=Paper=
{{Paper
|id=None
|storemode=property
|title=The Multilingual Procedural Semantic Web
|pdfUrl=https://ceur-ws.org/Vol-936/paper7.pdf
|volume=Vol-936
|dblpUrl=https://dblp.org/rec/conf/semweb/NirenburgM12
}}
==The Multilingual Procedural Semantic Web==
https://ceur-ws.org/Vol-936/paper7.pdf
The Multilingual Procedural Semantic Web
A Position Paper
Sergei Nirenburg and Marjorie McShane
University of Maryland Baltimore County
sergei@umbc.edu, marge@umbc.edu
The stated goal of the Semantic Web community is to turn the Web into a
richly annotated resource, making its content more amenable to applications
that involve machine reasoning. The most widely discussed language-oriented
aspect of this vision involves the creation and use of an inventory of markup
tags that indicate select semantic types. So, the “semantics” of the Semantic
Web is not the semantics of full texts or even full sentences, but rather of
select elements of text and extra-textual information. Moreover, the
annotations are expected to be largely carried out manually, so broad coverage
is unlikely, as are consistency and universal public-spiritedness on the part of
annotators (cf. Doctorow, no date). Compare this to the ideal semantic web,
which would be automatically generated from the unadorned web by
processors that would carry out lexical disambiguation, referential
disambiguation, and the interpretation of textual implicatures, such as the
recognition of irony and indirect speech acts. Such full semantic
interpretations of web content would serve as optimal input for machine
reasoners.
It is common practice in the field of AI to assume the availability of
such knowledge structures – in fact, practically all work on machine reasoning
over the past decades has used hand-crafted, complete, unambiguous
knowledge structures as input. How that could be achieved automatically was
always considered a separate issue, delegated to the NLP community. The
NLP community, however, by and large abandoned the task of deep semantic
analysis some 20 years ago, opting to pursue either (a) knowledge lean, “low-
hanging fruit” tasks that contribute to the configuration of better NLP
applications in the near term but do not contribute to the ultimate goal of
automatic text understanding or (b) method-oriented work, in which the
methods themselves are of first priority and natural language serves primarily
as a source of data sets.1
The Semantic Web community has largely followed the spirit of the
NLP majority by deeming full semantics to be too complex to be pursed. As
such, the semantics of the Semantic Web is effectively constrained to
selective annotation of text strings in ways that are considered feasible in the
short term. The preferences of the Semantic Web community are reflected in
1
Space does not permit a full motivation for these generalizations. For that see Nirenburg and
McShane, forthcoming as well as the historical references cited therein.
the selection of foci of work: the development of formal standards, metadata
tag sets, ontologies to be used as the content of tag sets, and so on. While we
appreciate the common preferences of the mainstream NLP and Semantic-
Web communities, and while the material below describes an attempt to
contribute to the near-term gains they seek, our contributions must be framed
within the research paradigm that we deem the most promising for the long-
term utility of any NLP, be it for the web or any other corpus: computational
deep-semantic processing. We will argue that one near-term results can be
achieved within a theory and methodology that seek full understanding of
texts, along with associated sophisticated behaviors, by intelligent agents.
Our research program is an outgrowth of the theory of Ontological
Semantics, which studies the processes of automatically extracting,
representing and manipulating meaning in natural language texts. Analysis by
the OntoSem text analyzer pursues all of the desiderata listed in the
introductory paragraph, seeking fully specified, unambiguous, ontologically
grounded meaning representations that are more amenable to machine
reasoning than highly ambiguous natural language texts (Nirenburg and
Raskin 2004). Of course, the automatically generated structures are not yet
perfect, as that would be well beyond the current state of the art. However, we
are making direct progress toward this goal, which suggests that the vision of
fully interpreted content delivered over the internet should not be neglected. A
prototype for this vision was demonstrated in the implemented SemNews
application (Java et al. 2007), which took web-delivered news feeds as input
and generated semantic interpretations of them represented as RTF structures.
Significantly, Ontological Semantics is a language-independent
theory, most of whose knowledge bases (e.g., ontology, fact repository, rule
sets for agent decision-making) and reasoning engines are language-
independent. In fact, in the intentionally provocatively titled “An NLP lexicon
as a largely language independent resource” (McShane et al. 2005), we
describe how much of the information found even in the lexicons used to
support OntoSem language processing can be directly reused across languages
(more on this below). Once the input strings from any language have been
interpreted using a battery of processors, the resulting text-meaning
representations can be reasoned over by a single set of engines. Language-
neutrality offers not only great savings in time for the acquisition of
knowledge resources and development of processors, it also offers
consistency of processing across languages.
The core point of this statement, which follows basic tenets of
configuring intelligent agents within the OntoAgent environment, is as
follows. The only realistic way to enhance the Web with useful semantic
annotations is automatically. Semantic analysis is, by its very nature,
procedural: a system – hereafter “agent” – receives some input, analyzes it in
context, and generates an interpretation. The component functions of this
process, like all functions, are subject to error; as a result, the agent must be
able to evaluate its confidence in the function’s output based on the overall
predictive power of the function as well as the confidence in each input
parameter value. Depending upon the calculated confidence in output, the
agent can decide whether or not to use the output in a given application. Since
many of the actual functions used to generate interpretations are identical (or
at least very similar) cross-linguistically, they should be reused to support
both efficiency and consistency in the treatment of Web content. Since
different functions take different types of parameter values as input – and
since some parameter values are quite easy to compute with high confidence
while others are much more difficult – it is possible to introduce procedural
semantic analyses to web content in a progressive manner, over time.
We will now illustrate how automatic annotations, generated using
cross-linguistically applicable functions, could be incorporated into the
Semantic Web over time. We will use as sample phenomena so-called
indexical expressions, which are strings whose absolute meaning can be
understood only with reference to a specific context: e.g., he, themselves, over
there, now, in a few minutes, the preceding paragraph. The reason why one
would want all these indexical expressions fully, locally resolved as
annotations to Semantic Web content should be self-explanatory: it is more
directly useful to an automatic reasoner to have access to the information
“John. W. Lacey III of Kansas City, Kansas died on July 5, 1974 in
Washington, D.C. from complications of heart disease” rather than an
expression that could be synonymous given the right context: “Yesterday, in
that same place, that happened to one of our local boys.”
There exists an unfortunate, in our opinion, tradition within the NLP
community to treat indexicals in a suboptimal way on at least three fronts. (1)
Unrealistic preconditions. Most work on automatic pronoun resolution, for
example, involves supervised learning (i.e., learning from manually annotated
corpora), whose resultant engines require that all future inputs be already
annotated, to perfection, in the expected way. (2) All-or-nothing
classifications. Indexicals are regularly (albeit often tacitly) categorized as
“easy” (e.g., he) or “too hard” (e.g., pronominal that), whereas the actual
easy/hard distinction is largely based on the contextual usage of the element.
(3) Language specificity. Most work on indexicals in NLP and descriptive
linguistics is language-specific, but many resolution functions are actually
cross-linguistically applicable. 2
Our proposal is to apply to the Semantic Web the same types of cross-
linguistically applicable indexical resolution functions that are already used in
the OntoSem environment. The key to successful realization of this proposal
2
Theoretical work, like that grounded in the tradition of theoretical syntax, typically lacks the
needed level of descriptive detail to be of practical utility for NLP.
involves classifying usage cases for indexicals with respect to which
parameter values are required for each decision function and how and with
what confidence those parameter values can be obtained and in each context.
Let us begin by considering some types, sources and confidence levels
of input parameters that might contribute to functions for resolving indexicals
found on the Web. The surface string: always available, maximally high
confidence. Semantic web annotations: sometimes available for some types
of entities; confidence varies depending on the source, type of tag, etc.
Traditional web annotations: typically available for html documents; some
types of tags (as for formatting) are of high confidence but might be noisy and
difficult to automatically interpret. Automatic “preprocessing” of text:
preprocessing (detecting tokens, proper names, dates, etc.) is a cornerstone of
NLP, but web content can be error-prone due to the metadata text, embedded
media, etc. Syntactic analysis: another mainstream NLP task though even the
best current parsers achieve far less than perfect results. Basic semantic
analysis (word sense disambiguation and the determination of
dependencies): carried out by few NLP systems, OntoSem being among
them; analyses tend to be extremely useful in supporting high-level tasks like
resolving indexicals, but they are error-prone. Procedural semantic routines to
resolve indexicals become more complex, and typically of lower confidence,
as they incorporate the latter types of features. But, centrally important for this
proposal, the difficulty of each usage case and its associated confidence level
can typically be automatically calculated, thus suggesting in which types of
applications the automatic results might best be used. Let us consider just a
few examples of indexical treatment.
Relative time expressions – such as today, now, three weeks from
tomorrow and in a little while – can readily be resolved to real times (month,
day, year, etc.) if (a) the “anchor time” – i.e., the time of the post (article, etc.)
– is known, and (b) the time expression is used outside of direct speech. The
former is expected to be recorded in Semantic Web tags, and the latter can
typically be determined with high confidence using a preprocessor. (If the
expression is within direct speech, then the time of speech must be
determined, which requires semantic analysis.) Within OntoSem, the actual
functions that can calculate, e.g., today vs. three weeks from tomorrow are
recorded in the “meaning procedures” zones of the respective lexicon entries
(McShane et al. 2004). As mentioned earlier, OntoSem lexicons are largely
language-independent, meaning that their semantic descriptions and
procedural semantic routines can be reused across languages (McShane et al.
2005); so the procedure already available for English today can be used to
derive the full meaning of Czech dnes or Hebrew – םהיוassuming, of course,
that preprocessors for these languages are available.
A similar example is the pronoun I, which can be resolved with high
confidence in one of two cases: (1) if it is used outside of direct speech and
the piece has a single author as indicated by a Semantic Web tag or (2) it is
used within an instance of direct speech that contains a preceding instance of
I. In this latter case, although the real-world referent cannot be confidently
distinguished, the coreference relationship between instances of I can be. Now
contrast I with its plural counterpart we. We is substantially more difficult to
interpret since a single author often affirms group membership – explicitly or
implicitly – then subsequently speaks on behalf of the group. Alternatively, a
piece can be written by more than one person, with we in a given context
referring either to a subset of the authors or to a larger community to which
they all, or a subset of them, belong. The extensive analysis required by
people to craft a robust function for resolving we underscores why we (yes,
we!) should take a cross-linguistic approach to developing procedural
semantic functions for the web: it will save the community time and foster
consistency of interpretations. Our initial work on the resolution of we within
OntoSem includes subfunctions for resolving I and we that involve different
kinds of heuristic evidence, some of which we can expect to be available for
any language in the short term and other aspects of which require full-blown
semantic analyses of the type we are working toward.
Let us conclude by stating that there are many more largely cross-
linguistically applicable procedural semantic routines beyond indexicals, for
example, the procedure for resolving very (as applied to different types of
expressions) are (McShane et al. 2004).
References Cited
Doctorow, C. (No date) Metacrap: Putting the torch to seven straw-men of the meta-
utopia. Available at http://www.well.com/~doctorow/metacrap.htm
Java, Akshay, Sergei Nirenburg, Marjorie McShane, Timothy Finin, Jesse English,
Anupam Joshi. 2007. Using a natural language understanding system to generate
Semantic Web content. International Journal on Semantic Web & Information
Systems, 3(4), 50-74. October-December 2007.
McShane, Marjorie, Sergei Nirenburg and Stephen Beale. 2005. An NLP lexicon as a
largely language independent resource. Machine Translation 19(2): 139-173.
McShane, Marjorie, Stephen Beale and Sergei Nirenburg. 2004. Some meaning
procedures of Ontological Semantics. Proceedings of LREC-2004.
Nirenburg, S. & McShane, M. Forthcoming. Natural Language Processing. To
appear in S.Chipman (ed.) The Oxford Handbook of Cognitive Science.
Nirenburg, S. & Raskin, V. 2004. Ontological Semantics. The MIT Press.