Semantic Role Labeling: An
Introduction to
the Special Issue
Introduction
What is semantics role labelling:an introduction to the special issue? The primary task of semantics role labelling(SRL)is to indicate exactly what semantics relations hold among a predicate and its associated participants and properties,with these relations drawn from a pre-spesified list of possible semantic role for that predicate (or class of predicates).In order to acomplish this,the role -bearing constituantes in a clause must be identified and their correct semantic role labels asssigned,as in:[The girl on the swing][whispered]to[the boy beside her].
This special issue of computational linguistics presents several articles representing the state of the art in SRL,and this overview is intended to provide a broader context for that work.First,we briefly discuss some of the linguistic views on semantics roles that have had the most influence on computational approaches to SRL tasks.Next,we show how the linguistic notions have influenced the development of resources that support SRL.We then provide on overview of SRL methods and describe the state of the art as well as current open problems in the fields.
Summary
Since the foundational work of Fillmore,considerable linguistic research has been devoted to the nature of semantic roles,such as agent and theme,there is no consensus on a definitive list of semantic roles,or even whether such as a list exist.
A major focus of work in the
linguistics community is on the mapping between the predicate–argument structure that
determines the roles, and the syntactic realization of the recipients of those roles (Grimshaw
1990; Levin 1993; Levin and Rappaport Hovav
2005) .Semantic role lists are
generally viewed as inadequate for explaining the morphosyntactic behavior of argument
expression, with argument realization dependent on a deeper lexical semantic representation
of the components of the event that the predicate describes .Although much
of the mapping from argument structure to syntax is predictable, this mapping is not
completely regular, nor entirely understood .An important question for SRL,
therefore, is the extent to which performance is degraded by the irregularities noted in
linguistic studies of semantic roles.
Nonetheless, sufficient regularity
exists to provide the foundation for meaningful generalizations .Much research has
focused on explaining the varied expression of verb arguments within syntactic positions
(Levin 1993) .A major conclusion of that work is that the patterns of syntactic
alternation exhibit regularity that reflects an underlying semantic similarity among verbs,
forming the basis for verb classes .Such classes, and the argument structure
specifications for them, have proven useful in a number of NLP tasks (Habash, Dorr, and Traum 2003;
Shi and Mihalcea 2005), including SRL (Swier and Stevenson 2004), and have provided
the foundation for the computational verb lexicon VerbNet (Kipper, Dang, and Palmer
2000).
This approach to argument
realization focuses on the relation of morphosyntactic behavior to argument semantics, and
typically leads to a general conceptualization of semantic roles .In frame semantics
(Fillmore 1976), on the other hand, a word activates
a frame of semantic knowledge that
relates linguistic semantics to encyclopedic knowl-edge. This effort has tended to
focus on the delineation of situation-specific frames (e.g.,an Arrest frame) and correspondingly
more specific semantic roles (e.g., Suspect and
Authorities) that codify the
conceptual structure associated with lexical items (Fillmore,Ruppenhofer, and Baker 2004) .With a
recognition that many lexical items could activate any such frame, this approach leads
to lexical classes of a somewhat different nature
than those of Levin (1993) .Whereas
lexical items in a Levin class are syntactically homogeneous and share coarse
semantic properties, items in a frame may syntactically vary somewhat but share
fine-grained, real-world semantic properties.
A further difference in these
perspectives is the view of the roles themselves .In defining verb classes that capture
argument structure similarities, Levin (1993) does not explicitly draw on the notion of
semantic role, instead basing the classes on behavior that is hypothesized to reflect the
properties of those roles .Other work also eschews the notion of a simple list of
roles, instead postulating underlying semantic structure that captures the relevant
properties (Levin and Rappaport Hovav 1998) .Interestingly,147 Computational Linguistics Volume 34, Number 2
as described in Fillmore,
Ruppenhofer, and Baker (2004), frame semantics also avoids a predefined list of roles, but for
different reasons .The set of semantic roles, called frame elements , are chosen for each frame, rather
than being selected from a predefined list that may not capture the relevant
distinctions in that particular situation .Clearly, to the extent that disagreement
persists on semantic role lists and the nature of the roles themselves, SRL may be working on a
shifting target. These approaches also differ in the
broad characterization of event participants (and their roles) as more or less
essential to the predicate .In the more syntactic-oriented approaches, roles are typically
divided into two categories:
arguments, which cap-ture a core relation, and adjuncts,which are less central .In frame
semantics, the roles are divided into core
frame elements (e.g., Suspect,
Authorities, Offense) and periph-eral
or extra-thematic
elements (e.g., Manner, Time,
Place). These distinctions carry over into SRL, where we see that
systems generally perform better on the more central arguments.
Finally, although predicates are
typically expressed as verbs, and thus much work both linguistics and SRL focuses
on them, some nouns and adjectives may be used predicatively, assigning their own
roles to entities (as in the adjective phrase proud that we finished the paper, where the subordinate clause is a
Theme argument of the adjective proud) .Frame semantics tends to include
in a frame relevant non-verb lexical items,due to the emphasis on a common
situation semantics .In contrast, the morphosyntactic approaches have focused on defining
classes of verbs only, because they depend on common syntactic behavior that may
not be apparent across syntactic categories.
Interestingly, prepositions have a
somewhat dual status with regard to role labeling.
In languages like English,
prepositions serve an important function in signaling the rela-
tion of a participant to a verb .For
example, it is widely accepted that to in give the book to Mary
serves as a grammatical indicator of
the Recipient role assigned by the verb, rather than as a role assigner itself .In
other situations, however, a preposition can be viewed as a role-assigning predicate in its
own right .Although some work in computational linguistics is tackling the issue of
the appropriate characterization of prepositions and their contribution to semantic role
assignment (as we see subsequently), much work remains in order to fully integrate
linguistic theories of prepositional function and semantics into SRL.
Weakness
The article is not complete.The article need to repair with a good word and need to add the knowledge more.
Strenghtness
This article is good to the readers and the readers can learn about semantics linguistics in article.
Conclusion
SRL systems have been shown to
perform reasonably well in some controlled
experiments,measures in the low 80s on standard
test collections for English.
Still, a number of important
challenges exist for future research on SRL .It remains
unclear what is the appropriate
level of syntax needed to support robust analysis of
semantic roles, and to what degree
improved performance in SRL is constrained by the
state-of-the-art in tagging and
parsing .Beyond syntax, the relation of semantic roles to
other semantic knowledge (such as
WordNet, named entities, or even a catalogue of
frames) has scarcely been addressed
in the design of current SRL models .A deeper
understanding of these questions
could help in developing methods that yield im-
proved generalization, and that are
less dependent on large quantities of role-annotated
training data.
Indeed, the requirement of most SRL
approaches for such training data, which is
both difficult and highly expensive
to produce, is the major obstacle to the widespread
application of SRL across different
genres and different languages .Given the degrada-
tion of performance when a
supervised system is faced with unseen events or a testing
corpus different from training, this
is a major impediment to increasing the application
of SRL even within English, a
language for which two major annotated corpora are
available .It is critical for the
future of SRL that research broadens to include wider
investigation of unsupervised and
minimally supervised learning methods.
In addition to these open research
problems, there are also methodological issues
that need to be addressed regarding
how research is conducted and evaluated .Shared
task frameworks have been crucial in
SRL development by supporting explicit compar-
isons of approaches, but such
benchmark testing can also overly focus research efforts
on small improvements in particular
evaluation measures .Improving the entire SRL
approach in a significant way may
require more open-ended investigation and more
qualitative analysis.
