Proceedings of the 3rd Workshop on Scalable Natural Language Understanding,pages73–80,
NewYorkCity,June2006. c©2006AssociationforComputationalLinguistics
A(very)BriefIntroductiontoFluidConstructionGrammar
LucSteels(1,2)andJoachimdeBeule(1)
(1)UniversityofBrussels(VUBAILab)
(2)SONYComputerScienceLab-Paris
steels@arti.vub.ac.be
Abstract
Fluid Construction Grammar (FCG) is a
new linguistic formalism designed to ex-
plore in how far a construction gram-
mar approach can be used for handling
open-ended grounded dialogue, i.e. dia-
logue between or with autonomous em-
bodied agents about the world as experi-
enced through their sensory-motor appa-
ratus. We seek scalable, open-ended lan-
guage systems by giving agents both the
ability to use existing conventions or on-
tologies, and to invent or learn new ones
as the needs arise. This paper contains a
brief introduction to the key ideas behind
FCGanditscurrentstatus.
1 Introduction
Construction grammar is receiving growing atten-
tion lately, partly because it has allowed linguists
to discuss a wide range of phenomena which were
difficult to handle in earlier frameworks (Goldberg,
1995; OstmanFried, 2005; Croft, 2001), and partly
because it has allowed psychologists to describe in
a more satisfactory way early language develop-
ment(TomaselloBrooks,1999). Therewerealready
some attempts to formalise construction grammar
(KayFillmore, 1999) and build a computational im-
plementation (BergenChang, 2003), but many open
problems remain and at this early stage of fun-
damental research, it makes sense to explore al-
ternative approaches. In our team, we focus on
open-ended grounded dialogue, in other words how
it is possible for a speaker to formulate an utter-
ance about the world and for a hearer to under-
stand what is meant (ClarkBrennan, 1991). The
present paper briefly reports on the formalisation
of construction grammar called Fluid Construc-
tion Grammar (FCG) that we have developed for
this research. Although the formalism is novel in
several fundamental aspects, it also builds heav-
ily on the state of the art in formal and computa-
tionallinguistics,particularlywithinthetraditionof
unification-basedfeaturestructuregrammarssuchas
HPSG (PollardSag, 1994). FCG has been under de-
velopment from around 2001 and an implementa-
tion on a LISP substrate has been released through
http://arti.vub.ac.be/FCG/ in 2005. The FCG core
engine (for parsing and production) is fully opera-
tionalandhasalreadybeenusedinsomelarge-scale
experimentsinlanguagegrounding(SteelsLoetzsch,
2006). We do not claim to have a complete solu-
tionforalllinguisticissuesthatariseinconstruction
grammar,andneitherdoweclaimthatthesolutions
wehaveadoptedsofararefinal. Onthecontrary,we
areawareofmanydifficulttechnicalissuesthatstill
remainunresolvedandwelcomeanydiscussionthat
wouldbringusforward.
2 Motivations
FCG grew out of efforts to understand the creative
basis of language. Language creativity is more than
the application of an existing set of rules (even if
the rules are recursive and thus allow an infinite set
ofpossiblesentences). Humanlanguageusersoften
stretch and expand rules whenever the need arises,
73
Figure 1: Typical experimental setup. The bot-
tom shows two robots moving around in an envi-
ronment that contains balls and boxes. The robots
are equiped with a complex sensory-motor system,
able to detect the objects and build an analog world
modeloftheirlocationandtrajectories(asshownin
therighttopcorner).
and occasionally invent totally new ones. So we
need to understand how new aspects of language
(new concepts and conceptualisations, new lexical
items, new syntactic and semantic categories, new
grammaticalconstructions,newinteractionpatterns)
may arise and spread in a population, the same way
biologiststrytounderstandhownewlifeformsmay
arise(Steels,2003).
This motivation leads immediately to some require-
ments. First of all we always use multi-agent sim-
ulations so that we can investigate the spreading of
conventions in a population. Agents take turns be-
ing speaker and hearer and build up competences in
conceptualisation and verbalisation (for production)
and parsing and interpretation (for understanding).
Theymustbeabletostoreaninventoryofrulesand
apply them in either processing direction, and they
must be able to expand their inventories both by in-
ventingnewconstructionsifnecessaryandbyadopt-
ing those used by others. Second, the agents must
have something to talk about. We are interested in
groundedlanguage,whichmeansdialogueaboutob-
jects and events in the world as perceived through a
sensory-motor apparatus. We take embodiment lit-
erally. Our experiments use physical robots (Sony
AIBOs)locatedinarealworldenvironment(seefig-
ure1from(SteelsLoetzsch,2006))Third,theagents
must be motivated to say and learn something. We
achievethisbyprogrammingtherobotswithscripts
to play language games. A language game sets up
a joint attentional frame so that robots share gen-
eral motives for interaction, a specific communica-
tive goal (for example draw attention to an object),
and give feedback to enable repair of miscommu-
nication (for example through pointing). We typi-
cally perform experiments in which a population of
agents starts with empty conceptual and linguistic
repertoiresandthenbuildsfromscratchacommuni-
cationsystemthatisadequateforaparticularkindof
language game. Agents seek to maximise commu-
nicativesuccesswhileminimisingcognitiveeffort.
One advantage of grounded language experiments
is that we can clearly monitor whether the capaci-
ties given to the agents are adequate for bootstrap-
ping a language system and how efficient and suc-
cessful they are. By starting from scratch, we can
alsotestwhetherourobjectiveofunderstandinglan-
guage creativity has been achieved. Of course such
experiments will never spontaneously lead to the
emergenceofEnglishoranyotherhumanlanguage,
butwecanlearnagreatdealabouttheprocessesthat
havegivenriseandarestillshapingsuchlanguages.
3 Meaning
The information about an utterance is organized in
a semantic and a syntactic structure. The seman-
tic structure is a decomposition of the utterance’s
meaning and contains language-specific semantic
re-categorisations (for example a put-event is cate-
gorised as a cause-move-location with an agent, a
patient and a location). The syntactic structure is
a decomposition of the form of the utterance into
constituentsandmorphemesandcontainsadditional
syntactic categorisations such as syntactic features
(like number and gender), word order constraints,
etc.
We follow a procedural semantics approach, in the
sense that the meaning of an utterance is a program
that the hearer is assumed to execute (Winograd,
1972; Johnson-Laird, 1997). Hence conceptualisa-
tion becomes a planning process (to plan the pro-
gram) and interpretation becomes the execution of
a program. For example, the meaning of a phrase
like ”the box” is taken to be a program that in-
volves the application of an image schema to the
flow of perceptual images and anchor it to a partic-
74
ular physical object in the scene. So we do not as-
sume some pre-defined or pre-processed logic-style
fact base containing the present status of the world
(as this is extremely difficult to extract and main-
tain from real world perception in a noisy and fast
changing world) but view language as playing an
active role in how the world is perceived and cate-
gorised. It is in principle possible to use many dif-
ferent programming languages, but we have opted
for constraint based processing and designed a new
constraint programming language IRL (Incremental
RecruitmentLanguage)andimplementedtheneces-
sary planning, chunking and execution mechanisms
of constraint networks (SteelsBleys, 2005). A sim-
pleexampleofaconstraintnetworkfor”thebox”is
asfollows
1
:
1. (equal-to-context ?s)
2. (filter-set-prototype ?r ?s ?p)
3. (prototype ?p [box])
4. (select-element ?o ?r ?d)
5. (determiner ?d [single-unique])
Equal-to-context, select-element,
etc. are primitive constraints that implement funda-
mentalcognitiveoperators. Equal-to-context
grabs the set of elements in the current context
and binds it to ?s. Filter-set-prototype
filters this set with a prototype ?p which is bound
in (3) to [box]. Select-element selects an
element ?o from ?r according to the determiner
?d which is bound to [single-unique] in
(5), meaning that ?r should be a singleton. The
constraints are powerful enough to be used both in
interpretation, when semantic objects such as pro-
totypes, determiners, categories, relations, etc. are
supplied through language and values need to be
found for other variables, and in conceptualisation,
when these values are known but the objective is
to find the semantic objects. Moreover, during
conceptualization the constraints may extend the
repertoire of semantic objects (e.g. introducing a
new prototype) if needed, allowing the agents to
progressivelybuilduptheirontologies.
1
We use prefix notation. Order does not play a role as the
constraint interpreter cycles through the network until all vari-
ablesareboundoruntilnofurtherprogresscanbemade. Sym-
bolsstartingwithaquestionmarkrepresentvariables.
Figure2: Left: decompositionoftheconstraintpro-
gram for “the ball” in the semantic structure. Right:
related syntactic structure. In reality both structures
containalotmoreinformation.
4 SyntacticandSemanticStructures
Asmentioned,FCGorganisestheinformationabout
an utterance in feature structures, similar to other
feature-structure based formalisms (as first intro-
duced by Kay (Kay, 1984)) but with some impor-
tant differences. An FCG feature structure contains
units which correspond (roughly) to words (more
preciselymorphemes)andconstituents.
Aunithasanameandasetoffeatures. Hierarchical
structureisnotimplicitlyrepresentedbyembedding
one unit in another one, but explicitly by the fea-
tures syn-subunits (for the syntactic structure) and
sem-subunits (forthesemanticstructure). Thereisa
strongcorrespondencebetweenthesyntacticandse-
mantic structurebuilt up forthe same utterance(see
figure 2) although there can be units which only ap-
pearinthesyntacticstructure(forexampleforgram-
matical function words) and vice versa. The cor-
respondence is maintained by using the same unit
names in both the semantic and syntactic structure.
Units in syntactic structures have three features: (1)
syn-subunits, (2) syn-cat which contains the syn-
tactic categories, and (3) form containing the form
associated with the unit. Units in semantic struc-
tures have four features: (1) sem-subunits, (2) sem-
cat containing the semantic categories, (3) meaning
whichisthepartoftheutterance’smeaningcovered
bytheunit,and(4)contextwhichcontainsvariables
that occur in the meaning but are ’external’ in the
sense that they are linked to variables occurring in
the meaning of other units. An example semantic
structure (in list-notation) for the left structure in
figure 2 is shown in figure 3. FCG is a completely
open-endedformalisminthesensethatalllinguistic
75
Figure3: Semanticstructureinlist-notation.
categories (syntactic or semantic) are open and in
principle language-specific (as in radical construc-
tion grammar (Croft, 2001).) Thus the set of lexical
categories (noun, verb, adjective, etc.), of possible
semanticroles(agent,patient,etc.),ofsyntacticfea-
tures (number, gender, politeness, etc.), and so on,
are all open. The value of the syn-cat and sem-cat
featuresconsistsofaconjunctionofpredicates(each
possibly having arguments.) New categories can be
introduced at any time and used as (part of) a pred-
icate. The form of the utterance is described in a
declarativemanner,usingpredicateslikeprecedesor
meets which define linear ordering relations among
theformofunitsoranyotheraspectofsurfaceform
includingprosodiccontourorstress.
5 Rules
A rule (also called template) typically expresses
constraints on possible meaning-form mappings.
Each rule has a score which reflects the success
that the agent has had in using it. All else be-
ing equal, agents prefer rules with higher scores,
thus reflecting frequency effects. A rule has two
poles. A left pole which typically contains con-
straints on semantic structure formulated as a fea-
ture structure with variables, and a right pole which
typically contains constraints on syntactic structure
again formulated as a feature structure with vari-
ables. Rules are divided into rule subsets which
help constrain the order of rule-application and de-
signlarge-scalegrammars. Thuswemakeadistinc-
tionbetweenmorph-rules,whichdecomposeaword
into a stem and pending morphemes and introduce
syntacticcategories;lex-stem-rules,whichassociate
meaning with the stem as well as valence informa-
tion and a role-frame; con-rules, which correspond
to grammatical constructions that associate parts of
semantic structure with parts of syntactic structure;
andsemandsyn-ruleswhichperforminferenceover
semantic or syntactic categories to expand semantic
orsyntacticstructure.
All rules are bi-directional. Typically, during pro-
duction, the left pole is ‘unified’ with the semantic
structure under construction, possibly yielding a set
ofbindings. Ifsuccessful,therightpoleis‘merged’
with the syntactic structure under construction. The
merge operation can be understood as a partial uni-
fication, but extending the structure with those parts
of the pole that were missing. During parsing, the
right pole is unified with the syntactic structure and
parts of the left pole are added to the semantic
structure. The unification phase is thus used to see
whetheraruleistriggeredandthemergephaserep-
resents the actual application of the rule. The FCG
Unify and Merge operators are defined in great for-
mal detail in (SteelsDeBeule, 2006). During pro-
duction lex-stem-rules are applied before the con-
rules and the morph-rules. During parsing the lex-
stem-rules are applied right after the morph-rules.
The con-rules then build higher order structure. It
is enormously challenging to write rules that work
in both directions but this strong constraint is very
helpfultoachieveacompactpowerfulgrammar.
6 BuildingHierarchy
One of the innovative aspects of FCG is the way it
handles hierarchy. Both the left-pole and the right-
pole of a construction can introduce hierarchical
structurewiththeJ-operator(DeBeuleSteels,2005).
Thisway,thesemanticpoleofconstructions(lexical
or grammatical) can decompose the meaning to be
expressed (which originally resides in the top node
ofthesemanticstructure)andthesyntacticpolecan
group units together into a larger constituent. Con-
straints governed by the J-operator do not have to
match during the unification phase. Instead they are
used to build additional structure during the merge
76
Figure 4: Example lexical entry for “put” and illus-
trationoftheJ-operator.
phase. This may include the construction of a new
unitaswellaspendingfromanexistingunitandab-
sorbingsomeotherunits.
Figure 4 shows an example which will be used fur-
ther in the next section. It is a lexical rule prepar-
ing a resultative construction (GoldbergJackendoff,
2004). The semantic pole of the rule combines
some stretch of meaning (the introduction of an
event-type,namelyaput-event)withaframe(cause-
move-locationwithrolesforagent,patientandloca-
tion). Theseareassociatedwithalexicalstem”put”
in the right pole which also adds a valence frame
SVOL (triggering the subject-verb-object-location
construction). In production, this rule triggers when
a‘put’event-typeispartofthemeaning(‘==’means
‘includes but may also contain additional expres-
sions’). When merging the semantic pole with the
semantic structure, a new unit hanging from ?top is
created and the specified value of the meaning fea-
ture copied down. The new unit also receives the
context and sem-cat features as specified by the J-
operator. At the same time, the syntactic pole is
mergedwiththesyntacticstructureandsothe?new-
unit (which is already bound) is added as a subunit
of ?top in the syntactic structure as well. The J-
operator will then add stem and valence informa-
tion. Thus the semantic structure of figure 5 will
betransformedintotheoneoffigure6. Andthecor-
responding syntactic structure becomes as in figure
7. In parsing, an existing syntactic unit with stem
((unit-2
(meaning
( ..
(event-type ev-type1
(put (put-1 o1) (put-2 o11)
(put-3 o22))) ... ))))
Figure 5: Semantic structure triggering the rule in
figure4inproduction.
((unit-2
(sem-subunits (... unit-3 ...)))
(unit-3
(meaning
((event-type
ev-type1
(put (put-1 o1) (put-2 o11)
(put-3 o22)))))
(context ((link ev-type1)))
(sem-cat
((sem-event-type
ev-type1
(cause-move-location
(agent o1) (patient o11)
(location o22))))))
... )
Figure 6: Resulting semantic structure after apply-
ing the rule in figure 4 to the semantic structure of
figure5.
((unit-2
(syn-subunits (... unit-3 ...)))
(unit-3
(form ((stem unit-3 "put")))
(syn-cat ((valence SVOL))))
... )
Figure 7: Resulting syntactic structure after apply-
ingtheruleinfigure4.
”put” is required to trigger the rule. If found, the
rule will add the valence information to it and on
77
thesemanticsidethemeaningaswellastheseman-
tic categorisation in terms of a cause-move-location
frameareadded.
7 ImplementingConstructions
Lexicalconstructionsprovideframeandvalencein-
formation for word stems and parts of meaning.
Grammatical constructions bind all this together.
Figure 8 shows an example of a grammatical con-
struction. It also uses the J-operator to build hier-
archy, both on the semantic side (to decompose or
add meaning) and on the syntactic side (to group
constituents together.) An example of a SVOL-
construct is Mary puts the milk in the refrigerator.
Beforeapplicationoftheconstruction,variousunits
should already group together the words making up
a nounphrase for the subject (which will be bound
to ?subject-unit), a nounphrase for the direct object
(boundtothe?object-unit)andaprepositionalnoun-
phrase(boundto?oblique-unit). Eachoftheseunits
also will bind variables to their referents, commu-
nicated as context to the others. On the semantic
side the cause-move-location frame with its various
roles aids to make sure that all the right variable
bindings are established. On the syntactic side the
construction imposes word-order constraints (ex-
pressedwiththemeets-predicate),thevalenceofthe
verb,andspecifictypesofconstituents(nounphrase,
verbphrase, prepositional nounphrase). The SVOL
construction operates again in two directions. In
production it is triggered when the semantic struc-
ture built so far unifies with the semantic pole, and
then the syntactic structure is expanded with the
missing parts from the syntactic pole. Constraints
on the syntactic pole (e.g. valence) may prevent
the application of the construction. In parsing, the
SVOL construction is triggered when the syntactic
structure built so far unifies with the syntactic pole
andthesemanticstructureisthenexpandedwiththe
missing parts from the semantic pole. Again ap-
plication may be constrained when semantic con-
straintsintheconstructionpreventit.
8 Fluidity,Conventionalisationand
Meta-grammars
Although FCG must become adequate for dealing
with the typical phenomena that we find in human
natural languages, our main target is to make scien-
tificmodelsoftheprocessesthatunderlytheorigins
of language, in other words of the creative process
by which language users adapt or invent new forms
toexpressnewmeaningsthatunavoidablyariseinan
openworldandnegotiatetacitlytheconventionsthat
they adopt as a group. We have already carried out
a number of experiments in this direction and here
only a brief summary can be given (for more dis-
cussion see: (Steels, 2004; DeBeuleBergen, 2006;
SteelsLoetzsch,2006)).
In our experiments, speaker and hearer are cho-
sen randomly from a population to play a language
game as part of a situated embodied interaction that
involves perception, joint attention and feedback.
When the speaker conceptualizes the scene, he may
construct new semantic objects (for example new
categories) or recruit new constraint networks in
order to achieve the communicative goal imposed
by the game. Also when the speaker is trying to
verbalise the constraint network that constitutes the
meaning of an utterance, there may be lexical items
missing or new constructions may have to be built.
We use a meta-level architecture with reflection to
organise this process. The speaker goes through the
normalprocessingsteps,usingwhateverinventoryis
available. Missing items may accumulate and then
the speaker moves to a meta-level, trying to repair
the utterance by stretching existing constructions,
re-using them by analogy for new purposes, or in-
troducing other linguistic items. The speaker also
engages in self-monitoring by re-entering the utter-
ance and comparing what he meant to say to inter-
pretationsderivedbyparsinghisownutterance. The
speaker can thus detect potential problems for the
listener such as combinatorial explosions in pars-
ing, equalities among variables which were not ex-
pressed, etc. and these problems can be repaired by
theintroductionofadditionalrules.
The hearer receives an utterance and tries to go as
far as possible in the understanding process. The
parserandinterpreterarenotgearedtowardscheck-
ing for grammaticality but capable to handle utter-
ances even if a large part of the rules are missing.
The(partial)meaningisthenusedtoarriveatanin-
terpretation, aided by the fact that the context and
communicative goals are restricted by the language
game. Ifpossible,thehearergivesfeedbackonhow
78
heunderstoodtheutteranceandwhetheraninterpre-
tationwasfound. Ifthereisfailureormiscommuni-
cationthehearerwillthenrepairhisinventorybased
on extra information provided by the speaker. This
can imply the introduction of new concepts extend-
ing the ontology, storing new lexical items, intro-
ducing new constructions, assigning certain words
to new syntactic classes, etc. Speaker and hearer
also update the scores of all rules and concepts. In
case of success, scores go up of the items that were
used and competitors are decreased to achieve lat-
eral inhibition and hence a positive feedback loop
between success and use. In case of failure, scores
go down so that the likelihood of using the failing
solution diminishes. In our simulations, games are
played consecutively by members of a population
andwehavebeenabletoshow–sofarforrelatively
simple forms of language– that shared communi-
cation systems can emerge from scratch in popula-
tions. Muchworkremainstobedoneinresearching
therepairstrategiesneededandwhentheyshouldbe
triggered. The repair strategies themselves should
also be the subject of negotiation among the agents
because they make use of a meta-grammar that de-
scribes in terms of rules (with the same syntax and
processingastheFCGrulesdiscussedhere)howre-
pairsaretobeachieved.
9 Conclusions
FCG is a tool offered to the community of re-
searchers interested in construction grammar. It al-
lowsthepreciseformaldefinitionofconstructionsin
a unification-based feature structure grammar style
and contains the necessary complex machinery for
building an utterance starting from meaning and
reconstructing meaning starting from an utterance.
FCG does not make linguistic theorising superflu-
ous, on the contrary, the formalism is open to any
framework of linguistic categories or organisation
ofgrammaticalknowledgeaslongasaconstruction
grammar framework is adopted. There is obviously
a lot more to say, not only about how we handle
various linguistic phenomena (such as inheritance
of properties by a parent phrasal unit from its head
subunit) but also what learning operators can pro-
gressively build fluid construction grammars driven
bytheneedsofcommunication. Wereferthereader
to the growing number of papers that provide more
detailsonthesevariousaspects.
10 Acknowledgement
This research was conducted at the Sony Computer
Science Laboratory in Paris and the University of
Brussels VUB Artificial Intelligence Laboratory. It
is partially sponsored by the EU ECAgents project
(FET IST-1940). FCG and the experiments in lan-
guage evolution are team work and major contribu-
tions were made by Joris Bleys, Martin Loetzsch,
Nicolas Neubauer, Wouter Van den Broeck, Remy
VanTrijp,andPieterWellens.

References
Bergen, B.K. and N.C. Chang. (2003) Embodied Con-
struction Grammar in Simulation-Based Language
Understanding. Technical Report 02-004, Interna-
tionalComputerScienceInstitute,Berkeley.
Clark, H. and S. Brennan (1991) Grounding in com-
munication. In: Resnick, L. J. Levine and S. Teasley
(eds.)PerspectivesonSociallySharedCognition.APA
Books,Washington.p.127-149.
Croft, William A. (2001). Radical Construction Gram-
mar; SyntacticTheoryinTypologicalPerspective.Ox-
ford: OxfordUniversityPress.
De Beule, J. and L. Steels (2005) Hierarchy in Fluid
Construction Grammar. In: Furbach, U. (eds) (2005)
Proceedings of KI-2005. Lecture Notes in AI 3698.
Springer-Verlag,Berlin.p.1-15.
De Beule, J. and B. Bergen (2006) On the Emergence
of Compositionality. Proceedings of the Evolution of
LanguageConferenceVI,Rome.
Goldberg, A.E. (1995) Constructions.: A Construction
Grammar Approach to Argument Structure. Univer-
sityofChicagoPress,Chicago
Goldberg, A. and R. Jackendoff (2004) The English Re-
sultative as a Family of Constructions. Language 80
532-568.
Johnson-Laird, P.N. (1997) Procedural Semantics. Cog-
nition,5(1977)189-214.
Goldberg, A. (2003) Constructions: A new theoretical
approach to language Trends in Cognitive Science.
Volume7,Issue5,May2003,pp.219-224.
Kay, P. and C. Fillmore (1999) Grammatical construc-
tions and linguistic generalizations: the Whats X do-
ing Y? construction. Language75(1),133.
Kay, M. (1984) Functional unification grammar: A for-
malism for machine translation. Proceedings of the
International Conference of Computational Linguis-
tics.
Pollard, C.and I. Sag (1994) Head-driven Phrase Struc-
ture Grammar. CSLIStanfordUniv,Calif.
Ostman, Jan-Ola and Mirjam Fried (eds.) (2005) Con-
struction Grammars: Cognitive grounding and theo-
retical extensions. W.Benjamins,Amsterdam.
Steels, L. (2003) Evolving grounded communication for
robots. Trends in Cognitive Science. Volume 7, Issue
7,July2003,pp.308-312.
Steels, L. (2004) Constructivist Development of
Grounded Construction Grammars. In D. Scott, W.
DaelemansandM.Walker(Eds.),ProceedingsAnnual
MeetingofAssociationforComputationalLinguistics
Conference.Barcelona: ACL,(pp.9-16).
Steels, L. and J. De Beule (2006) Unify and Merge in
FCG.In: Vogt,P.et.al.(eds.)ProceedingsofEELCIII.
Lecture Notes in Computer Science. Springer-Verlag,
Berlin.
Steels,L.andJ.Bleys(2005)PlanningWhatToSay: Sec-
ond Order Semantics for Fluid Construction Gram-
mars. In: Proceedings of CAEPIA 2005. Santiago de
Compostella.
Steels,L.andM.Loetzsch(2006) Perspective Alignment
in Spatial Language. In: Coventry, K., J. Bateman
andT.Tenbrink(2006)SpatialLanguageinDialogue.
OxfordUniversityPress.Oxford.
Tomasello, M. and P.J. Brooks (1999) Early syntactic
development: A Construction Grammar approach In:
Barrett,M.(ed.)(1999)TheDevelopmentofLanguage
PsychologyPress,London.pp.161-190.
Winograd, T. (1972) Understanding natural language.
NewYork,AcademicPress.
