The Grice Club

Welcome

The Grice Club

The club for all those whose members have no (other) club.

Is Grice the greatest philosopher that ever lived?

Search This Blog

Wednesday, January 9, 2013

Grice on metaphor: You're the cream in my coffee

Speranza
 
Coomentary on V. Bambini & D. Resta, "Metaphor and Experimental Pragmatics:
When Theory Meets Empirical Investigation"

Grice was NEVER interested in METAPHOR. He was interested in PHILOSOPHY, and, as a neo-Plotinian, he KNEW that philosophers, wrongly, have to rely, sometimes, meta-categorially, with METAPHOR, or SIMILE, or ANALOGY. The study of these philosophical modes, since they are so high-sounding, Grice refers to as "PHILOSOPHICAL ESCHATOLOGY".

On the other hand, when in the first lecture to "Logic and Conversation" (Harvard, 1967), he proposed to make fun of a number of philosophers who should have known better as to how to minimise the number of 'senses' of words in the analysis of this or that philosophical claim, he concocted the 'conversational implicature'.

He realised, since he was a classicist (of the Cliftonian school, too boot) that this 'conversational implicature' shared a few features with classical figures of 'rhetoric' -- such as METAPHOR. As the only example of a metaphor, he employs the title of a 1920s song:

The only necessity is you,

whose first line goes

"You're the cream in my coffee".

He analyses this as a categorial mistake, and thus, as implicating, via 'conversational implicature':

"You are my pride and joy".

He noted that, via irony, it could well mean

"You are my bane".

-----

Why "YOU" are the "CREAM IN MY COFEE" is a metaphor, in unpublications, Grice relates to his equally (if not more) genial idea of DISIMPLICATURE.

If with implicature you mean MORE than you say, with DISIMPLICATURE you mean less.

It's like the dropping of entailments (or logical, not pragmatic) implications.

Thus, "You're the cream in my coffee" is best seen as a DISIMPLICATURE -- where the utterer means less than he means.

He realised that people, once used to the idea of conversational implicature, would find trouble in re-organising their thoughts and make room for conversational disimplicature. And he was happy with it!


In this contribution Bambini and Resta focus on one phenomenon that has a special role
in pragmatic theorizing, namely metaphor, and select two issues
deriving from theoretical models and prone to be tested experimentally.


The first issue concerns the comprehension procedure, that is whether
access to metaphorical meaning goes through a mandatory literal stage
and thus is indirect, as predicted by a Griceian inspired account, or
rather is retrieved directly.

The question will be discussed by referring
to behavioral and neurophysiological studies, which advanced our
understanding of the time course of metaphorical interpretation but
proved not fully suitable to answer the question.

The second issue
revolves around the cognitive architecture of the pragmatic system as it
operates in the case of metaphor comprehension. We will illustrate the
contribution that functional neuroimaging, coupled with clinical
investigations can provide to fine-tune the architecture of the system
responsible for metaphor processing. Some outstanding questions are
highlighted in the final part, aiming at sketching our interpretation of
the experimental pragmatic enterprise.


Keywords: metaphor, experimental pragmatics, neuropragmatics,
literal-first hypothesis, mind-reading


Research in pragmatics follows a thread that originated in philosophy,
developed in linguistics and cognitive science, and more recently experienced
an expansion indeed a real turn into experimental investigation. This
introduction briefly traces this evolution, with the aim of presenting the
background upon which metaphor studies have moved and are still moving.
The study of pragmatics finds its origin in the philosophical tradition, and
mostly in the attempt of distinguishing formal language from actual linguistic
usage.

 First references can be traced back to Charles Morris’s semiotic theory
(Morris, 1938), who defined pragmatics as the study of the relations between
signs and their users, in opposition to syntax (the relation of signs to one
another) and semantics (the relation of signs to what they denote). The field
developed in the 1970s after the contribution of philosophers adhering to the
so-called ordinary language philosophy, namely the late Wittgenstein, Austin
(with his famous lecture “How to do things with words”) and Strawson,
highlighting the complexity of meanings and the variety of forms of verbal
communication, and assuming an unbridgeable gap between the semantics of
formal language and that of natural languages.
 
In his William James Lectures at
Harvard in 1967, Herbert Paul Grice claimed that the gap between formal and
natural language could be reduced by way of distinguishing linguistic meaning
from speaker’s meaning (Grice, 1989). He showed that, when considered in
specific contexts of use, linguistic meaning can convey richer and fuzzier
speaker’s meanings, made up not only of “what is said”, but also of “what is
implicated”. In doing so, he introduced new conceptual tools, in particular the
notion of “implicature” that became the foundation for modern pragmatic
theories.
 
One of the merits of Grice is to have shown that pragmatics identifies
a form of behavior, and thus, ultimately, deals with mind faculties. His
Cooperative principle (“Make your conversational contribution such as is
required, at the stage at which it occurs, by the accepted purpose or direction
of the talk exchange in which you are engaged”; Grice, 1989) was intended as
a general law of human rationality. In this frame, meaning became primarily a
psychological phenomenon, and only derivatively a linguistic one. Although it
is widely recognized that Grice intended to offer only a rational reconstruction
rather than a real-time description of pragmatic processes (Bach, 2005), his
work paved the way to a consideration of pragmatics at the interface between
language and cognition.

A full account of pragmatics in psychologically plausible terms developed in
the late 1980s, when pragmatic processes were considered in a mentalistic
perspective, and the term “cognitive pragmatics” started to circulate. Major
issues became the description of comprehension procedure and mechanisms,
and the identification of the pragmatic system with respect to the general
architecture of cognition (Kasher, 1984; Bara & Tirassa, 1999). Upon this
background, Relevance Theory appears as the most full-fledged account of
pragmatic processes, and specifically verbal communication.1 By revisiting the
Gricean maxims in terms of Relevance, i.e., as a function of processing efforts
and contextual effects, and by promoting Relevance to the status of a key
principle in human cognition, Relevance Theory aims at explaining every
possible meaning phenomenon in communication (Sperber & Wilson,
1986/1995). In this framework, “pragmatics is a capacity of the mind, a kind
of information-processing system, a system for interpreting a particular
phenomenon in the word, namely human communicative behavior. It is a
proper object of study itself, no longer to be seen as simply an adjunct to
natural language semantics. The components of the theory are quite different
from those of Gricean and other philosophical descriptions; they include online
cognitive processes, input and output representation, processing effort
and cognitive effects” (Carston, 2002).

Despite Relevance Theory considers pragmatics as a process and not as a
set of abstract and formal relationships (Wilson & Sperber, 2012) and is
concerned with real time processes (Wilson & Sperber, 2004), for a long time
it remained exclusively theoretical, and the same holds true for other pragmatic
hypotheses and research traditions. Theories were tested by using as evidence
a mixture of intuitions about interpretation and observations of behaviors. In
this sense, an experimental approach can strongly help in disentangling
alternative theoretical accounts and their implications regarding underlying
cognitive mechanisms. Psycholinguistics has developed sophisticated
experimental methods in the study of language processing. These were applied
to pragmatic phenomena, but not systematically used to test pragmatic
theories. In the last decade, the publication of the book “Experimental
Pragmatics” (Noveck & Sperber, 2004) marked what became known as the
“experimental turn” in pragmatics (Noveck & Reboul, 2008). Experimental
1 Another model of pragmatic interpretation is Bara’s Cognitive Pragmatics, which accounts
for communicative processes, both linguistic and extralinguistic (i.e., gestural), by positing
special emphasis on the description of the communicative agents’ mental states (Bara, 2010).

pragmatics presents itself as a new field that “draws on pragmatics,
psycholinguistics and also on the psychology of reasoning” (Noveck &
Sperber, 2004) and opens new scenarios in the formulation of testable
predictions derived by theory. More specifically, experimental pragmatics aims
at furthering linguistic and pragmatic theoretical assumptions by better
describing the cognitive factors and mechanisms playing a role in
communicative exchanges and by testing the validity of pragmatic theoretical
proposals for a number of specific phenomena (Noveck & Reboul, 2008).
Developing the experimental side of pragmatics entails deriving testable
pragmatic hypotheses from theoretical assumptions – based on intuitions and
observations – and thus leading theory to refine itself in the light of
experimental evidence.
The experimental side of pragmatics is explored also within the field of
neuropragmatics (Bambini, 2010; Bambini & Bara, 2012), aimed at
describing the neural underpinnings of pragmatic processes by applying, for
example, functional neuroimage. Even though the identification of the aspects
of the pragmatic theory to be addressed in brain’s terms is not definite yet
(Bertuccelli, 2010), most neuropragmatic studies “go neuro” by using
neuroscientific techniques to validate cognitive models and hypotheses. The
importance of this research direction is undeniable, because models are able to
support investigations that otherwise will be only be mere empiricism (Van
Berkum, 2010). However, it should be noticed that the levels of analysis
proposed in the linguistic field may be not always suitable to reflect brain
functioning due to incommensurable units between the linguistic and the
neural level (Poeppel & Embick, 2005; Grimaldi, 2012). Thus, for “going
neuro” it would be desirable also starting from “neuro” hypotheses and
reconsidering research issues with special attention to brain mechanisms and
functioning Van Berkum (2010), which holds for pragmatics as well.
Overall, the experimental turn puts forward a two-way relation between
theory and empirical confirmation. On the one hand, theoretical issues ask for
more detailed experimental evidence and, on the other hand, experimental
evidence puts pressure on theorizing, which in turn leads to refinements and
expansions of theoretical models. As regards the objects under the lens of
experimental pragmatics, several phenomena fit into the paradigm. Some of
them, as for example scalar inference and reference resolution have received
systematic attention (Noveck & Reboul, 2008), whereas for others, as for
example metaphor, experimental evidence is still fragmentary and needs to be

gathered more systematically. In the present work, we will focus precisely on
metaphor, assuming its definition as a “paradigmatically pragmatic
phenomenon” “involving a gap between the conventional meaning of words
and their occasion-specific use”, able to motivate the distinction between
pragmatics and semantics (Camp, 2009), and to approach the psychological
reality of pragmatics as well. We will discuss two research questions that
concern fundamental unit of construction in a pragmatic model such as
Relevance Theory. The first one concerns the comprehension process and,
more precisely, whether there is a mandatory literary step while accessing
metaphorical meanings. The second issue concerns the mental architecture
and how of the pragmatic system operates in the case of metaphor
comprehension cognitive components.
1. Issue 1: Metaphor comprehension: a direct or indirect process?
The assumption that metaphor is a deviation from literal meaning dates back to
Aristotle’s Poetics, where metaphor was considered as a departure from the
“literal norm”. This position has been widely developed within the so-called
standard pragmatic model, associated primarily with philosophers Grice
(1975, 1978) and Searle (1979). Grice assumed metaphor to violate one of
the cooperative principles that govern communication, namely the first Maxim
of Quality, i.e., “Do not say what you believe to be false”, and thus to be
defective in communicative terms compared to literal language. In order to
success in metaphor comprehension, the hearer relies on a set of inferential
rules and pragmatically works out speaker’s meaning, which is separated from
the linguistic meaning. Therefore, part of the conveyed meaning is explicitly
communicated, while another part needs to be inferred (“implicature”).
Moreover, whereas literal meaning is accessed directly, metaphor
interpretation requires subsequent different stages: first, literal meaning is
accessed; second, literal interpretation is detected as defective; and third, the
search for another interpretation is pursued. Accessing literal meaning is thus
a mandatory stage, and metaphor interpretation occurs indirectly. A similar
position is held by Searle (1979), who argued that in understanding a
metaphor the hearer first determines whether seeking for a metaphorical
interpretation or not, then uses a particular set of strategies for computing the
values of the intended metaphorical meaning and, finally, uses a particular
strategy for restricting these values. Interestingly, also the psychologist George
42 Humana.Mente – Issue 23 – December 2012
Miller (1979) argued that metaphor generates a “recognition problem”, i.e., a
discrepancy against the previous context, which is the first step, followed by
reconstruction and interpretation towards metaphor comprehension. Note that
the standard pragmatic model provides a rationalization of the processes used
in deriving figurative meanings but it is in no ways a psychological or cognitive
model. No speculation about the nature or the temporal development of
cognitive processes behind metaphor is provided.
In more recent times, Relevance Theory (Sperber & Wilson, 1986/1995),
within a broad framework for explaining human cognition and communication,
challenged the standard pragmatic model arguing that the same inferential
processes determining implicatures may contribute also to the explicit level of
communication. Metaphorical meaning is claimed to belong to this explicit
level, namely to “what is said”, rather than to “what is implicated” as postulated
by Grice. What guides inferential comprehension is the expectation of
relevance raised by an utterance against the provided context (Sperber &
Wilson, 2008). Inferential processes allow us to automatically “adjust” lexical
concepts and construct ad hoc concepts during online interpretation without
any conscious effort. The concept inferentially derived may have a “broader”
or “narrower” denotation compared to lexically encoded concept. Contrary to
the standard pragmatic model, the relevance-theoretic approach does not
recognize a mandatory literal step in metaphor interpretation. The lexical
encoded meaning is a mere point of access to an array of encyclopedic
information from which the hearer selects in order to achieve a satisfactory
interpretation. Interpretative hypotheses are tested in order of accessibility and
are driven by the achievement of optimal relevance (Wilson & Sperber, 2004).
Turning into the experimental field, these different accounts of the
comprehension process allow for a wide margin for empirical translation of
theoretical hypotheses into time-sensitive predictions. Predictions have been
derived mainly from the standard pragmatic model and were based on the
assumption that literal meanings have absolute priority, and that literal
intended meanings should be easier and faster to understand than figurative
meanings (the so-called “literal first hypothesis”). Longer reaction times for
metaphors were assumed as an indirect support for this hypothesis, assuming
that the process of accessing literal meaning, rejecting it, and searching for a
figurative interpretation requires longer times than directly accessing literal
meaning. In a few words: more stages require more time. This assumption
seemed to be confirmed by behavioral measurements showing that reading
Metaphor and Experimental Pragmatics 43
metaphors is more time consuming than reading literal utterances (Janus &
Bever, 1985). Similar findings, however, were soon challenged by evidence
showing that, when contextually supported, metaphors are comprehended as
quickly as literal counterparts (Gibbs & Gerrig, 1989; Inhoff, Lima, & Carroll,
1984), opening the way to the formulation of an alternative model, known as
“direct access view”.
The dichotomy indirect versus direct access to metaphorical meaning is
mitigated by the Graded Salience Hypothesis (Giora, 2003), a general view of
language comprehension that postulates the activation of salient meaning (i.e.,
the most prominent and faster retrievable from the mental lexicon) in the first
stage of language processing without additional pragmatic interferences and
regardless of context. When literal and non-literal utterances converge in the
degree of salience, the initial process is the same and a direct access occurs. On
the contrary, in the case of novel – non salient – metaphors, the access to
metaphorical meaning is not direct. Gibbs (2001) defined the Graded Salience
Hypothesis as a “hybrid theory of figurative language” because it combines
both a direct access view (for salient meanings) and an indirect access view (for
non salient meanings). The Graded Salience Hypothesis seems to be
confirmed by reading times that showed equal reading times for both salient
literal and metaphorical meanings (Giora, 2003).
Equal reading times, however, are not sufficient per se neither to support
nor to discredit the standard pragmatic model, because several elements may
induce an “incorrect illusion” of equivalence (Gibbs & Gerrig, 1989).
Importantly, McElree and Nordlie (1999) showed that even if the process of
understanding literal and figurative language might be equal in time,
differences in accessibility might be significant. A number of techniques are
able to offer online, more fine-grained data on metaphor understanding
process than those provided by reading times. Among these, the recording of
brain activity through Event-Related Potential (ERP) technique seems
especially promising in disentangling the direct versus indirect dichotomy as it
is able to give millisecond precise temporal pattern of comprehension process
time-locked to the word of interest.
Up to now, about fifteen ERP studies on metaphor have been published and
among them a great variability of metaphorical structures, languages,
experimental protocols, and observed ERP components is visible (among
others, Arzouan, Goldstein, & Faust, 2007; De Grauwe, Swain, Holcomb,
Ditman & Kuperberg, 2010; Lai, Curran, & Menn, 2009; Pynte, Besson,
44 Humana.Mente – Issue 23 – December 2012
Robichon, & Poli, 1996; Tartter, Gomes, & Dubrovsky, 2002). It is beyond
the purpose of the present paper to review all available ERP contributions on
metaphor. What falls within our aims is to overview how the literal-first
question has been addressed in this domain. The dichotomy direct versus
indirect access models seems indeed to be a compulsory step for introducing
ERP studies on figurative language but, actually, only in a few cases precise
experimental predictions are provided and discussed. Among the available
amount of ERP studies on metaphor, we selected two studies that clearly
transposed the theoretical assumptions of the standard pragmatic model into
ERP predictions, in terms of modulation of components: Pynte et al. (1996)
and De Grauwe et al. (2010).
The first one (Pynte et al., 1996) investigated the difference between
familiar and unfamiliar metaphorical sentences both with and without
contextual support. The second one (De Grauwe et al., 2010) investigated the
difference between conventional metaphorical sentences and literal or
anomalous sentences. Both studies predicted a modulation of two components
as neural evidence for the literal first hypothesis. Namely, the N400 – a
negative-going component that peaks at approximately 400 ms after the onset
of the target stimulus – taken as an index of the difficulty in semantic
processing, and the Late Positive Component (LPC) – a positive going
potential following the N400 – taken as a marker of additional processing or
meaning reanalysis. However, due to the variability in adopted protocols,
conclusions proved highly different.
More specifically, Pynte et al. (1996) ran four ERP experiments comparing
familiar metaphors (“Those fighters are lions”), unfamiliar metaphors (“Those
apprentices are lions”) and literal sentences (“Those animals are lions”). The
authors predicted that if the literal-first hypothesis holds, metaphors should
elicit an N400 effect (i.e., a different amplitude compared to literal controls)
indexing the access to literal meaning, followed by an LPC effect indexing the
access to the metaphorical meaning. Otherwise, if the parallel hypothesis (or
direct access) holds, the access to literal and metaphorical meaning should
occur in the same latency band and the N400 should be modulated by factors
that usually influence metaphor comprehension, such as familiarity. Finally,
they introduced a third hypothesis, namely the context dependent-hypothesis
arguing that metaphorical meaning is accessed directly when relevant to the
preceding context. If this last hypothesis is valid, the N400 to metaphor should
be reduced in presence of a supportive context. Results showed that, when
Metaphor and Experimental Pragmatics 45
compared to literal sentences, familiar metaphor elicited a more enhanced
N400, consistently with the literal-first hypothesis (experiment 1). But when
familiar metaphors were compared to unfamiliar metaphor, no N400 was
visible (experiment 2). In this case, N400 proved not sensitive to familiarity
modulation thus disconfirming, according to the predictions, a parallel
hypothesis. When context was added and again familiar and unfamiliar
metaphors were compared, results run as follows. When context was
supportive in the case of familiar metaphors (“They are not talkative: these
counselors are carps”) and unsupportive in the case of unfamiliar metaphors
(“They are not obedient: these engineers are carps”), unfamiliar metaphors
elicited both N400 and LPC effects (experiment 3). The authors claimed that
this result could be explained both in terms of the literal-first hypothesis
because of the biphasic pattern N400-LPC reflecting different stages, and in
terms of the context-dependent hypothesis, because context increased the
difference between familiar and unfamiliar metaphors, not visible in the out of
context condition. But, when context was unsupportive for familiar metaphors
(“They are not naive: Those fighters are lions”) and supportive for unfamiliar
metaphors (“They are not cowardly: this apprentices are lions”), surprisingly
an N400 effect was elicited by familiar metaphors, while an LPC effect was
observed for unfamiliar metaphors. The authors concluded that the crucial
effect of context visible in N400 modulation definitely supports the contextdependent
hypothesis.
De Grauwe et al. (2010) compared conventional nominal metaphors
(“Unemployment is a plague”), literal (“Cholera is a plague”) and semantically
anomalous sentences (“Metal is a plague”). Like in Pynte et al. (1996), here
the authors suggested that if the literal-first hypothesis is valid, an N400 effect
to metaphors should be elicited – indexing a preliminary difficulty in semantic
mapping – followed by an LPC effect reflecting an additional effort in accessing
metaphorical meaning or integrating it in context. Moreover, they took into
account the Graded Salience Hypothesis (Giora, 2003) and predicted that if it
holds no difference in the N400 amplitude should be recorded between
conventional metaphors and literal sentences, as in both cases the critical word
is salient in the context and should be activated immediately. The authors
claimed that an LPC effect might be consistent with the Graded Salience
Hypothesis, as well, because also in familiar metaphors the literal meaning
could be accessed in addition to the metaphorical meaning until considered in
conflict with a plausible interpretation thus eliciting a reanalysis indexed by the
46 Humana.Mente – Issue 23 – December 2012
LPC. Results showed that semantically anomalous phrases elicited a greater
N400 effect than the other two experimental conditions, while the difference
between metaphors and literal counterparts was only transient. The N400,
indexing the access to literal meaning, was followed by an LPC effect triggered
by the conflict between the correct metaphorical meaning and the still retained
literal meaning, in other words by a kind of alignment process. To confirm
these results, the authors designed a second experiment where the critical
word was in mid-sentence position and thus wrap-up effects due to finalsentence
positions were avoided. Importantly, also when presented in midsentence
position, metaphorical critical words elicited an N400 (even if
localized) followed by an LPC effect. Overall, the authors found thus support
for the literal-first hypothesis, while the Graded Salience Hypothesis seemed to
be not supported.
In sum, both reviewed ERP studies confirmed that metaphor evoked a
biphasic pattern (N400-LPC), like other pragmatic phenomena (e.g.
metonymy, Schumacher, 2011) but they differed in interpreting the results
with respect to the direct versus indirect dichotomy: while Pynte et al. (1996)
interpreted their results in terms of a context-dependent, i.e., direct
hypothesis, De Grauwe et al. (2010) claimed to offer support in favor of serial,
i.e., indirect access. Again, the literal-first hypothesis is neither confirmed nor
disconfirmed definitely, as it is not possible to determine whether there is a
mandatory initial stage.
Therefore, the question is whether available predictions and adopted
experimental paradigms are actually suited to ascertain the presence of a literal
step with sufficient accuracy. While an initial step with a fully accessed literal
meaning seems implausible, it is possible that core aspects of the literal
meaning are automatically activated even in metaphorical-biased contexts.
Remarkable evidence in this direction comes from a cross modal priming study
(Rubio Fernandez, 2007). This study showed that core features of word
meaning (i.e., literal meaning) are always activated even in context biased
towards metaphorical interpretation and remain activated beyond the
recovering of the metaphorical meaning. On the contrary, irrelevant features
were deactivated between 400 and 1000 ms through a suppression process
that involves high-level cognitive processes similar to those involved in
ambiguity resolutions. In its turn, also theory is moving towards a convergence
with experimental data by incorporating the idea of a “lingering” of literal
meaning throughout the metaphor comprehension process (Carston, 2010) in
Metaphor and Experimental Pragmatics 47
perfect accordance with experimental pragmatics’ aims. In this scenario, the
question about whether the access to metaphor passes through a mandatory
literal step seems to be rather interpreted as “what is the role of literal
meaning”. Some promising suggestions rely on new paradigms – as for
example the combination of masked priming and ERP recording (Schumacher,
Bambini, Weiland in press) – that are able to tap the very early phases of
processing and could be profitable used for investigating also the construction
of metaphorical meanings.
2. Issue 2: Cognitive components in metaphor comprehension
The second issue we would like to consider is related to the cognitive
architecture of the metaphor comprehension process, i.e., the description of
the cognitive systems that participate in metaphor understanding and of their
neural underpinning. Here the experimental pragmatics perspective blends
with clinical pragmatics and neuropragmatics, as it enters matters and adopts
methodologies belonging to neuropsychology and cognitive neuroscience. To
our view, here too it is possible yet profitable to stick with the guideline of a
strict dialogue between theoretical hypotheses and empirical data.
Going back to Grice, he was certainly far away from approaching the
neurocognitive description of metaphor. However, the assimilation of
metaphor to implicature and the description of how implicatures are derived,
namely through the recognition of intentions, makes it plausible to assume
that, in a Gricean inspired cognitive model, metaphor comprehension – as all
cases of implicated meaning – would require mind-reading operations, i.e.,
what has become known as Theory of Mind, defined as the ability of attributing
mental states to others. The Gricean view is that pragmatic interpretation is
ultimately an exercise in metapsychology, in which the hearer infers the
speaker’s intended meaning from evidence she has provided for this purpose
(Sperber & Wilson, 2002).
In maintaining the inferential nature of pragmatic processes at the implicit
level and extending it to the explicit level of communication, Relevance Theory
too assumes the involvement of mind-reading mechanisms. More specifically,
Relevance Theory considers mind-reading as a dedicated inferential module,
and pragmatics as a sub-module of the mind-reading module, with its own
special-purpose principles and mechanisms, especially dedicated to verbal
communication (Wilson, 2005). The relevance-theoretic comprehension
48 Humana.Mente – Issue 23 – December 2012
procedure automatically constructs hypotheses about the speaker’s meaning
on the basis of a description of the utterance plus available contextual
information. In the case of metaphor, the hypotheses about the speaker’s
meaning cannot be based on the default rules of literalness, being metaphorical
expressions alternative routes to reach optimal relevance, determining
additional efforts in terms of intention recognition (Happé, 1993).
This aspect of the relevance-theoretic perspective found an empirical test
bed in the autistic condition: autistic patients lacking mind-reading abilities
should be impaired in deriving metaphorical meanings as compared to literal
interpretation. The data collected by Happé (1993) supported this view, by
showing that autistic children who do not pass the first-order theory of mind
test are impaired in metaphor comprehension but not in simile
comprehension, as the latter condition can be interpreted literally. Several
studies confirmed this piece of evidence, although it has been showed that
Theory of Mind is necessary but not sufficient to understand metaphor
(Norbury, 2005), the specificity of pragmatic disorders in autism is a matter of
current debate (Giora, 2012). Recently, Wearing (2010) highlighted the
apparent conflict between the idea of metaphor as loose use, based on ad hoc
concepts and not involving special mechanisms, and the special mind-reading
effort assumed to be necessary for metaphor interpretation. The conflict is
solved by clarifying that the mind-reading effort in metaphor processing lies in
the type of source the hearer must rely upon to derive the ad hoc concept and
solve the interpretive process successfully: not only the world at large, as
typically in the literal case, but the beliefs that the speaker is likely to hold and
not to hold. The relevance of a metaphor is guaranteed by the speaker’s mental
states, rather than by the world, which is what autistic individuals are unable to
process.
Importantly, as evident from the description of the relevance-theoretic
comprehension procedure, mind-reading mechanisms operate against a wider
cognitive background which includes other systems. As already mentioned, the
comprehension procedure follows a path of least effort in formulating
hypotheses about the speaker’s meaning, aiming at the maximization of
relevance and “using whatever contextual information is most highly activated
by the automatic workings of the cognitive system at the time” (Wilson, 2005).
The maximization of relevance is guaranteed by three notably characteristics of
the human cognition: the constant monitoring of the environmental features,
the permanent availability of a huge amount of memorized data, and the
Metaphor and Experimental Pragmatics 49
attentional resources, which are able to handle only a limited amount of
information at any given time. The efficiency of the process is a matter of being
able to select, from the environment and from memory, the most relevant
information for attentional processing (Sperber & Wilson, 2002).
On this basis, if we could take a global picture of the cognitive systems at
work in pragmatic interpretation, one should expect not only mind-reading
mechanisms, but also the attentional system to be involved, possibly with a
higher effort for metaphor as compared to literal interpretation, related to
higher costs in selecting the relevant information from context and possibly
also the appropriate meaning among competing ones. Moreover, the
construction of the ad hoc concepts – accessing concepts and modulating their
denotations through broadening and narrowing – implies operations that
should be reflected in the conceptual system.
In a similar scenario, functional neuroimaging techniques appear as the
methodology of choice for exploring the cognitive architecture of metaphor
comprehension, being able to provide information concerning “where” the
processes take place, and allowing for anatomo-functional correlations
between brain structures and cognitive systems. The neuroimaging of
metaphor comprehension is relatively vast, including around twenty studies
published in the last decade. Metaphor nicely suits into the requirements of
standard neuroimaging paradigms, i.e., subtractive: while for structural aspects
of language we need to device sound experimental designs, for instance
introducing anomalies (Moro, 2008), metaphor can be easily compared to its
literal equivalent, offering a window on what systems are recruited to adjust
meaning pragmatically. However, most of the existing literature has not paid
attention on describing the architecture of the process, being interested
instead in assessing hemispheric involvement, following the classic hypothesis
that pragmatic aspects of language are processed in the right hemisphere
(Joanette, Goulet, & Hannequin, 1990; Tompkins, 1995). In contrast with
the classic view, most studies on metaphor interpretation report bilateral
patterns of activations, where frontal and temporal areas stand out in
particular, as shown in recent comparative analyses (Bohrn, Altmann, &
Jakobs, 2012; Rapp, Mutschler, & Erb, 2012). The involvement of the right
hemisphere seems to be especially modulated by the conventionality vs. novelty
of the metaphorical expressions, with greater right activations involvement for
novel and unfamiliar expressions (Schmidt et al., 2010). This is in line with the
theoretical tenets of Graded Salience Hypothesis, which specifically aims at
50 Humana.Mente – Issue 23 – December 2012
accounting for the difference between familiar and unfamiliar expressions,
assuming two different processing styles (Bohrn et al., 2012). The cognitive
architecture of the comprehension process, however, is not the main topic in
the neuroimaging of metaphor.
In a recent fMRI study, we tried to overcome the right versus left debate to
focus on the neurofunctional description of metaphor comprehension, based
on the cognitive architecture put forward in the cognitive pragmatics literature
(Bambini et al., 2011). The building blocks of the metaphor comprehension
process assumed were the following:
a) To start with, we hypothesized the crucial involvement of the conceptual
component , reflecting conceptual access and the construction of ad hoc
concepts through context-based inferences.
b) Second, we hypothesized the involvement of the attentional component ,
in charge of selecting the information worth bring together, filtering
contextual features and competing meanings.
c) Finally, the mind-reading component is assumed to participate in
supporting the recognition of the speaker’s communicative intentions, and
specifically, along Wearing (2010), the recognition of the speaker’s belief on
which ad hoc concepts are based.
We constructed an experimental paradigm where participants were
presented with paired passages including literal and metaphorical expressions
(e.g., “Do you know what that insect is? A dragonfly” vs. “Do you know what
that dancer is? A dragonfly”), intermixed with fillers, in order to reduce the
proportion of figurative language. Participants were instructed to read the
passages, and then perform an adjective matching task, making metaphor
comprehension an implicit task. The analysis showed that metaphor as
compared to literal comprehension produced greater activations in a number
of regions, distributed bilaterally and involving especially the frontal and
temporal lobes: the bilateral inferior frontal gyrus (extending over left and right
BA 45 and left BA 47) and other prefrontal regions (right BA 9 and left BA 8),
the left angular gyrus (BA 39), the cingulated cortex bilaterally (BA 24 and BA
32) and the right posterior superior temporal gyrus (BA 22).
Metaphor and Experimental Pragmatics 51
Figure 1. Brain areas activated for metaphor comprehension as compared to literal sentences. Coronal
(on the left) and sagittal view of the left hemisphere (center) and of the right hemisphere (right); the
lines in the coronal image correspond to the location of the sagittal slices. AC: anterior cingulate; IFG:
inferior frontal gyrus; MidFG: middle frontal gyrus; STS: superior temporal sulcus; AG: angular gyrus
Adapted from Bambini et al. (2011).
The anatomo-functional correlations seemed to confirm our predictions:
a) Activity at the level of the inferior frontal gyrus bilaterally is likely to
reflect the activation of the conceptual system. The activation of bi-frontal
areas is probably the most robust results across the literature on metaphor, and
in discourse processing as well, i.e., when word meaning needs to be
integrated with world knowledge and the wider context. The same function
seems to be supported by the angular gyrus (BA 39), which too is greater
activated in metaphor as compared to literal process, and which is considered
as an area supporting high-level conceptual processing.
b) The activity observed in the cingulate cortex, as well as activity in
prefrontal areas, is likely to reflect attentional mechanisms . These regions are
implicated in cognitively demanding tasks involving stimulus-response
selection in the face of competing streams of information, including Color
Stroop and Stroop-like tasks and many working memory tasks.
c) Finally, the posterior part of the right superior temporal gyrus might
reflect mind-reading mechanisms , as this area has been implicated in
monitoring the protagonists’ perspective and attributing intentions to agents.
Interestingly, in our study the regions along the superior temporal sulcus show
greater activity in response to unfamiliar than familiar metaphor, extending to
the left hemisphere for unfamiliar items, pointing to a strong relation between
novelty and mind-reading efforts.
Two important considerations should be added here. First, this
decomposition is highly compatible with the clinical literature on metaphor
deficit, which is not limited to the autistic population. Deficit in metaphor
52 Humana.Mente – Issue 23 – December 2012
comprehension are indeed vulnerable to several neurological and psychiatric
conditions (Thoma & Daum, 2006; Rapp & Wild, 2011), which suggests that
a complex cognitive architecture is involved: the underlying cause of deficits in
pragmatic interpretation, far from being unitary, might find different
explanations in different populations (Stemmer, 2008). For instance,
difficulties in demented patients might be related to attentional deficits in the
case of demented population (Amanzio et al., 2008). Second, this
decomposition shares similarities with the findings reported for other cases of
figurative language interpretation, for instance idiom processing (Papagno &
Romero Lauro, 2010) which suggests common mechanisms and is in line with
the Relevance Theory’s idea that metaphor doesn’t require specific operations
to be processed.
Overall, the main achievement of our study is the neutrally-plausible
decomposition of the metaphor comprehension process in a network of
functional components, which are candidate to represent basic blocks in a fullfledged
neurocognitive model of metaphor processing. Cognitive pragmatics
provided the foundation for the study, by helping formulating predictions and
interpreting results. To this respect, Relevance Theory seems to offer a
comprehensive framework for understanding figurative language
comprehension, by sketching a model which is grounded in general cognition
and takes into account different aspects of the process. Besides, Relevance
Theory might hopefully bear on neuroimaging evidence to further detail the
architecture of the system. Similar considerations, however, are not limited to
the Relevance Theory framework, but extend to other cognitive models of
metaphor comprehension: we will go back to this point in the final remarks.
One can expect that major future achievement could come from the attempt to
go beyond the decomposition toward a proper neurocognitive model: much
promises lie in the study of connectivity between brain areas, in order to
explore the delicate interplay of the components, the order of activation, and
their specific roles.
Conclusion
In this contribution, we attempted to show how theoretical assumptions in
pragmatics can be translated into experimental paradigms, and indeed can
profitably guide empirical investigations in formulating predictions and
interpreting results. As emerged in the discussion, the process is far from
Metaphor and Experimental Pragmatics 53
being complete. Rather, the experimental turn in metaphor research is in the
phase of sharpening research questions and experimental paradigms for closer
addressing key theoretical points.
One of these central questions regards the description of the time course of
metaphor interpretation, bearing in mind the standard pragmatic view and the
hypothesis of a mandatory initial literal step. Through the consideration of
behavioral data and ERP results collected over the last three decades, we
argued that available experimental data are still not able to give a decisive
answer to this question. Promising hints seem to come from a refinement of the
theory, incorporating the notion of lingering of the literal meaning, and from
the employment of new paradigms able to tap the very early stages of
processing (e.g., masked priming combined with ERP).
Another case is represented by the cognitive architecture of the metaphor
comprehension process. The involvement of mind-reading mechanisms, since
long assumed in the pragmatic tradition, is confirmed by clinical evidence on
autistic patients. Moreover, functional neuroimaging helped decomposing the
process of metaphor interpretation in a number of cognitive components
which include also the conceptual and the attentional system. Guided by
theoretical modeling, we are stepping forward in describing the
neurofunctional architecture of metaphor interpretation.
A framework such as Relevance Theory – explicitly modeled on online
processes of utterance interpretation and the nature of the systems behind
(Wilson & Sperber, 2004) – offers many more aspects to be addressed
experimentally. We would like to mention just two of them, which are of
special interest to our view and on which our group is working within an
experimental (neuro)pragmatics perspective.
First, recent developments in the relevance-theoretic account of metaphor
suggest that the comprehension of creative and extended metaphors might
imply a meta-representative process of extracting the intended meaning from
both literal meaning and the evoked imagery rather than a process of ad hoc
concept construction (Carston & Wearing, 2011). To now, very little
experimental evidence has been collected on literary metaphor, which might
represent a case very high processing costs are compensated by high cognitive
benefits, possibly in terms of aesthetic appreciation. Early results showed that
literary metaphor is a complex, multidimensional phenomenon whose
comprehension is influenced and mediated by a number of psycholinguistic
variables (Resta, Bambini, & Grimaldi, submitted). This domain appears to us
54 Humana.Mente – Issue 23 – December 2012
as a good candidate for a rich dialogue between experimental evidence and
theory, pragmatics and possibly also cognitive poetics.
Second, figurative language has been used in pragmatics as a label for
several different phenomena, and only recently theory is starting to clarify the
differences across phenomena, by developing a finely-grained lexical
pragmatics account. Although we assume that all figurative instances require
pragmatic inferences, there might be different underlying processes linked to
different operations at conceptual level. For example, there is early evidence
supporting a distinction between metaphor, metonymy, and approximation –
which are claimed to vary in the degree of underlying adjustment – in terms of
the interpretation availability and costs (Bambini, Ghio, & Schumacher,
submitted). We believe that along this line a psychologically grounded
taxonomy of figurative language might be reached, which might account for the
alternative routes speakers might choice to communicate meaning.
Besides Relevance Theory, interesting suggestions might come from other
theoretical proposals which move in different fields than pragmatics. Above all,
Cognitive Linguistics – grown out from the work of Lakoff and colleagues –
suggested that metaphor is not a specific linguistic device, but a conceptual
phenomenon deriving from bodily grounded mapping operations (Lakoff &
Johnson, 1980). Interesting experimental predictions about the involvement
of the sensory motor system in the processing of metaphor meaning might be
derived in this scenario, which are being elaborated but still require further
exploration. Furthermore, the mechanisms of mapping might play a role in
understanding some types of metaphor, although the theory is still
underdetermined in terms of comprehension procedure. Importantly,
Cognitive Linguistics appears to be not incompatible with Relevance Theory’s
main claims and the possibility of a combination of the two to contribute to a
comprehensive theory of metaphor has been already explored, as they might
target different aspects and thus be complementary (Tendahl & Gibbs, 2008;
Wilson, 2011). In this light, experimental pragmatics might be at the forefront
in solving theoretical disputes within the wider perspective of sharpening
theory to account for experimental evidence and allowing experimental
evidence to sharpen theory in its turn.
Metaphor and Experimental Pragmatics 55
ACKNOWLEDGEMENTS
Although the paper reflects opinions of both authors, V. Bambini is
mainly responsible for sections I and III, while D. Resta for section II.
Part of the contents reported in the paper was discussed as preparatory
work of a research project funded by the European Science Foundation’s
EURO-XPRAG Program and run in collaboration with Petra B.
Schumacher and Marta Ghio. Discussions with them are gratefully
acknowledged. We would also like to thank Francesca Ervas and
Elisabetta Gola for giving us the opportunity to contribute to this Special
Issue.

REFERENCES


Amanzio, M., Geminiani, G., Leotta, D., Cappa, S. (2008). Metaphor comprehension
in Alzheimer’s disease: novelty matters. Brain and Language , 107, 1–10.
Arzouan, Y., Goldstein, A., Faust, M. (2007). Brainwaves are stethoscopes: ERP
correlates of novel metaphor comprehension. Brain Research, 60, 69–81.
Bach, K. (2005). The Top 10 Misconceptions about Implicature. In B. Birner, G.
Ward (Eds.), A Festschrift for Larry Horn, Amsterdam: John Benjamins.
Bambini, V. (2010). Neuropragmatics: A foreword. Italian Journal of
Linguistics/Rivista di Linguistica , 22, 1–20.
Bambini, V., Bara, B.G. (2012). Neuropragmatics. In J.-O. Ă–stman J. Verschueren
(eds.), Handbook of Pragmatics, Amsterdam/Philadelphia: John Benjamins.
Bambini, V., Gentili, C., Ricciardi, E., Bertinetto, P. M., Pietrini, P. (2011).
Decomposing metaphor processing at the cognitive and neural level through
functional magnetic resonance imaging. Brain Research Bulletin, 86, 203–
216.
Bambini, V., Ghio, M., Schumacher P. (submitted). Differentiating among pragmatic
uses of words through timed sensicality judgments: Metaphor, metonymy and
approximation.
Bara, B.G. (2010). Cognitive Pragmatics. Cambridge (MA), MIT Press.
Bara, B.G., Tirassa, M. (1999). A mentalist framework for linguistic and
extralinguistic communication. Proceedings of the 3rd European Conference
56 Humana.Mente – Issue 23 – December 2012
on Cognitive Science (ECCS '99 — Siena, Italy, 27-30 October 1999),
Bagnara, S. (ed.), Roma, Istituto di Psicologia del Consiglio Nazionale delle
Ricerche, 285–290.
Bertuccelli Papi, M. (2010). How does pragmatics fit with the brain? New challenges
from complex systems theories. Italian Journal of Linguistics/Rivista di
Linguistica, 22, 209–228.
Bohrn, I. C., Altmann, U., Jacobs, A.M. (2012). Looking at the brains behind
figurative language – A quantitative meta-analysis of neuroimaging studies on
metaphor, idiom, and irony processing. Neuropsychologia, 50, 2669-2683.
Camp, E. (2009). Metaphor. In L. Cummings (Ed.), The Pragmatics Encyclopedia ,
Oxford: Routledge.
Carston, R. (2002). Linguistics meaning, communicated meaning and cognitive
pragmatics. Mind and Language, 17, 127-148.
Carston, R. (2010). XIII-Metaphor: Ad hoc concepts, literal meaning and mental
images. Proceedings of the Aristotelian Society (Hardback) , 110, 295-321.
Carston, R., Wearing, C. (2011). Metaphor, hyperbole and simile: A pragmatic
approach. Language and Cognition, 2, 283-312.
De Grauwe, S., Swain, A., Holcomb, P. J., Ditman, T., Kuperberg, G. R. (2010).
Electrophysiological insights into the processing of nominal metaphors.
Neuropsychologia , 48, 1965-1984.
Gibbs, R.W. (2001). Evaluating Contemporary Models of Figurative Language
Understanding. Metaphor and Symbol, 16, 317-333.
Gibbs, R. W., Gerrig, R. J. (1989). How context makes metaphor comprehension
seem special. Metaphor and Symbolic Activity , 4, 145-158.
Giora, R. (2003) On our Mind: Salience, Context, and Figurative Language . New
York: Oxford University Press.
Giora, R. (2012). Introduction: Different? Not Different? Metaphor and Symbol ,
27, 1-3.
*******************************************************************
Grice, H.P. (1938). Negation. The Grice Papers, UC/Berkeley, Bancroft Library.
(1975). Logic and conversation. In P. Cole J. Morgan (Eds.), Syntax and
Semantics, Volume 3. New York: Academic Press, 41-58.
Grice, H.P. (1978). Further notes on logic and conversation. In P. Cole (Ed.), Syntax
and Semantics, Volume 9. New York: Academic Press, 113-127.
Grice, H.P. (1989). Studies in the way of words. Cambridge: Harvard University Press.

***************************************

Grimaldi, M. (2012). Toward a neural theory of language: Old issues and new
perspectives. Journal of Neurolinguistics , 25, 1-24.
Happé, F. (1993) Communicative competence and theory of mind in autism: a test of
Relevance Theory. Cognition, 48, 101-119.
Horn, L. R. A brief history of negation.
Inhoff, A. W., Lima, S. D., Carroll, P. J. (1984). Contextual effects on metaphor
comprehension in reading. Memory & Cognition, 12, 558-567.
Janus, R. A., Bever, T. G. (1985). Processing of metaphoric language: An
investigation of the three-stage model of metaphor comprehension. Journal of
Psycholinguistic Research, 14, 473-487.
Joanette, Y., P. Goulet and D. Hannequin. (1990). Right Hemisphere and Verbal
Communication. New York: Springer-Verlag.
Kasher, A. (1984). On the psychological reality of pragmatics. Journal of Pragmatics,
8, 539-557. [Reprinted as Pragmatics and the modularity of mind. Davis, S.
(ed.) 1991. Pragmatics: A Reader. Oxford, Oxford University Press. 567-
582.]
Lai, V. T., Curran, T., Menn, L. (2009). Comprehending conventional and novel
metaphors: An ERP study. Brain Research, 1284, 145-155.
Lakoff, G., Johnson, M. (1980). Metaphors We Live By . Chicago: University of
Chicago Press.
McElree, B., Nordlie, J. (1999). Literal and figurative interpretations are computed in
equal time. Psychonomic Bulletin & Review, 6, 486-494.
Miller, G.A. (1979). “Images and models, similes and metaphors”. In Metaphor and
Thought, Ortony, A. (ed.), New York, Cambridge University Press, 202-250.
Moro, A. (2008). Boundaries of Babel. The Brain and the Enigma of Impossible
Languages. Cambridge, Mass.: MIT Press.
Morris, C. (1938). Foundations of the Theory of Signs . Chicago: Chicago University
Press.
Norbury, C.F. (2005). The relationship between Theory of Mind and metaphor:
Evidence from children with language impairment and autistic spectrum
disorder. British Journal of Developmental Psychology , 23, 383-399.
Noveck, I., Reboul, A. (2008). Experimental pragmatics: a Gricean turn in the study
of language. Trends in Cognitive Sciences , 12, 425-431.
Noveck I., Sperber, D. eds. (2004) Experimental Pragmatics. San Diego: Palgrave
MacMillan.
58 Humana.Mente – Issue 23 – December 2012
Papagno, C., Romero Lauro, L. (2010) The neural basis of idiom processing:
Neuropsychological, neurophysiological and neuroimaging evidence. Italian
Journal of Linguistics/Rivista di Linguistica , 22, 21-40.
Poeppel, D., Embick, D. (2005). The relation between linguistics and neuroscience.
In A. Cutler (Ed.), Twenty-first century psycholinguistics: Four cornerstones .
New Jersey: Lawrence Erlbaum, 1-16.
Pynte, J., Besson, M., Robichon, F.-H., Poli, J. (1996). The time-course of metaphor
comprehension : an event-related potential study. Brain and Language , 316,
293-316.
Rapp, A. M., Mutschler, D. E., Erb, M. (2012). Where in the brain is nonliteral
language? A coordinate-based meta-analysis of functional magnetic resonance
imaging studies. NeuroImage, 63, 600-610.
Rapp, A. M., Wild, B. (2011). Nonliteral language in Alzheimer dementia: A review.
Journal of the International Neuropsychological Society, 17, 207-218.
Resta, D., Bambini, V., Grimaldi, M. (submitted). What counts in literary metaphor
comprehension? A normative study in metaphorical expressions from Italian
literature.
Rubio Fernández, P. (2007). Suppression in metaphor interpretation: differences
between meaning selection and meaning construction. Journal of Semantics,
24, 345-371.
Schmidt, G. L., Kranjec, A., Cardillo, E. R., Chatterjee, A. (2010). Beyond laterality:
a critical assessment of research on the neural basis of metaphor. Journal of the
International Neuropsychological Society, 16, 1-5.
Schumacher, P. B. (2011). The hepatitis called …: Electrophysiological Evidence for
Enriched Composition. In J. Meibauer M. Steinbach (Eds.). Experimental
Pragmatics/Semantics, Amsterdam/Philadelphia: John Benjamins, 199-219.
Schumacher, P., Bambini, V., Weiland, H. (in press). Event-related brain potentials of
masked repetition and semantic priming while listening to sentences.
Neuroscience Letters. doi: 10.1016/j.neulet.2012.09.057
Searle, J. (1979). Metaphor. In A. Ortony, A. (Ed.), Metaphor and Thought,
Cambridge: Cambridge University Press, 92-123.
Speranza, Join the Grice Club.
Sperber, D., Wilson, D. (1986/1995). Relevance: Communication and Cognition.
Oxford: Blackwell.
Sperber, D., Wilson, D. (2002). Pragmatics, Modularity and Mind-reading. Mind &
Language, 17, 3-23.
Metaphor and Experimental Pragmatics 59
Sperber, D., Wilson, D. (2008). A Deflationary Account of Metaphor. In R.W. Gibbs
(Ed.), The Handbook of Metaphor and Thought, Cambridge: Cambridge
University Press, 171-203.
Stemmer, B. (2008) Neuropragmatics, Disorders and Neural Systems. In B. Stemmer
H.A. Whitaker (Eds.),Handbook of the Neuroscience of Language , Cambridge
(MA), Elsevier, 367-379.
Tartter, V., Gomes, H., Dubrovsky, B. (2002). Novel metaphors appear anomalous at
least momentarily: Evidence from N400. Brain and Language, 80, 488-509.
Tendahl, M., Gibbs, R. W. (2008). Complementary perspectives on metaphor:
Cognitive linguistics and relevance theory. Journal of Pragmatics, 40, 1823-
1864.
Thoma, P., Daum, I. (2006). Neurocognitive mechanisms of figurative language
processing - Evidence from clinical dysfunctions. Biobehavioral Reviews, 30,
1182-1205.
Tompkins, C.A. (1995). Right Hemisphere Communication Disorders: Theory and
Management. San Diego: Singular.
Van Berkum, J.J.A. (2010). The brain is a prediction machine that cares about good
and bad - Any implications for neuropragmatics?. Italian Journal of
Linguistics/Rivista di Linguistica , 22, 181-208.
Wearing, C. (2010). Autism, metaphor and relevance theory. Mind and Language, 25,
196-216.
Wilson, D. (2005). New directions for research on pragmatics and modularity.
Lingua, 115, 1129-1146.
Wilson, D. (2011). Parallels and differences in Relevance Theory and cognitive
linguistics. Studia Linguistica , 128, 195-213.
Wilson, D., Sperber, D. (2004) Relevance Theory. In L. Horn Ward G. (Eds.), The
Handbook of Pragmatics, Oxford: Blackwell, 607-632.
Wilson, D., Sperber, D. (2012). Meaning and Relevance. Cambridge: Cambridge
University Press

Automated Grice

Speranza

Commentary on E. Gola, S. Federici, N. Ruimy, and J. Wade, "Automated Translation Between Lexicon and Corpora"

Grice hated computers and anything to do with automata.

The reason was simple. When he was given a word processor, he noted that the word processor could not process his neologism (of sorts): pirot. The word processor would systematically correct that to "parrot".

Similarly with "sticky wicket".

"That's enough!", he thought.

His anti-computational thoughts he shared with Haugeland, of Berkeley, and they would discuss anti-automatism (i.e. the negation of Kantian freedom) to no end.

In this paper, the authors argue for the relevance of the lexicon in the translation process and the need to dispose of wide coverage and high quality lexical resources. The access to a rich lexical knowledge is in fact a fundamental requirement for a computational system to correctly analyze a text and generate its translation.
To this purpose, we will present an Italian lexicon that meets the requirements of MT systems, and we will show how its lexical information can be used in a translation process.
It must, however, be emphasized that building a large coverage lexicon is a very costly and time consuming process. That is the reason why Computational Lexicography is today mostly oriented toward the development of methodologies and strategies that make the creation of lexicons easier and faster with the automatic acquisition of data from corpora, from the Web, or by induction from existing resources. In this paper, we will show a bootstrapping method, based on a machine learning technique, that allows us to build at the same time a corpus-based lexicon and a tagged corpus, that grow incrementally together in a semi-automated way.
1. Machine Translation: historical overview and state of the art
Since the beginning of Artificial Intelligence (AI) and Natural Language Processing (NLP) , studies and research were devoted to realizing the dream of Machine Translation.
During the first decades of MT research, an articulated panorama of methodologies and strategies started shaping. Classifying all the approaches is almost impossible, given that perspectives change along with the adopted parameters.
There is now a variety of MT systems which almost defies any neat classification. It is still often legitimate to apply the labels of the 1960s: practical vs. theoretical, empirical vs. perfectionist, direct vs. indirect, interlingual and transfer. But now there are new labels and new perspectives: interactive vs. fully automatic, ‘try-anything’ systems vs. ‘restricted language’ systems, mainframe systems vs. microcomputer or word-processor systems, AI-based systems vs. linguistics-oriented systems (Hutchins, 1986, p. 19).
Automated Translation between Lexicon and Corpora 63
To our purposes, we will focus on the distinction between direct and indirect strategies that belong respectively to first and second generation MT systems.
Until the sixties, MT systems, called first generation systems, followed a so-called direct strategy, in which a direct correspondence was established between Source Language and Target Language (henceforth, SL and TL). In this strategy, the SL was only analyzed from a morphological point of view. The output of the morphological analysis constituted the access point to the bilingual lexicon. In this way, a text could only be translated word-by-word. This strategy failed therefore to cope with the translation of ambiguous sentences or sentences with different SL and TL syntactic structures, such as the Italian sentence Questo ragazzo piace a Maria (lit. this boy likes to Maria), whose English structure: Maria likes this boy is quite different.
During the sixties, the second generation systems adopted an indirect strategy, in which two approaches were followed. Firstly, a two-phase process defined as the Interlingua approach and, secondly, a three-phase process defined as the Transfer approach.
In the Interlingua approach, a formal, abstract and language-independent representation interfaces source and target languages: a SL text is analyzed into an interlingual representation which is then synthesized into a TL text. In this view, a conceptual lexicon is required, the building of which is an extremely complex and controversial task.
For this reason, more realistic strategies, based on Transfer, are adopted. In this case, the translation steps are the following:
 analysis of a SL text into a SL formal representation;
 transfer of the SL formal representation into a TL formal representation;
 generation of a TL text from the TL formal representation.
In the Transfer approach, the structural analysis of the SL text is performed in different steps and leads to the building of a formal representation of the SL structures that, in the transfer phase, is mapped onto a formal representation of the TL structures. As to the lexical transfer, the SL lexical units are translated into TL lexical units, using an electronic bilingual dictionary. During the
64 Humana.Mente – Issue 23 – December 2012
synthesis phase, the TL formal representation is turned, following different steps, into a TL text. In this perspective, cultural aspects of different languages are taken into account.
In spite of this innovation, disappointment with the feasibility of MT was growing, due to the “semantic barriers” that researchers encountered and that proved difficult to overcome.
Furthermore, in 1964, the US government sponsors asked the National Science Foundation to constitute a committee in order to evaluate the progress made in NLP in general and in the MT state-of-the-art in particular. The commission produced in 1966 a “(in)famous report”, as John Hutchings (1996) defined it, the ALPAC report (from the name of the committee: Automatic Language Processing Advisory Committee). The ALPAC report stated that MT systems were slower, less precise and more expensive than human translators. The verdict then was: “there is no immediate or predictable prospect of useful machine translation” (ALPAC, 1966).
It should be noticed, though, that the ALPAC committee, in its report, took into consideration only direct strategy systems, evaluating them negatively. For the next ten years, this assessment caused the U.S. Government to reduce its funding in this area dramatically. As a direct consequence, research in this field stopped in the US for over a decade, while it carried on in Canada, Germany and France.
It is only in the middle of the Seventies that we find a renewed interest for automated translation, with the emergence of third generation systems based on Artificial Intelligence.
Starting from the 1990s, a new methodological approach emerges, that makes use of large bodies of text (corpora) (Hunston, 2002). Among the corpus-based systems, the most common approaches are statistics-based systems (SBMT) and example-based systems (EBMT).
SBMT follows strategies in which SL and TL sentences are tentatively aligned on the basis of the probability that each word in the SL sentence corresponds to one or more words in the TL sentence. On the contrary, the example-based methodology, suggested by Nagao in 1984 but implemented only in the 1990s, gives a translation by analogy, comparing the input sentence with a bilingual dictionary that includes examples and matching those that are more similar to the input (Nagao, 1984; Brown, 1999, Turcato et al., 1999).
In the same years, the rule-based systems move away from syntax-based representations to more 'lexicalist' approaches. At its extreme, the essence of
Automated Translation between Lexicon and Corpora 65
the lexicalist approach in MT system design is to reduce transfer rules to simple bilingual lexical equivalences. Such a drastic reduction was first put forward in the CRITTER project (Isabelle et al., 1988). The approach has been explored in the ACQUILEX project devoted primarily to the construction of multilingual lexicons for transfer-based MT (Sanfilippo et al., 1992), and is probably best known as the 'shake-and-bake' method described by Whitelock (1992). The requirement for structural representations - common to both transfer and interlingua approaches - is abandoned in favour of sets of semantic and syntactic constraints on lexical items. Translation involves the identification of TL lexical items which satisfy the semantic constraints attached to the SL lexical equivalents.
The 'bag' of target lexical items is then 'shaken' to generate an output text consistent with the syntax and semantics of the target language (Hutchins, 1993).
This ‘lexicalist’ turn led the MT community to an increasing interest for computational lexicons.
Today, Machine Translation systems usually follow either a corpus-based or a rule-based approach. In the first trend, we find statistical approaches and example-based approaches. In the second one, emphasis is given to lexical resources. In the following section of the paper, we will propose an integration of these two approaches.
2. Relevance of the lexicon in MT
In order to produce a good translation it is necessary to understand correctly the input text. It is precisely for this reason that Machine Translation is deemed one of the most difficult tasks in the field of AI language applications. Any translation process implies, in fact, the resolution of a whole range of problems regarding both the analysis and the generation of texts. In this context, the lexicon plays a crucial role. A robust translation system should be able to cope with a wide range of issues inherent to the complexity of natural language, such as the various types of ambiguity, non literal uses, polysemy and so on. A poor lexicon fails to support these challenging tasks.
66 Humana.Mente – Issue 23 – December 2012
3. Lexicon and lexical problems in MT
Table 1 illustrates some of the most typical and frequent lexical problems that are encountered during a translation process and that a lexicon tailored for an MT system should be able to deal with.
The lexicons used in MT systems must have wide coverage and provide, for each lexical entry, a large range of rich and various information spanning all levels of linguistic description.
Direct strategy MT systems used a unique, very complex bilingual lexicon containing all grammatical information concerning both the SL and the TL lexical units, as well as the conditions for selecting the appropriate translation in case there are different alternatives possible.
Transfer-based MT systems, by contrast, use different monolingual lexicons (morphological, syntactic and semantic) containing all relevant information for each level of linguistic description for both the analysis and generation phases. In the transfer phase a bilingual lexicon is used. The transfer bilingual lexicon consists of lexical rules setting i) the correspondences between the lexical units described in the SL and TL monolingual semantic lexicons and ii) the conditions imposed on those equivalences. For example, in case of a SL word translatable by different TL words, the lexical transfer rule selects the appropriate TL equivalent, on the basis of the information provided by the two translational equivalents in their respective monolingual description.
In the domain of computational lexicography, a significant number of electronic lexical resources are now available, even though not all languages are equally represented. Most lexicons deal with a single level of linguistic description; some describe a unique part of speech or are strictly theory-dependent. Some are created in order to describe the vocabulary of a particular domain; others in order to meet the requirements of a specific application.
Automated Translation between Lexicon and Corpora 67
Level
Phenomenon
Example
Phonology
Homography
 it. pĂ©sca = en. fishing
 it. pèsca = en. peach
Morphology
Homonymy
 it. legge, porta, sbarra: N & V
 it. appunto: N. & ADV
Syntax
Syntagmatic realization
 en. know + NP = it. conoscere
 en. know + WH-clause = sapere
Semantics
Homonymy
 fr. louer = en. to praise
 fr. louer = en. to rent
Polysemy
 en. set up = it. piantare, erigere,
mettere su, causare, installare, allestire, formare, etc.
Conceptual division
 en. corner = sp. rincĂłn (internal),
esquina (external)
Lexical gaps
 it. fuoricorso, consuocero :
not lexicalized in English and in French
Table 1.
Very few lexical resources, however, have the required features to be used in an MT system. As a matter of fact, besides providing a rich and various amount of information, a lexicon must guarantee completeness and coherence of the encoded lexical data. Moreover, it must be conceived as a dynamic resource, and not as a static and crystallized repertory of lexical information. Such a resource should be simple to update and expand not only manually but essentially through the automatic acquisition of information from textual resources, so as to reflect the continuous evolution of languages and to meet the new needs and answer the problematic issues which might emerge from the translation process. In this perspective, a generic lexical model and a modular architecture are essential for an electronic lexicon to be profitably exploitable.
A large computational lexical resource for the Italian language was developed at the Istituto di Linguistica Computazionale of the National
68 Humana.Mente – Issue 23 – December 2012
Research Council in Pisa from 1996 to 2003, which presents these characteristics.
4. The Lexical Resource
The computational lexicon PAROLE-SIMPLE-CLIPS (Ruimy et al., 1998; 2002; 2003), elaborated in the framework of three different projects1, provides a wide-coverage, four-level description of the Italian language. This lexical resource was built according to a multifunctional and multilingual perspective and in compliance with the international standards set out in the PAROLE-SIMPLE lexical model (Ruimy et al., 1998; Lenci et al., 2000).
This model, based on the EAGLES recommendations (San Filippo et al., 1998) and on an extended version of the GENELEX model (Antoni-Lay et al., 1994), is at the forefront of the field of Computational Lexicography for some outstanding and innovative features. The flexible architecture of the model as well as the building methodology allow the coherent encoding of a wide range of highly structured information, at the desired granularity level. Consensually adopted at a European level for the building of twelve harmonized monolingual electronic lexicons, the PAROLE-SIMPLE lexical model became a de facto standard and subsequently strongly inspired the ISO standard for NLP lexicons, the metamodel Lexical Markup Framework2.
The PAROLE-SIMPLE-CLIPS lexicon offers, therefore, the outstanding advantage of being compatible with eleven other lexicons developed for European languages, with which it shares the theoretical and representational model, the working methodology as well as a kernel of entries.
The lexicon is articulated in four independent but interrelated modules, which correspond respectively to the phonological, morphological, syntactic and semantic levels of linguistic representation. The complete description of a lexical unit consists therefore in a minimum of four interconnected entries, each one providing a structured set of information relevant to the description level that hosts it.
A phonological entry accounts for the phonetic and phonological features of a lexical unit while a morphological entry informs on its grammatical category and inflectional paradigm. A syntactic entry describes both the
1 The European projects LE-PAROLE and LE-SIMPLE and the Italian project Corpora e Lessici dell’italiano Parlato e Scritto (CLIPS)
2 ISO-24613:2008
Automated Translation between Lexicon and Corpora 69
intrinsic and contextual properties of a lexical unit in one specific syntactic structure. The subcategorization frame is modelled in terms of syntactic category, grammatical function, optionality and morphosyntactic, syntactic and lexical restrictions of the governed elements. Systematic frame alternations, such as the causative-inchoative variation, are represented in a complex entry whereby the correspondence between the constituents of the two structures is specified.
The adopted theoretical framework for the representation of semantic information is based on the fundamental principles of the Generative Lexicon theory (Pustejovsky, 1995). In a generative lexicon, a semantic unit is modelled through four different levels of representation3 that account for the componential aspect of meaning, define the type of event denoted, describe its semantic context and set its hierarchical position with respect to other lexicon units.
The semantic lexicon is structured in terms of an ontology of semantic types (the SIMPLE ontology). In a semantic entry, which encodes a single meaning of a lexeme, the membership in an ontological type represents the primary and most relevant information. Besides the ontological classification, the semantic unit is endowed with information concerning its domain of use; the type of event it denotes, where relevant; some distinctive semantic features; its links with other lexical units - among which synonymy and morphological derivation links - and membership in a class of regular polysemy. The semantic frame of predicative units is also described in terms of semantic role and selectional restrictions of the arguments.
To express the links holding among sense units, the SIMPLE lexicographers benefited from a remarkably efficient expressive means, the Extended Qualia Structure. This representational tool was derived from the Qualia Structure, a four-role4 structure which is considered a mainstay in the Generative Lexicon theory for representing the multidimensionality of a word’s meaning. The extended structure was created by defining, for each of the four Qualia roles, a subset of semantic relations. Such relations obviously allowed a much sharper expression of both the multidimensional aspect of a word sense and the nature of its syntagmatic and paradigmatic links to other lexical units. To give but one example, considering the telic role that informs
3 Namely Qualia structure, Event Structure Argument Structure and Lexical Typing Structure.
4 Formal, constitutive, agentive and telic.
70 Humana.Mente – Issue 23 – December 2012
about the function or purpose of an entity, the most appropriate relation may be selected among the following ones: ‘used_for’, ‘used_by’, ‘used_as’, ‘used_against’, ‘is_the_activity_of’, ‘object_of_the_activity’ and so on.
Moreover, in a new and revised version of the lexical-semantic database, called Simple_PLUS, the semantic representation has been enriched with significant information concerning the relationships holding between events and their participants and among co-participants in events (Ruimy, 2010).
This lexicon offers, therefore, a wide range of very rich and interesting information, especially at the semantic level. It is our deep conviction that an MT system could greatly benefit from such a wealth of lexical data, for both the granularity of the information provided and its explicit formulation.
5. Lexical Semantics for the resolution of some MT problems
A translation process presupposes the understanding of the many and various aspects that characterize the input text. Besides the morphological and syntactic aspects, it is necessary to disambiguate the logical form of the sentence, checking the coherence among semantic restrictions and preferences of words. To establish an equivalence between a source and a target text a translator should also understand other semantic and pragmatic aspects (for example conversational implicatures, metaphors, ironic contexts, etc.), that are not easily detectable. In the following, we will briefly show how Lexical Semantics plays a central role in the resolution of problems that typically emerge in Machine Translation.
Word sense ambiguity is a pervasive characteristic of natural language. It is one of the main reasons for poor performance of Information Retrieval systems. In MT, lexical ambiguity may occur both in the analysis and the transfer phases. Its resolution, which is therefore considered a major problem, requires a large amount of rich lexical knowledge.
5.1.1. Polysemy / homonymy and domain knowledge
A SL polysemic word or two SL homonyms may translate in two different ways according to their usage domain (see Table 2). Matching the information concerning the topic of the source text and the indication, in the monolingual lexicon, of the different domains of use of the ambiguous word enables the selection, in the bilingual lexicon, of the appropriate translation.
Automated Translation between Lexicon and Corpora 71
en.
mouse

it.
(gen.)
(inform.)
topo
mouse
it.
borsa

en.
(gen.)
(econ.)
bag
stock exchange
it.
calcolo

en.
(gen.)
(med.)
calculation
gallstone
Table 2.
5.1.2. Polysemy / homonymy and ontological classification
The semantic classification of a word sense is generally sufficient to discriminate among its different meanings or among homonyms and therefore to enable the selection of the relevant one from its different possible translational equivalents, as shown in Table 3 for Italian-English and Italian-French translations.
Italian

English
French
ala : [PART]

wing
aile
ala : [BODY_PART]

wing
aile
ala : [ROLE]

winger
ailier
espresso [ARTIFACT_DRINK]

espresso
express
espresso [VEHICLE]

express (train)
express
espresso [SEMIOTIC_ARTIFACT]

express (letter)
exprès
Table 3.
5.1.3. Polysemy / homonymy and contextual links
More complex situations emerge when two readings of a lemma cannot be disambiguated through their semantic classification or other paradigmatic information. In this case, syntagmatic and therefore contextual links may be used. In the following example reported in Table 4, means for selecting the appropriate translation are provided by the domain of use, but also by semantic relations linking each ambiguous term to the predicate denoting its function.
72 Humana.Mente – Issue 23 – December 2012
Italian

English
ferri_1
[INSTRUMENT] used_for sferruzzare (to knit)

knitting needles
ferri_2
[INSTRUMENT] used_for operare (to operate)

surgical instruments
Table 4.
5.1.4. Polysemy / homonymy and semantic frame
The semantic frame description may also provide clues for solving lexical ambiguities. Two homonym predicates may be distinguished by a different argument structure, either by the number of arguments they require (Table 5, first example) or by the semantic restrictions imposed on those arguments (Table 5, second example).
Italian

English
avvertire1: arg0, arg1, arg2

to inform, to warn
avvertire2: arg0, arg1

to feel, to notice
Italian

English
camminare1: arg0 = + animate

walk
camminare2:arg0 = - animate

work
Table 5.
It is worth noting that the whole range of lexical semantic information used for solving the above cases of ambiguity is encoded in the lexicon presented in the previous section.
6. The Corpus-based Approach
In order to briefly illustrate how a corpus-based approach may work, we have decided to focus our attention on one specific example, the translation of the English phrasal verb ‘set up’ into Italian, gathering our samples from electronic texts on the Web and analyzing them with a KeyWord in Context (KWIC) tool.
The experiment outlined here was carried out using the following procedure. A lexical item was selected, for the purposes of this analysis the
Automated Translation between Lexicon and Corpora 73
English phrasal verb ‘set up’ (Wade & Federici, 2006), since this item is problematic from a semantic point of view. It provides an interesting example of the highly polysemic nature of the English language, characterised by “remarkable range, flexibility and adaptability” (Crystal, 1988, p. 39). In this case the translator, for example, is required to consider the context specific nature of the lexical item (Eco, 2003, p. 29) and where areas of “inherent fuzziness” (Bell, 1991, p. 102) are found in establishing equivalence between one language and another. Indeed, ‘set’ alone has about 120 different meanings (cf. Collins Cobuild English Dictionary, 1995). With regard to ‘set up’, it was decided to first examine its meanings using a traditional bi-lingual dictionary Ragazzini-Zanichelli (2009). Secondly, a small sample of examples was collected from the web with a specifically designed search tool, followed by the manual examination and analysis of the gathered data and comparison with the information provided in the dictionary. The analysis was then extended through the analogical comparison of the initial manual analysis, allowing the further extraction of a wider sample of data.
To perform the kind of analysis described above, a tool was developed to acquire word-concordances directly from the web. The tool is a combination of several web/linguistic tools:
 a web spider that acquires a predefined number of web pages;
 a segmenter that splits acquired web pages;
 a rule-based lemmatiser;
 a KWIC (KeyWord In Context) tool;
 a self-learning analogy-based engine.
The web spider (cf. Federici, Wade, 2007) extracts web pages starting from a given web address, thus providing “a random snapshot of the current state of the Internet in a given language” (Sharoff, 2006, p. 437). The spider filters out all unneeded web overstructure (see Figure 1).
74 Humana.Mente – Issue 23 – December 2012
Figure 1
Then the lemmatiser associates each word form contained in the extracted
web pages to the corresponding lemma. After the corpus has been cleaned and
lemmatised, the KWIC will read the corpus by indexing all the lemmas. This is
illustrated in Figure 2, where the word forms or lemmas are in the keyword
area on the left (2a), and clicking on the keyword which is the focus of our
interest the concordances are created (2b).
Figure 2a
Automated Translation between Lexicon and Corpora 75
AREA KEYWORDS AREA CONCORDANZE
Figure 2b
While this approach is certainly useful as it enables the linguist to capture
the real usage of a given word, it also suffers from a number of limitations:
 the manual analysis of data is extremely time-consuming;
 it is often not practicable to analyse all the examples, especially in
large corpora, so only a selected number of examples are chosen as
representative;
 there is the risk of human error and inconsistency in manual
analysis.
7. Corpus vs. Dictionary
Our starting point was an analysis of the word senses provided in bi-lingual
English-Italian dictionary Ragazzini-Zanichelli (2009). The result is illustrated
in Figure 3.
76 Humana.Mente – Issue 23 – December 2012
to set up (verbo transitivo)
1. mettere su; alzare; erigere; piantare; montare; installare; allestire
2. mettere su; montare; installare; allestire
3. mettere su; mettere in piedi; istituire; fondare; costituire; formare; aprire (un ufficio); avviare (un’azienda)
4. sistemare; mettere (q.) in affari (o politica ecc.); aiutare (q.) finanziariamente (politicamente ecc.)
5. lanciare (un grido)
6. causare; provocare; dare l’avvio (o il via) a
7. stabilire (sport)
8. comporre (tipog.)
9. tesare, arridare (naut.)
10. (fam.) rimettere in salute (o in forze; in sesto); tirare su
11. (fam.) montare un’accusa contro (q.); incastrare; mettere contro; mettersi a fare; fornire; essere forte; essere ben fornito
Figure 3
It is to be noted that only a restricted number of entries provide contextualised examples of usage.
An initial analysis of the corpus created with the tools described above, on the other hand, reveals significantly richer contextualised source. In fact, it becomes immediately apparent that there are cases which are not included in the dictionary, such as the meaning ‘creare’, which is the appropriate translation of ‘set up’ in the case illustrated below:
[…] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was […] useful information on the British Council's website, which was set up set upset upset upset up specifi specifispecifispecifispecifispecifically for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. cally for assistants to use in their placement countries. (“Foreign (“Foreign (“Foreign (“Foreign (“Foreign (“Foreign (“Foreign (“Foreign assistance” assistance” assistance” assistance” assistance” assistance” assistance” by Katie PhippsKatie Phipps Katie PhippsKatie PhippsKatie PhippsKatie PhippsKatie PhippsKatie PhippsKatie PhippsKatie Phipps, «Education Guardian online» «Education Guardian online»«Education Guardian online»«Education Guardian online» «Education Guardian online» «Education Guardian online»«Education Guardian online»«Education Guardian online»«Education Guardian online» «Education Guardian online»«Education Guardian online»«Education Guardian online» «Education Guardian online»«Education Guardian online» «Education Guardian online»«Education Guardian online»«Education Guardian online»«Education Guardian online»«Education Guardian online», August AugustAugustAugust 23th 23th 23th 23th 23th 2005)
In an experiment that analysed 600 contexts of ‘set up’, only 8 out of 17 translations were attested in the dictionary. While it may be argued that the entries in the dictionary could be the most frequent usages of ‘set up’, it does not seem to be the case if we consider that the dictionary covers only about 47% of the translations of ‘set up’ occurring in our corpus (see Figure 4).
Automated Translation between Lexicon and Corpora 77
Translations present in the dictionary
Translations not present in the dictionary
ALLESTIRE
AVVIARE
COSTITUIRE
FONDARE
FORMARE
INSTALLARE
ISTITUIRE
STABILIRE
APPRONTARE
ATTUARE
CREARE
DEFINIRE
DIPINGERE
IMPOSTARE
ORGANIZZARE
PREPARARE
REALIZZARE
Figure 4
From these analyses it emerges that examples extracted from real texts may be useful (i) to extend coverage of the lexicon and (ii) to refine semantic entries.
8. Extending the study: the ‘bootstrapping’ process
In order to extend the study and refine the data gathered, we need to use some type of Artificial Intelligence engine that (semi-)automatically carries out the annotation task. The procedure applied for the purposes of this study is called ‘bootstrapping’.
In the first step a small portion of the corpus was annotated manually, assigning a translation to each sample (see Figure 5):
Manually annotated concordances:
1. […] websites have also been set up/CREARE by the LSC […]
2. […] Websites have also been set up/CREARE and open days organised […]
3. […] an appeal panel has been set up/COSTITUIRE by the Dept. […]
4. […] a panel, task force, set up/COSTITUIRE by Harvard […]
Figure 5
78 Humana.Mente – Issue 23 – December 2012
In the second step, the annotation is extended automatically to the remaining concordances for ‘set up’ in the corpus. At this stage it is found that not all of the translations assigned are correct (see Figure 6).
Manually annotated concordances:
1. […] websites have also been set up/CREARE by the LSC […]
2. […] Websites have also been set up/CREARE and open days organised […]
3. […] an appeal panel has been set up/COSTITUIRE by the Dept. […]
4. […] a panel, task force, set up/COSTITUIRE by Harvard […]
New concordances (Automatic annotation)
1. There are […] much more useful information on the British Council’s website, which was set up/CREARE […] for assistants […] (CORRECT)
2. […] a committee was set up/ISTITUIRE to arrange […] (WRONG)
Figure 6
During this automatic annotation step the first occurrence of ‘set up’ is automatically annotated as ‘CREARE’, which is correct, while the second is automatically annotated as ‘ISTITUIRE’, which is incorrect. This is because the algorithm in this case failed to provide the appropriate translation for lack of evidence.
In the third step, therefore, further manual revision is necessary. During this last phase the correct interpretation is manually assigned to those keywords that have been wrongly annotated.
9. Practical application of the procedure
We tested this procedure by setting up an experiment in which 600 contexts from a 1.5 million word corpus were manually annotated by assigning a translation to each concordance with ‘set up’, 400 new contexts were then automatically annotated and finally revised manually.
The results were encouraging, since the correctness of the automatically assigned translations was about 49%. That is, almost half of the time the
Automated Translation between Lexicon and Corpora 79
procedure assigns the correct translation, even starting from a relatively small set of training samples. This is acceptable when compared to the high number of possible translations (17) and less thorough baselines, such as the ones that could be obtained by assigning a random interpretation (1/17=6%) or just the most frequent one (that is “avviare”, that accounts for only 12% of the cases).
Conclusions
In our hypothesis, the corpus-based process outlined above might prove to be very useful in enhancing lexical resources. This study aimed to create a dynamic cyclical process, in which the lexicon, in the case of our web-based experiment, is enhanced by a corpus-based analysis, and the corpus-based analysis can then be automatized thanks to the availability of richer and more precise lexical knowledge. This would appear to be necessary when dealing with a dynamic process as opposing a static lexicon which fails to provide a complete descriptive picture of current language use. With the application of automated methods, a wide set of new lexical data and knowledge can be collected and analyzed.
There is the need, therefore, for the implementation of systems which are able to dynamically extend/enhance/update lexicons with information acquired from large corpora and from the web. Our objective should be to set up a new generation of large-size, dynamic lexical resources that fully capture current language usage (how language is materially manifested) and use (the way in which language forms are used as a means of communication) (Widdowson, 1978, pp. 18-19).
ACKNOWLEDGEMENTS

This work is the outcome of a collaborative effort. However, for the specific concerns of Italian academy, Elisabetta Gola is responsible for paragraphs 1-3, Stefano Federici is responsible for 8-9, Nilda Ruimy is responsible for 4-5 and John Wade is responsible for 6-7.
80 Humana.Mente – Issue 23 – December 2012

REFERENCES
Antoni-Lay, M.-H., Francopoulo, G., Zaysser, L. (1994). Generic Model for Reuseable Lexicons: The Genelex Project, Literary and Linguistic Computing, 9(1), 47-54.
Bell, R.T. (1991). Translation and Translating: Theory and Practice. Harlow: Longman.
Brown, P.F., Cocke, J., Della Pietra, S., Della Pietra, V., Jelinek, F., Mercer, R., Roossin, P. (1990). A statistical approach to language translation, Computational Linguistics, 16, 79-85.
Brown, R.D. (1999). Example-based machine translation, accessed on http://www.cs.cmu.edu/afs/cs.cmu.edu/user/ralf/pub/WWW/ebmt/ebmt.html [2000, January 6].
Collins Cobuild English Dictionary (1995). London: Harper Collins Publisher.
Crystal, D. (1988). The English Language, London: Penguin.
Dizionario Inglese-Italiano, Italiano Inglese Ragazzini-Zanichelli. Zanichelli Editore (2009).
Eco, U. (2003). Dire quasi la stessa cosa: Esperienze di traduzione. Milano: Bompiani.
Federici, S., Wade J.C. (2007). Letting in the light and working with the Web: A dynamic corpus development approach to interpreting metaphor. In M. Davis, P. Rayson, S. Hunston P. Danielsson, eds. Proceedings of Corpus Linguistics Conference 2007, University of Birmingham (UK), http://corpus.bham.ac.uk/corplingproceedings07/paper/207_Paper.pdf.
Grice, H. P. (1938). Negation. The Grice Papers, UC/Berkeley, Bancroft Library.
Grice, H. P. (1989). Studies in the way of words.
Horn, L. R. A brief history of negation.
Hunston, S. (2002). Corpora in Applied Linguistics. Cambridge, UK: Cambridge University Press.
Hutchins, J. (1996). ALPAC: the (in)famous report, MT News International 14, June 1996, 9-12.
Hutchins, J. (1986). Machine Translation: Past, Present, Future. Chichester: Ellis Horwood Series in Computers and their Applications.
Hutchins, J. (1993). Latest Developments in Machine Translation technology: Beginning a New Era in MT research. In Proceedings MT Summit IV.:
Automated Translation between Lexicon and Corpora 81
International cooperation for global communication, July 20-22, 1993, Kobe, Japan, 11-34.
Hutchins, J. (2003). Machine translation: general overview. In R. Mitkov (ed.), The Oxford Handbook of Computational Linguistics, Oxford: Oxford University Press, 501-511.
Isabelle, P., Dymetman, M., Macklovitch, E. (1988). CRITTER: a translation system for agricultural market reports”, Proceedings of the 12th conference on Computational linguistics - Volume 1, Budapest, 261-266.
Lenci, A., Bel, N., Busa, F., Calzolari N., Gola, E., Monachini, M., Ogonowski, A., Peters I., Peters, W., Ruimy, N., Villegas, M., Zampolli, A. (2000). SIMPLE: A General Framework for the Development of Multilingual Lexicon, International Journal of Lexicography, special issue, Dictionaries, Thesauri and Lexical-Semantic Relations, 13(4), 249-263.
Nagao, M. (1984). “A framework of a mechanical translation between Japanese and English by analogy principle”, Artificial and Human Intelligence: edited review papers at the International NATO Symposium on Artificial and Human Intelligence sponsored by the Special Programme Panel held in Lyon, France, October, 1981, Amsterdam: Elsevier Science Publishers, 173-180.
Pustejovsky, J. (1995). The Generative Lexicon, , Cambridge, MA: The MIT Press.
Ruimy, N., Corazzari, O., Gola, E., Spanu, A., Calzolari, N., Zampolli, A. (1998). The European LE-PAROLE Project: The Italian Syntactic Lexicon, LREC (1998) First International Conference on Language Resources and Evaluation Proceedings, I, Granada, Spain, 241-248.
Ruimy, N., Monachini, M., Distante, R., Guazzini, E., Molino, S., Ulivieri, M., Calzolari, N., Zampolli, A. (2002). CLIPS, a Multi-level Italian Computational Lexicon, LREC (2002) Third International Conference on language resources and evaluation proceedings, III, Las Palmas de Gran Canaria, 792-799.
Ruimy, N., Monachini, M., Gola, E., Calzolari, N., Del Fiorentino, M.C., Ulivieri, M., Rossi, S. (2003). A computational semantic lexicon of Italian: SIMPLE,. In A. Zampolli, N. Calzolari, L. Cignoni (Eds.), Computational Linguistics in Pisa. Linguistica Computazionale, Special Issue, XVIII-XIX (II), Pisa-Roma: IEPI, 821-864.
Ruimy, N. (2010). Simple_PLUS: a network of lexical semantic relations Simple_PLUS: una red de relaciones léxico-semánticas. In: Procesamiento del
82 Humana.Mente – Issue 23 – December 2012
Lenguaje Natural, 44, Sociedad Española para el Procesamiento del Lenguaje Natural, 99-106.
Sanfilippo, A., et al. (1998) EAGLES Preliminary recommendations on semantic encoding, The EAGLES Lexicon Interest Group, http://www.ilc.cnr.it/EAGLES/EAGLESLE.PDF.
Sharoff, S. (2006). Open-source corpora: using the net to fish for linguistic data, The International Journal of Corpus Linguistics, 11(4), 435-462.
Speranza, Join the Grice Club.
Turcato, D., Mcfetridge, P., Popowich, F., Toole, J. (1999). A unified example-based and lexicalist approach to machine translation. In Proceedings of the 8th International Conference on theoretical and methodological issues in Machine Translation (TMI '99), Chester.
Wade, J.C., Federici, S. (2006). Struttura-significato. Il processo di traduzione. In R. Pititto, S. Venezia (Eds.), Tradurre e comprendere: pluralitĂ  dei linguaggi e delle culture (Atti del XII Congresso Nazionale della SocietĂ  di Filosofia del Linguaggio, Piano di Sorrento 2005), Roma: Aracne, 307-332.
Whitelock, P. (1992). Shake-and-bake translation, Proceedings of the 14th conference on Computational linguistics, 2, Nantes, 784-791.
Widdowson H.G. (1978) Teaching language as communication. Oxford: Oxford University Press