Scribd
Upload
44 views2 pages

Factual Information at All, This Is Very Likely To Take The Form of Generic Knowledge. Indeed

Uploaded by

Rodrigo Sanz
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
44 views2 pages

Factual Information at All, This Is Very Likely To Take The Form of Generic Knowledge. Indeed

Uploaded by

Rodrigo Sanz
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 2

Frege in space.

2.6 Meaning and reference

The symbol grounding challenge raised by philosophers and cognitive scientists pertains to the perceptual
underpinnings of our generic knowl- edge of concepts (you need to have seen a dog to truly grasp the meaning of
the word dog). The dominant tradition in formal semantics stresses instead another type of relation between
linguistic signs and the external world, characterizing meaning in terms of reference to a specific state of the
world, above and beyond our ability to perceive it. Knowing the meaning of the statement Marco is a dog is
knowing under which outside-world conditions this statement would be true. To calculate this, standard
denotational semantics takes as its primitives individuals (objects or events), truth values and propositions
(possible worlds, states of a↵airs).

The focus of denotational semantics and DSMs is very di↵erent, and so are their strengths and weaknesses. In
denotational semantics, proper names are the simple cases, those that directly point to indi- viduals, unary
predicates refer to sets of individuals having a certain property at a given time and world, binary ones refer to
sets of ordered pairs of individuals, and so forth. In turn, quantifiers express relations between sets, modifiers
typically reduce the size of sets, predicate con- junction intersects them, etc. (see, e.g., Heim and Kratzer 1998).
This model of meaning has been designed to express episodic knowledge— facts that are true of specific
individuals at specific places and times. Capturing the meaning of generic sentences—statements about laws,
regularities or tendencies of whole classes of objects—requires a com- plex quantificational apparatus (Krifka et
al. 1995; Cohen 2004) and is a widely debated but still ill-understood topic

DSMs, on the other hand, are extracted from large corpora where proper names, common
nouns and other predicates refer to states of the world and events spanning a large chunk of
time and space, reflecting different points of view, etc. So, if they are able to extract any
factual information at all, this is very likely to take the form of generic knowledge. Indeed,
a typical application of corpus-based semantics is the extraction of commonsense-
knowledge “factoids” that are generally useful while not universally true: bananas are
yellow, birds fly, etc. (e.g., Eslick 2006; Schubert and Tong 2003).

Copestake and Herbelot (2012) to provide an extensional semantics for DSMs

Two things should be noted here

First, it is perfectly possible to give a reference-based interpretation of DSMs, the question is


what it tells us. Suppose we characterize the meaning of a word w in the corpus C in terms of
its sentential context, expressed via a vector of binary values: for every word g, vectorw(g) = 1
if g appears along with w in some sentence in C, 0 otherwise. Suppose we now call S the set
of all words for which vectorw gives 1, i.e. the set of all words that cooccur with w some
sentence, and treat the lexicon L of C as our domain of reference. Now S, built as an
approximation of the ‘meaning of w’ in a DSM, is at the same time the denotation in L of the
expression “the lexical sentential context of word w” in denotational semantics.

With some added complexity, it is straightforward to give a referential translation of vectors


which do not contain just binary numbers, but any integer value. What this example shows is
that the divide between DSM and denotational semantics is not reference/lack- of-reference,
but rather reference to linguistic strings (which we can easily record) or to objects (which we
cannot). The question becomes what the referential meaning of the noun phrase the linguistic
context of “dog” can tell us about the referential meaning of dog. As we hope to show, quite a
lot

But now what about the linguistic context of “John”? How can DSMs deal with objects that
are often described as being “purely referential”, empty of descriptive content? (proper
names, demonstratives, personal pronouns, etc.; Kripke 1980). We believe that in DSMs
there is no prin- cipled distinction between proper names and common nouns, but there are
very far-reaching practical ones. If we consider names like Barack Obama and bare nouns
like presidents, the di↵erence is small; both will appear in highly informative contexts;
people will write contrasting things about Barack Obama, but so they will about presidents
or just about any common noun. Jut like common nouns, proper names can be polysemous
(Italy lost to Spain—the soccer team; Italy is boot-shaped— the land, etc.), and the same
techniques mentioned above for common nouns can be used to make the right facets of
meaning emerge in the proper combination. But moving on the scale of referential
expressions from Barack Obama to Obama, then to Barack, to that person, to him (or here,
now, any finite tense marker), the dimension of homonymy increases dramatically. Pure
referential expressions are infinitely more ambiguous than descriptive ones, and this causes
a proliferation of ap- parent inconsistencies.

At least for the time being, we will just treat denotational semantics and DSMs as covering complementary
aspects of meaning. To exem- plify, suppose we hear the sentence A dog is barking. Our distributional- feature-
based representation of its constituents will provide us with a sketch of typical contexts in which it can be
uttered truthfully, which can orient our perceptual system to pick up the relevant cues to de- termine if a dog is
indeed barking right now, so that we can evaluate the referential meaning of the sentence. Indeed, to step into
science fiction for a moment, given that state-of-the-art computational image analysis systems produce
vectorial representations of objects (see, e.g., Grauman and Leibe 2011), the process of verifying the state of
a↵airs described by an utterance against perceptual input could take the form of operations on distributional
and perceptual vectors, the former rep- resenting the (parts of the) utterance, the second representing objects
and possibly events in the perceived world. This idea gains further plausibility if we adopt distributional vectors
that record perceptual information coming from vision and other senses, as briefly discussed at the end of the
previous subsection.

You might also like