Appears in AAAI 2006. 1
Societal Grounding is Essential to Meaningful Language Use
David DeVault
1Department of Computer Science
Rutgers University
Picataway, NJ 08845-8020
David.DeVault@rutgers.edu
Iris Oved
Department of Philosophy
Rutgers University
New Brunswick, NJ 08901-1411
irisoved@eden.rutgers.edu
Matthew Stone1,2
2Human Communication Research Centre
University of Edinburgh
Edinburgh EH8 9LW, UK
Matthew.Stone@rutgers.edu
Abstract
Language engineers often point to tight connections be-
tween their systems’ linguistic representations and accumu-
lated sensor data as a sign that their systems really mean what
they say. While we believe such connections are an impor-
tant piece in the puzzle of meaning, we argue that perceptual
grounding alone does not suffice to explain the specific, sta-
ble meanings human speakers attribute to each other. Instead,
human attributions of meaning depend on a process of soci-
etal grounding by which individual language speakers coor-
dinate their perceptual experience and linguistic usage with
other members of their linguistic communities. For system
builders, this suggests that implementing a strategy of soci-
etal grounding would justify the attribution of bona fide lin-
guistic meaning to a system even if it had little perceptual
experience and only modest perceptual accuracy. We illus-
trate the importance and role of societal grounding using an
implemented dialogue system that collaboratively identifies
visual objects with human users.
Introduction
In this paper, we treat meaningful language use as an explicit
design goal for conversational systems: they should mean
what they say. We argue that achieving this goal requires
that implementations explicitly connect system meaning to
societal standards. Part of using language meaningfully, we
claim, is working to keep your meaning aligned with what
others mean in your community. Our case rests on strong
intuitions about speaker meaning in specific contexts. We
draw on these intuitions to understand and to criticize imple-
mentations that construe meaning in exclusively perceptual
terms, and to articulate an alternative approach.
Systems that use language face a traditional line of objec-
tion that any meaning in a computer program’s “utterances”
is merely parasitic on the intentions and interpretations of its
programmers. In its strongest form (Searle 1980), the prob-
lem is seen as endemic to computation: none of the sym-
bols in a computer program, including any linguistic repre-
sentations it may have, are ever intrinsically meaningful to
the system; at best, it is argued, we engineer and arrange
them in such a way that they seem meaningful to us. Other
Copyright c© 2006, American Association for Artificial Intelli-
gence (www.aaai.org). All rights reserved.
well-known arguments dispute the meaningfulness of lan-
guage use in specific extant systems; the symbols they use
to achieve linguistic meaning have been variously held to be
hopelessly impoverished due to the limited range of infer-
ences the system is able to draw using them (Dreyfus 1979),
limited problematically to the “narrow micro-world” of the
programmer’s chosen domain theory (Winograd & Flores
1986), or effectively meaningless due to the lack of any cou-
pling with the external world via perception (Harnad 1990).
The general thrust of these objections is that people find
it hard to ascribe genuine meaning, of any sort, to dis-
embodied, decontextualized, perceptionless computer pro-
grams. A common response has been to see the key to mean-
ing as lying in a process of perceptually grounding a com-
puter program’s representations in real sensor data (Harnad
1990). While originally formulated as a strategy for imbu-
ing arbitrary internal symbols with meaning, this approach
has been thought to apply straightforwardly to symbols that
link words to the world; indeed, many AI researchers ex-
plicitly advocate achieving linguistic meaning through per-
ceptual grounding (Oates, Schmill, & Cohen 2000; Roy &
Pentland 2002; Cohen et al. 2002; Yu & Ballard 2004;
Steels & Belpaeme 2005).
This paper contests this view. Perceptual grounding, as
it has been understood, is neither necessary nor sufficient to
justify the attribution of linguistic meaning. Human inter-
locutors achieve stable, specific meanings in linguistic com-
munication by coordinating their perceptual experience with
linguistic usage across their community through a process
we call societal grounding. Participating in this process is
a prerequisite for intuitive ascriptions of speaker meaning,
and yet realizing this process robustly would justify saying
a system meant what it said even if it had little perceptual
experience and only modest perceptual accuracy.
Our argument links questions about meaning in imple-
mented systems to the phenomenon of meaning borrow-
ing described in the philosophy of language (Kripke 1972;
Putnam 1975). We use this connection to develop a system-
atic characterization of societal grounding, and then sketch
an architecture for realizing societal grounding within the
information-state approach to dialogue management (Lars-
son & Traum 2000). Our concluding discussions position
computation as a framework that can continue to inform the
study of meaningfulness in machines and in humans.

Appears in AAAI 2006. 2
c
a b
d
(U1) C: it’s a square
(U2) U: you mean a
rectangle?
(U3) C: no
(U4) U: which square?
(U5) C: the square
(U6) U: there is no square
C: [ ignores this ]
(U7) C: the square is solid
Figure 1: User interaction with the COREF agent. The user
(U:) can see the four displayed objects, but not COREF’s
(C:) private labels {a,b,c,d} for them. The target in this
example is object a.
A motivating example
Figure 1 shows an excerpt of an interaction with COREF,
an implemented dialogue agent designed to collaboratively
identify visual objects with human users (DeVault et al.
2005). In this interaction, COREF’s goal is to get the user
to identify object a, the solid red rectangle. COREF begins
by utteringU1, it’s a square — even though the target object
is not a square. In order to diagnose the remainder of this
dialogue, and in pursuit of our goal of meaningful language
use, we would like to understand what COREF means by its
uses of the term ‘square’ in U1, U5, and U7. If it does not
mean what its human users mean when they say ‘square’,
we would like to understand why not.
As it happens, COREF represents the meaning of its uses
of ‘square’ with an internal symbol, square. COREF clas-
sifies any object x as square when
(i) rectangle(x)
& (ii) length(x)width(x) ≤ 1+ ε
(1)
In this case, COREF classifies object a as square, because
a is classified under COREF’s rectangle symbol, and
its dimensions satisfy (ii). Any perceptual classifier would
need a tolerance threshold such as ε to allow for noise in its
length estimates. The exact value ε takes is unimportant; we
caricature COREF’s misclassification in our depiction of a.
We will use this interaction to gradually develop a fine-
grained understanding of the interplay between linguistic
meaning, mental meaning, and system design. For nota-
tion, we will write square for the property that COREF’s
user means by ‘square’, for example in utterances U4 and
U6. The questions we now wish to examine carefully are:
Does COREF mean square when it says ‘square’? If not,
is this because COREF’s internal symbol square does not
mean square? What engineering would be required in order
to make COREF mean square?
Linguistic meaning and speaker’s beliefs
To make our argument for the role of societal grounding in
linguistic meaning, we will examine a series of alternative
conditions that might be thought necessary in order for an
agent A to meanM by a use of linguistic term T . Each condi-
tion can be seen as an attempt to specify under what circum-
stances A could be said to “understand” what is allegedly
meant. For example, according to the historically influential
“description theory of meaning”, competent speakers know,
for each term, some description or set of properties that iso-
lates the term’s meaning; see e.g. (Devitt & Sterelny 1999).
We might endorse this theory with the following condition:
A does not mean M by T unless A associates
T ′s meaning with a set of properties that
uniquely identifies M.
(2)
The properties that COREF associates with the meaning
of ‘square’ in (1) do not uniquely distinguish instances of
square (as evidenced by object a), so the description theory
counsels that COREF does not mean square by ‘square’.
However, in the 1970s, compelling arguments emerged in
the philosophy literature that, contrary to the description the-
ory, a speaker’s meaning in using a term depends not just on
what the speaker privately believes about the term’s meaning
but also strongly depends on perceptual and social factors
(Kripke 1972; Putnam 1975; Burge 1979). From the stand-
point of AI, the most important arguments against the de-
scription theory are what philosophers have called the prob-
lems of ignorance and error (Devitt & Sterelny 1999): peo-
ple mean what they say even when they lack complete and
correct knowledge of meaning. For example, Putnam (1975)
argues persuasively that a human speaker can use the words
‘elm’ and ‘beech’ to mean elm and beech (the two types
of tree) despite being unaware of any (non-linguistic) prop-
erty that distinguishes elm trees from beech trees. When
such a speaker asks of a tree “Is that an elm?”, he still in-
quires meaningfully whether it is an instance of elm. Nor
do erroneous beliefs seem to undermine linguistic meaning;
a human speaker can still mean elm when he says “Is that
an elm?” even if he falsely believes that every tree to which
‘elm’ applies is less than 30 meters tall.
In designing a meaningful agent, we take it as a method-
ological constraint that we ought not impugn the meaning-
fulness of the system on grounds that could also impugn the
meaningfulness of its human users. Thus, (2) is not a gen-
eral principle according to which we can deny the mean-
ing square to COREF’s utterances of ‘square’. The inaccu-
rate properties in (1) that COREF associates with ‘square’
could just be analogous to the properties a meaningful hu-
man speaker erroneously associates with ‘elm’.1
Nevertheless, something is clearly wrong with COREF:
it doesn’t seem to mean square, and this fact seems to ex-
plain the miscommunication in Figure 1. Philosophers have
suggested that the mechanism by which a human speaker
is able to mean elm when he says ‘elm’, despite being un-
able to characterize that meaning accurately, is that there is
a division of linguistic labor within a community (Putnam
1Parallel possibilities of ignorance and error rule out (2) as a
condition on human meaning for mass terms like ‘gold’ or ‘water’
(Kripke 1972; Putnam 1975), medical terms like ‘arthritis’ (Burge
1979), artifact terms like ‘pencil’ (Putnam 1975) or ‘chair’ or ‘sofa’
(Devitt & Sterelny 1999), or for names like ‘John’ or ‘Socrates’
(Kripke 1972), among others.