<ul><li><p>d binding:</p><p>ud</p><p>model of the neural architecture of language, with a focuson the role of Brocas area. The framework connects</p><p>Moreover, in recent linguistic theories, the distinction</p><p>ist Program [4] plays a similar role.</p><p>Opinion TRENDS in Cognitive Sciences Vol.9 No.9 September 2005taking during conversation.and Control. The Memory component comprises aspecification of the different types of language informationstored in long-term memory, as well as the retrievaloperations. The Unification component refers to theintegration of lexically retrieved information into arepresentation of multi-word utterances. The Controlcomponent relates language to action, and is invoked, forinstance, when the correct target language has to beselected (in the case of bilingualism), or for handling turn</p><p>Although the tripartite architecture holds equally forlanguage production and language comprehension, in theremainder of this article I will focus on languagecomprehension, as this has been the subject of mostneuroimaging studies of language. Moreover, the focuswill be on unification, because this is what Brocas areaand adjacent cortex are especially relevant for.</p><p>Syntactic unificationcomponents of language processing: Memory, Unification</p><p>From a design stance, one can distinguish three functional variables. The operation MERGE in Chomskys Minimal-psycholinguistically motivated processing components totheir neuronal substrate, guided by knowledge aboutbrain function across domains of cognition. The aim is toaccount for neuroimaging studies on language from thedesign perspective rather than the experimental taskperspective (e.g. phoneme identification, rhyme judgment,verb generate task, lexical decision, etc.). I believe that tomake sense of the rapidly increasing number of imagingstudies on language, a design perspective is urgentlyneeded. This perspective requires a specification of theprocessing requirements for mapping sound onto meaning(when listening), or vice versa (when speaking).</p><p>The design stance componentsOn Broca, brain, ana new frameworkPeter Hagoort</p><p>F.C. Donders Centre for Cognitive Neuroimaging &amp; NICI, Radbo</p><p>In speaking and comprehending language, word infor-mation is retrieved from memory and combined intolarger units (unification). Unification operations takeplace in parallel at the semantic, syntactic and phono-logical levels of processing. This article proposes a newframework that connects psycholinguistic models to aneurobiological account of language. According to thisproposal the left inferior frontal gyrus (LIFG) plays animportant role in unification. Research in other domainsof cognition indicates that left prefrontal cortex has thenecessary neurobiological characteristics for its involve-ment in the unification for language. I offer here apsycholinguistic perspective on the nature of languageunification and the role of LIFG.</p><p>IntroductionIn this article I will outline a new general framework for abetween lexical items and traditional rules of grammar isvanishing. For instance, Jackendoff proposes that the onlyremaining rule of grammar is UNIFY PIECES, and allthe pieces are stored in a common format that permitsunification. ([2], p. 180). The unification operations cliptogether lexicalized patterns containing one or moreUniversity Nijmegen, The Netherlands</p><p>In principle, the new MUC (Memory, Unification,Control) framework applies to both language productionand language comprehension, although details of theirfunctional anatomy within each component will bedifferent. I will first sketch the psycholinguistic motiva-tion behind the MUC framework.</p><p>Psycholinguistic backgroundIn psycholinguistic models of language processing, there isa general agreement that the language system has atripartite architecture (cf. [13]), with levels of sound,syntax and meaning as the core aspects of our languagefaculty. For all these levels the following dichotomyapplies: the basic information components are retrievedfrom long-term memory (the mental lexicon in psycho-linguistic terms), whereas additional information isderived from combinatorial operations (unification) thatassemble the basic components into larger structures.</p><p>Most current models of language processing agree that,in on-line sentence processing, different types of con-straints are very quickly taken into consideration duringspeaking and listening (or reading). Constraints on howwords can be structurally combined operate alongsidequalitatively distinct constraints on the combination ofword meanings, on the grouping of words into phono-logical and intonational phrases, and on their referentialbinding into a discourse model (see Figure 1).syntactic analysis. Here, again, a distinction can be madebetween the retrieval of syntactic frames from memory</p><p>Corresponding author: Hagoort, P. (peter.hagoort@fcdonders.ru.nl).Available online 27 July 2005</p><p>www.sciencedirect.com 1364-6613/$ - see front matter Q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.tics.2005.07.004Classically, accounts of unification in language focus on</p></li><li><p>ROpinion TRENDS in Cognitive Sciences Vol.9 No.9 September 2005 417O OON N NN O OO N NC C</p><p>R RR R</p><p>R</p><p>a</p><p>XX X</p><p>XX</p><p>b</p><p>Phonological structure(a)and their unification. The account of syntactic unifica-tion as proposed in the MUC framework [5] is inspiredby Aravind Joshis Tree-Adjoining Grammars and arelated explicit computational model of syntactic pro-cessing [6]. In this proposal, each word form in themental lexicon (Memory) is associated with a structuralframe. This structural frame consists of a three-tieredtree, specifying the possible structural environment ofthe particular lexical item (see Figure 2). The top layerof the frame consists of a single phrasal node (e.g. NP).This so-called root node is connected to one or morefunctional nodes (e.g. Subject, Head, Direct Object) inthe second layer of the frame. The third layer contains</p><p>Cl ClWd</p><p>Wd Wd</p><p>Wd Wdc e f gd</p><p>Syntactic structure</p><p>Semantic/conceptual structure</p><p>Spatial structure 2</p><p>cDet3 N3 singcount</p><p>fVAP</p><p>dA4</p><p>NP2</p><p>aS1</p><p>V6 Infl</p><p>pres73 sing</p><p>e 5</p><p>PRES7 BE6[TYPE:STAR]5DEF3[PropLITTLE]4Situation State Thing Pla2 ,</p><p>(b)</p><p>(c)</p><p>(d)</p><p>Figure 1. The tripartite architecture of the language system. The example gives the phono</p><p>little stars beside the big star. (d) shows an approximate spatial structure of the refere</p><p>www.sciencedirect.comSegmental</p><p>Prosodicstructure</p><p>Syllabicstructure</p><p>OON N NCC C C</p><p>R R R</p><p>XXX Xagain phrasal nodes to which lexical items or otherframes can be attached.</p><p>This parsing account is lexicalist in the sense that allsyntactic nodes (e.g. S, NP, VP, N, V, etc.) are retrievedfrom the mental lexicon. In other words, chunks ofsyntactic structure are stored in memory and there areno syntactic rules that introduce additional nodes.</p><p>In the on-line comprehension process, lexical items areretrieved sequentially, driven by the time course of theinput. The structural frames associated with the indivi-dual word forms thus enter the unification workspaceincrementally, in the order that is imposed by the input. Inthis workspace constituent structures spanning the whole</p><p>structure</p><p>Morpho-phonology</p><p>Cl Wd Wdh i j</p><p>Wd</p><p>10 8</p><p>bPP8</p><p>gP9</p><p>hDet11 N3 singcount</p><p>NP10</p><p>AP</p><p>iA12</p><p>VP</p><p>j 13</p><p>[TYPE:STAR]13INDEF11[PropBIG]12</p><p>BESIDE9ce Thing 10 18</p><p>logical (a), syntactic (b) and semantic/conceptual (c) structures for the sentence The</p><p>nce objects. Reprinted from [1] with permission of Oxford University Press.</p></li><li><p>the woman</p><p>P</p><p>NP</p><p>oot </p><p>the</p><p>l Ph</p><p>oten</p><p>P ro</p><p>Opinion TRENDS in Cognitive Sciences Vol.9 No.9 September 2005418utterance are formed by a unification operation. Thisoperation consists of linking up lexical frames withidentical root and foot nodes, and checking agreementfeatures (number, gender, person, etc.).</p><p>The resulting unification links between lexical framesare formed dynamically, which implies that the strength ofthe unification links varies over time until a state ofequilibrium is reached. Because of the inherent ambiguityin natural language, alternative binding candidates willusually be available at various points in the parsing</p><p>NP</p><p>det hd mod</p><p>DP N PP</p><p>man</p><p>P</p><p>hd </p><p>prep </p><p>with</p><p>Figure 2. Syntactic frames in memory (the mental lexicon). Frames are retrieved on</p><p>binoculars. DP: Determiner Phrase; NP: Noun Phrase; S: Sentence; PP: Prepositiona</p><p>object. Syntactic ambiguity can be seen for the root node PP (with), which has three p</p><p>lateral inhibition, one of these PP foot nodes gets selected for unification with the PDP</p><p>hd</p><p>art</p><p>NP</p><p>det hd mod</p><p>DP N PP</p><p>Root node</p><p>Fprocess (see Figure 2). Typically, one phrasal configurationresults, and this requires that among the alternativebinding candidates only one remains active. This isachieved through a process of lateral inhibition betweentwo or more alternative unification links. The outcome ofthe unification process is thus achieved via a selectionmechanism (i.e. lateral inhibition) that chooses betweendifferent unification options.</p><p>The advantage of this unification account is that: (i) it iscomputationally explicit, (ii) it is compatible with a largeseries of empirical findings in the sentence processingliterature, and in the neuropsychological literature onaphasia [6], and (iii) it belongs to the class of lexicalistparsing models that have found increasing support inrecent years [2,79].</p><p>Semantic and phonological unificationUnification operations take place not only at the syntacticprocessing level. Combinatoriality is a hallmark of langu-age across representational domains (cf. [2]). Thus, also atsemantic and phonological levels, lexical elements arecombined and integrated into larger structures. Anexample of semantic unification is the integration ofword meaning into an unfolding discourse representationof the preceding context. For instance, the majority ofcommon English words have more than one meaning</p><p>www.sciencedirect.com(see www.wordsmyth.net). In the interaction with thepreceding sentence or discourse context, the appropriatemeaning is selected, so that a coherent interpretationresults.</p><p>At the level of phonology, lexical elements are unifiedinto intonational phrases, which are the parts of speechthat are spanned by one intonational contour. Theintonational phrase ends with a high or low boundarytone and is further marked by pausing, lengthening orsegmental variation [10]. The characteristics of the</p><p> obj</p><p> NP</p><p>det hd mod</p><p>DP N PP</p><p>binoculars</p><p>TRENDS in Cognitive Sciences </p><p>basis of the word form input for the sentence The woman sees the man with the</p><p>rase; art: article; hd: head; det: determiner; mod: modifier; subj: subject; dobj: direct</p><p>tial connection sites (PP foot nodes in woman, sees andman). Through a process of</p><p>ot node.S</p><p>subj hd dobj mod</p><p> NP V NP PP</p><p>sees</p><p>nodeintonational phrase co-determine which aspects of theutterance get focus. This again, cannot be determined onthe basis of information retrieved from memory, butrequires an analysis of how lexical elements are unifiedinto phonological structures spanning a multi wordutterance [2,3].</p><p>Although explicit computational unification models areless well developed for semantics and phonology than forsyntax, unification is as relevant for semantics andphonology as it is for syntax. There are good reasons toassume that in language comprehension syntactic,semantic and phonological unification processes operateconcurrently and interact to some extent [2,6].</p><p>Neurobiological requirements for unification inlanguageThe need for combining independent elements into acoherent overall representation is not unique for languagecomprehension. It also holds for the visual system. Invisual neuroscience this is referred to as the bindingproblem. However, a major difference between objectperception and language comprehension is that visualbinding is more or less instantaneous (in the order of a fewhundred milliseconds; [11]), whereas language compre-hension is extended over time (in the order of seconds).Crucially, one core feature of the binding problem for</p></li><li><p>Electrophysiological recordings in the macaque monkey</p><p>the proposal that the contribution of LIFG to languageprocessing can be specified in terms of unificationoperations. In short, the left inferior frontal cortex recruitslexical information, mainly stored in temporal lobe struc-tures that are known to be involved in lexical processing[22], and unifies them into overall representations thatspan multi-word utterances.</p><p>For syntactic unification, supporting evidence comesfrom numerous PET and fMRI studies (cf. [23]). In a recentmeta-analysis of 28 neuroimaging studies, Indefrey [24]found two areas that were crucial for syntactic processing,independent of the input modality (visual in reading,auditory in speech). These two areas were the leftposterior superior temporal gyrus and the left posteriorinferior frontal cortex. The left posterior temporal cortex isknown to be involved in lexical processing. This part of thebrain might be important for the retrieval of the syntacticframes that are stored in the lexicon. I suggest that theunification space where individual frames are connectedinto a phrasal configuration for the whole utterance hasthe LIFG as a crucial node.</p><p>Opinion TRENDS in Cognitive Sciences Vol.9 No.9 September 2005 419language is how information that is not only processed indifferent parts of cortex (as in visual processing), but alsoat different time scales and at relatively widely spacedparts of the time axis, can be unified into a coherentrepresentation of a multi-word utterance. A neurobiologi-cally plausible model of language, therefore, presupposesthe availability of cortical tissue that is particularly suitedfor maintaining information on-line while unificationoperations take place. As we will see, prefrontal cortexseems to be especially well-suited for doing exactly this. AsI will argue below, Brocas area and adjacent cortex havethe right kind of neurobiological properties to play acrucial role in unification.</p><p>Brocas area extendedBrocas area is a crucial component in all classicalneurobiological models of language. However, there is noneuroanatomically motivated reason to restrict language-relevant frontal cortex to the ill-defined region that isgiven the name Brocas area.</p><p>Despite some disagreement in the literature (see [12]),most authors agree that Brocas area comprises Brodmannsareas (BA) 44 and 45 of the left hemisphere. At themacroscopic level these areas involve the pars opercularis(BA 44) and the pars triangularis (BA 45) of the thirdfrontal convolution. However, detailed cytoarchitectonicanalysis [13] shows that the borders of areas 44 and 45 donot neatly coincide with the sulci that were assumed toform their boundaries in macroscopic anatomical terms.More fundamentally, one has to ask what the justificationis to subsume these two cytoarchitectonic areas under theoverarching heading of Broca, rather than, say, Brod..</p></li></ul>