCT.
Generic Events: finally, generic events were not skipped. Annotators followed the guideline that generic events can be ordered with respect to each other, but vague with respect to nearby non-generic events.

4 TimeBank-Dense: corpus statistics

We chose a subset of TimeBank documents for our new annotation: TimeBank-Dense. This provided a reliable labeling of events and time expressions. Using the tool described above, we annotated 36 random documents with at least two annotators each to create a training set, development set, and test set. These 36 documents led to an annotation of 4 times as many relations as the entire 183 document TimeBank.

The four authors of this paper were the four annotators. All four authors annotated the same initial document, conflicts and disagreements were discussed, and guidelines were updated accordingly. The rest of the documents were then annotated independently. Document annotation was not random, but we mixed pairs of authors where time constraints allowed.

Final annotator agreement is shown in \Creftab:agree. Average agreement was 65.1% (not far from the TimeBank’s 71% agreement, which did not include the more difficult vague relation). By far, the most significant disagreements pertained to choosing between the vague relation and a specific temporal relation. Since there were only 2 annotators for most documents (and since we require the majority of annotators to agree), in general, the vague label was applied to the final graph if either annotator chose it. This seems appropriate since a disagreement between two annotators directly implies that the relation is vague. \Creftab:specific-agree breaks down the individual disagreements.


tab:relcounts shows the individual relation counts in the final corpus. The vague relation makes up 46% of the relations. This is the first empirical count of how many temporal relations in news articles tend to be truly vague.

Annotated Relation Count

before 2590
after 2104
includes 836
included in 1060
simultaneous 215
vague 5910
Total Relations 12715

Table 2: The number of relations of each type in the TimeBank-Dense corpus.
Annotators # Links Agreement
Ann 1 and Ann 2 9282 0.650
Ann 1 and Ann 4 1605 0.715
Ann 2 and Ann 4 279 0.703
Ann 3 and Ann 4 1549 0.652
Table 3: Agreement between different annotators.
b a i ii s v
b 1776 22 88 37 21 192
a 17 1444 32 102 9 155
i 71 34 642 45 23 191
ii 81 76 40 826 31 230
s 12 8 25 28 147 29
v 500 441 289 356 64 1197
Table 4: Relation agreement between the two main annotators. Most disagreements involved whether a relation should be vague. b=before, a=after, i=includes, ii=is included, s=simultaneous, v=vague

5 Discussion

Both the annotation tool and the TimeBank-Dense corpus is publicly available at XXX. We hope these efforts will encourage a shift in the temporal ordering community to consider the entire document when making local decisions. Further work is needed to balance the difficulty in labeling all possible pairs with the ambiguity inherent in the vague relation. This paper describes our attempt to formalize an annotation framework that can produce corpora with formal guarantees about the structure of the annotated temporal graph. This is the first corpus with guarantees like connectedness, consistency, and the semantics of an unlabeled edge. We look forward to evaluating current algorithms on this dense corpus.


  • [1] S. Bethard, W. J. Corvey, S. Klingenstein and J. H. Martin(2008) Building a corpus of temporal-causal structure.. Cited by: 3.2.
  • [2] S. Bethard(2013-06) ClearTK-timeml: a minimalist approach to tempeval 2013. Atlanta, Georgia, USA, pp. 10–14. External Links: Link Cited by: 1.
  • [3] P. Bramsen, P. Deshpande, Y.K. Lee and R. Barzilay(2006) Inducing temporal graphs. pp. 189–198. Cited by: 2.
  • [4] N. Chambers and D. Jurafsky(2008) Jointly combining implicit constraints improves temporal ordering. pp. 698–706. Cited by: 3.
  • [5] O. Kolomiyets, S. Bethard and M. Moens(2012-07) Extracting narrative timelines as temporal dependency structures. Jeju Island, Korea, pp. 88–97. External Links: Link Cited by: 2.
  • [6] J. Pustejovsky, P. Hanks, R. Sauri, A. See, R. Gaizauskas, A. Setzer, D. Radev, B. Sundheim, D. Day and L. Ferro(2003) The timebank corpus. Vol. 2003, pp. 40. Cited by: 1, 2.
  • [7] N. UzZaman, H. Llorens, L. Derczynski, J. Allen, M. Verhagen and J. Pustejovsky(2013-06) SemEval-2013 task 1: tempeval-3: evaluating time expressions, events, and temporal relations. Atlanta, Georgia, USA, pp. 1–9. External Links: Link Cited by: 1, 2.
  • [8] M. Verhagen, R. Gaizauskas, F. Schilder, M. Hepple, G. Katz and J. Pustejovsky(2007) Semeval-2007 task 15: tempeval temporal relation identification. pp. 75–80. Cited by: 1, 1, 2.
  • [9] M. Verhagen, R. Sauri, T. Caselli and J. Pustejovsky(2010) SemEval-2010 task 13: tempeval-2. pp. 57–62. Cited by: 1, 2.
  • [10] K. Yoshikawa, S. Riedel, M. Asahara and Y. Matsumoto(2009) Jointly identifying temporal relations with Markov Logic. pp. 405–413. Cited by: 3.