Information Technology Reference
In-Depth Information
After this initial work, research on semantic role labeling (SRL) has grown
steadily, and in the years 2004 and 2005 [3, 4] a shared task at the CoNLL
2
was
defined, in which several research institutions compared their systems. In the mean-
time, besides FrameNet, another corpus with manually annotated semantic roles has
been prepared, PropNet [21], which differs from FrameNet in the fact that it has
general semantic roles not related to semantic frames. PropNet is also the corpus
used for training and evaluation of research systems on the SRL shared task. A
similar corpus to FrameNet for the German language has been created by the Salsa
project [10], and a discussion on the differences and similarities among these three
projects is found in [9].
Frame
Evidence
Definition
:The
Support
, a phenomenon or fact, lends support to a claim or proposed course of
action, the
Proposition
, where the
Domain of Relevance
mayalsobeexpressed.
Lexical units
:
argue.v
,
argument.n
,
attest.v
,
confirm.v
,
contradict.v
,
corroborate.v
,
demonstrate.v
,
dis-
prove.v
,
evidence.n
,
evidence.v
,
evince.v
,
from.prep
,
imply.v
,
indicate.v
,
mean.v
,
prove.v
,
reveal.v
,
show.v
,
substantiate.v
,
suggest.v
,
testify.v
,
verify.v
Frame Elements
:
Proposition [PRP]
This is a belief, claim, or proposed course of action to which the
Support lends validity.
Support is a fact that lends epistemic support to a claim, or that
provides a reason for a course of action.
Support [SUP]
...
Examples
:
And a [
SUP
sample tested
]
REVEALED
[
PRP
some inflammation
].
It says that [
SUP
rotation of partners
]doesnot
DEMONSTRATE
[
PRP
independence
].
Fig. 4.4.
Information on the frame
Evidence
from FrameNet.
SRL is approached as a learning task. For a given target verb in a sentence, the
syntactic constituents expressing semantic roles associated to this verb need to be
identified and labeled with the right roles. SRL systems usually divide sentences
word-by-word or phrase-by-phrase and for each of these instances calculate many
features creating a feature vector. The feature vectors are then fed to supervised
classifiers, such as support vector machines, maximum entropy, or memory-based
learners. While adapting such classifiers to perform better on this task could bring
some improvement, better results can be achieved by constructing informative fea-
tures for learning. A thorough discussion of different features used for SRL can be
found in [14, 22].
4.3.3 Frames and Roles for Annotating Cases
On the one hand, in knowledge engineering there are knowledge tasks and knowledge
roles to represent knowledge; on the other hand, in natural language understanding
there are semantic frames and semantic roles to represent meaning. When knowledge
2
Conference of Natural Language Learning
Search WWH ::
Custom Search