Abstract
Lucia Donatelli, Michael Regan, William Croft, Nathan Schneider. Proceedings of the Joint Workshop on Linguistic Annotation, Multiword Expressions and Constructions (LAW-MWE-CxG-2018).
2018.
Introduction: The "Time-Blindness" of AMR
Abstract Meaning Representation (AMR) has emerged as a powerful semantic framework, representing sentence meaning as a rooted, labeled graph. By design, AMR abstracts away from surface morphosyntax to focus on "who did what to whom." However, this abstraction has historically rendered AMR "time-blind," treating past, present, and future events as identical.
For example, in standard AMR, the sentences "Your brother is in the hospital" and "Your brother was in the hospital" result in the same graph. For critical applications like information extraction or automated decision-making, the distinction between a current emergency and a past event is vital. This paper introduces a coarse-grained framework to annotate tense and aspect within AMR, filling this gap without sacrificing the simplicity and scalability of the original system.
A Framework for Tense and Aspect
The authors propose augmenting AMR with a small set of non-core roles focusing specifically on finite predications (verbs that carry tense). These roles are semantic rather than grammatical; for instance, "I leave for Paris tomorrow" is marked as future time despite the present-tense verb.
1. Temporal Location (Tense)
Instead of complex logic, the framework utilizes a four-way distinction relative to the speech time ("now"):
- Present:
:time (n / now) - Past:
:time (b / before :op1 (n / now)) - Future:
:time (a / after :op1 (n / now)) - Continuative:
:time (u / up-to :op1 (n / now))(e.g., "The wildfire has spread.")
2. Event Structure (Aspect)
To describe how an event unfolds, the authors introduce four categorical labels. These labels allow the graph to distinguish between inherent properties (e.g., being a doctor) and transitory states (e.g., being tired).
As shown in the definitions above, the framework handles a variety of aspectual nuances:
:stable(+/-): Distinguishes permanent states from temporary ones.:ongoing(+/-/?): Indicates if an event is viewed from the "inside" as being in progress.:complete(+/-): Used for "telic" events (those with a natural endpoint) to show if that endpoint was reached.:habitual(+): Marks regularly recurring events.
Validation and Key Insights
The authors validated this schema using The Little Prince corpus. Pilot studies involving both expert and novice annotators were conducted on 200 existing AMRs to ensure the categories were intuitive and robust.
Several key insights emerged from this process:
- Abstracting the "Perfect": The English perfect (e.g., "I have eaten") is broken down into three semantic categories: continuative (continues to now), existential (happened at least once), and recent (just happened).
- Stativity Matters: The distinction between
:stable +and:stable -provides a major semantic upgrade, encoding the difference between inherent properties and transitory states. - Simplicity over Completeness: By avoiding a full temporal calculus and sticking to coarse-grained categories, the authors ensure the annotation process remains fast enough for large-scale production.
Limitations and Future Work
While the pilot studies showed the framework is highly applicable to English sentences, several areas remain for future development. The current focus is strictly sentential, meaning it does not yet handle discourse-level temporal relations (how time flows across a paragraph). Additionally, the framework currently ignores non-finite forms like infinitives and gerunds, as well as complex "irrealis" events (events that might have happened but didn't). Finally, the authors acknowledge that extending this system to languages without overt tense, such as Chinese, will require further research.
Summary updated