Temporal information conveyed by language describes how the world around us changes through time. Events, durations and times are all temporal elements that can be viewed as intervals. These intervals are sometimes temporally related in text. Automatically determining the nature of such relations is a complex and unsolved problem. Some words can act as “signals” which suggest a temporal ordering between intervals. We use these signal words to improve the accuracy of a recent approach to classification of temporal links.

1.
Using Signals to Improve Automatic
Classification of Temporal Relations

2. Time in language
● In natural languages, significant effort is devoted to
describing time (tense, aspect, adverbials).
● We have occurrences and states which we can temporally
“position” or ask about.
● Events, points and periods in time can all be related with
language.
● Understanding and processing this information is difficult.
● How can we formally describe time in language?

3. Temporal Annotation
What to annotate?
● Events and time expressions (intervals)
● Temporal, aspectual and subordinate links between intervals
● Signals that indicate recurrence or temporal ordering
TimeML is a formal specification for annotating these kinds of
entity
TimeBank is an annotated corpus of ~65 000 tokens

4. Temporal signals
Temporal links between intervals (TLINKs) specify a relation –
BEFORE, AFTER, INCLUDES
Implicit sources of information for TLINKs:
● Tense/aspect ­ “I had showered and I left.”
● World knowledge ­ “Ben pulled out a gun. The smell of
cordite filled the air.”
We also have explicit information from signals.
● “The road was built. Subsequently, travel time improved.”
We hypothesize that signal words can be described as
features to improve TLINK classification.

5. Baseline and corpus
The TLINK classification task is difficult:
● TempEval­1: 59%
● TempEval­2: 61%
● Most­common­class: 50­55%
We replicated established recent work, using a merge of
TimeBank v1.2 and the AQUAINT TimeML corpus.
● Event­event TLINK classification: 60%

6. Feature set
To augment the baseline's features with information about signals:
● Signal phrase
Textual position of event word and signal can affect temporal
interpretation of a relation:
● “I run before I sleep”
● “Before I run I sleep”
We capture ordering using these features:
● Arg1 / signal order
● Signal / Arg2 order
● Token distance between arg1 / signal / arg2

7. Results
Adding signals to our feature set improved overall performance
from 60.32% to 61.46% ­ marginal improvement.
Explanation:
Only 5.1% of TLINKs in evaluation data use a signal – the corpora
show strong evidence of under­annotation.
So, we split our data into TLINKs with and without signals.
● On unsignalled links, performance remained the same
● On signalled links, we saw a jump to 82.19% classification
accuracy

8. Conclusion
Signals as described by us are helpful to the TLINK
classification task.
Future work
● Automatically annotate signals
● Add more sophisticated features for signals
● Repair some under­annotation in TimeBank
Thanks ­ Any questions?