Information Technology Reference
In-Depth Information
TLINK track systems usually use machine learning approaches such as Sup-
port Vector Machine (SVM) [6] and Maximum Entropy (ME) [1]. Some systems
have used rule-based approaches [1]. Although generally rule-based approaches
do not outperform machine learning [7], some linguistic rules can be especially
useful in processing patient recordsfor example, a rule stating that an event in
a patients hospital course section usually appears after the admission time and
before the discharge time. In this paper, we explain an approach that combines
linguistic rules and CRFs to carry out temporal information extraction.
2 Related Work
Chang et al. [1] proposed a hybrid method to identify TLink type. They com-
bined a rule-based approach and an ME-based approach and developed an al-
gorithm to integrate the results of the two approaches. Tang et.al. [8] divided
the TLINK extraction task into three sub-tasks: 1. Identify TLinks between
events and section time; 2. Identify TLinks between events/times within one sen-
tence (sentence-internal TLinks); 3. Identify TLinks between events/times across
sentences (cross-sentence TLinks). They built classifiers for each of the three
categories of TLinks above. Then they applied event position information, bag-
of-words, part-of-speech (POS) tags, dependency-related features, time-related
features, event-related features, distance features and conjunction features. Their
system, based on linear chain CRFs and the SVM model, achieved a F-score
69.32% in the TLINK-only track in the i2b2 2012 challenge.
3 Approaches
Our TLINK extraction approach includes a rule-based approach and a CRFs-
based approach. The rule-based approach uses Tangs definition [8] to separate
time-event and event-event pairs into three categories, including section-event,
within-sentence, and cross-sentences pairs. According to the linguistic charac-
teristics of these categories, our rule-based system assign TLink types to these
pairs. Since some events/times are related to their surrounding event/times, we
consider TLink extraction a sequence labeling problem and use the CRFs model
to solve it.
3.1 Rule-Based Approach
We formulate TLink extraction as a relation-classification task. Given an
event/time pair (
i,j
), which might be one of the event-event, time-event or
section-event pairs, we need to determine the TLink type of this pair e.g., BE-
FORE, which means that
i
is before
j
.Weusethepredicate
TLink
(
i,j,tlink
)
to describe that
i
has the relation type
tlink
for
j
,where
tlink
∈{
BEFORE,
. For example,
TLink
(
i,j,
“
AFTER
)meansthat
OVERLAP, AFTER, etc ...
}
i
occured after
j
.
Search WWH ::
Custom Search