Modeling the meaning of words : Neural correlates of abstract and concrete noun processing

nouns Participant F44 Control Participant F58 Control Participant F25 Control Participant M78 Control glädje ‘joy’ kärlek ‘love’ barn ‘children’ lycka ‘joy’ sorg ‘sorrow’ tradition ‘tradition’ jul ‘Christmas’ jul ‘Christmas’ jul ‘Christmas’ följd ‘consequence’ början ‘beginning’ dagen ‘day’ start ‘start’ slut ‘end’ slutet ‘end’ ilska ‘anger’ oro ‘unease’ arg ‘angry’ raseri ‘rage’ aggression ‘aggression’ lojalitet ‘loyalty’ trogen ‘faithful’ godsint ‘kindhearted’ mod ‘courage’ gemenskapen ‘fellowship’ undantag ‘exception’ krig ‘war’ alla ‘all’ ibland ‘sometimes’ regel ‘rule’ depression ‘depression’ jag ‘me’ ledsen ‘sad’ psykofarmaka ‘psychopharmalogical drugs’ nedstämdhet ‘blues’ 474 F. Mårtensson et al. moral ‘morale’ etik ‘ethics’ etisk ‘ethic’ misstro ‘distrust’ etik ‘ethics’ reaktion ‘reaction’ snabbt ‘quick’ observant ‘observant’ illdåd ‘misdeed’ uppfattning ‘comprehension’ hat ‘hatred’ avsky ‘disgust’ avsky ‘disgust’ glädje ‘joy’ agg ‘grudge’ ärlighet ‘honesty’ positivt ‘positive’ trogen ‘faithful’ trogen ‘faithful’ ärlig ‘honest’ fördel ‘advantage’ bra ‘good’ förmån ‘benefit’ tårta ‘cake’ nos ‘nose’ osäkerhet ‘uncertainty’ olustigt ‘unpleasant’ tveksam ‘doubtful’ blygsam ‘modest’ vaghet ‘vagueness’ samvete ‘conscience’ rent ‘clean’ --tankar ‘thoughts’ rent ‘clean’ kris ‘crisis’ oro ‘worry’ katastrof ‘disaster’ livskris ‘crisis of life’ osäkerhet ‘insecurity’ passion ‘passion’ kärlek ‘love’ förkärlek ‘fondness’ kärlek ‘love’ lycka ‘joy’ förolämpning ‘insult’ otrevlig ‘unpleasant’ styggt ‘naughty’ vett ‘code of conduct’ motsägelse ‘contradiction’ ideal ‘ideal’ förebild ‘role model’ förebild ‘role model’ livsmål ‘aim in life’ synpunkt ‘viewpoint’ entusiasm ‘enthusiasm’ iver ‘keenness’ --sprallig ‘peppy’ ivrig ‘keen’ lydnad ‘obedience’ nödvändighet ‘necessary’ respekt ‘respect’ trofast ‘faithful’ respekt ‘respect’ kombination ‘combination’ möjlighet ‘possibility’ både ‘both’ lås ‘lock’ sammansättning ‘composition’ humör ‘mood’ glädje ‘joy’ glad ‘happy’ tankar ‘thoughts’ ilska ‘anger’ prestige ‘prestige’ kämpa ‘struggle’ --vilja ‘desire’ noggrann ‘scrutinous’ omständighet ‘circumstance’ situation ‘situation’ besvärligt ‘troublesome’ --svårighet ‘difficulty’ rädsla ‘fear’ obehag ‘unpleasantness’’ oro ‘worry’ olyckor ‘accidents’ skraj ‘harish’ ansvar ‘responsibility’ makt ‘power’ skyldighet ‘onus’ plikt ‘duty’ rädd ‘careful’ framtid ‘future’ tillförsikt ‘reassurance’ snart ‘soon’ nutid ‘present’ ålderdom ‘old age’ optimism ‘optimism’ positivt ‘positive’ glad ‘happy’ pessimism ‘pessimism’ glädje ‘joy’ fientlighet ‘hostility’ obehag ‘unpleasantness’ agg ‘grudge’ vänlighet ‘kindness’ aggression ‘aggression’ tendens ‘tendency’ förmåga ‘ability’ orsak ‘cause’ tyckas ‘seem’ tänker ‘thinks’ Neural correlates of word meaning 475 Appendix 2c: Word associations for the concrete nouns produced by aphasic participants Concrete nouns Participant 1 Anterior Participant 2 Anterior Participant 3 Anterior Participant 4 Posterior katt ‘cat’ mus ‘mouse’ hund ‘dog’ hund ‘dog’ kvinna ‘woman’ stol ‘chair’ bord ‘table’ bord ‘table’ bord ‘table’ sitta ‘sit’ krokodil ‘crocodile’ giraff ‘giraffe’ fanta* ‘elefant’ ödla ‘lizard’ trädgård ‘garden’ pincett ‘tweezers’ finne ‘pimple’ peang ‘forceps’ ögonbryn ‘eyebrow’ huvud ‘head’ klarinett ‘clarinet’ gitarr ‘guitar’ trumpet ‘trumpet’ spela ‘play’ instrument ‘instrument’ jordgubbe ‘strawberry’ mjölk ‘milk’ hallon ‘raspberry’ hallon ‘raspberry’ mat ‘food’ dragkedja ‘zipper’ knappar ‘buttons’ dragsko ‘~grommet’ jacka ‘jacket’ mage ‘belly’ makaron ‘macaroni’ pasta ‘pasta’ --pasta ‘pasta’ efterrätt ‘dessert’ fönster ‘window’ dörr ‘door’ karm ‘frame’ putsa ‘polish’ utsikt ‘view’ sköldpadda ‘tortoise’ bur ‘cage’ våreld ‘kalanchoe flower’ hav ‘sea’ --persika ‘peach’ frukt ‘fruit’ apelsin ‘orange’ banan ‘banana’ efterrätt ‘dessert’ termometer ‘thermometer’ Fahrenheit feber ‘fever’ kallt ‘cold’ temperatur ‘temperature’ stuga ‘cottage’ hus ‘house’ sommarnöje ‘summer relaxation’ skog ‘forest’ boning ‘dwelling’ leopard ‘leopard’ tiger ‘tiger’ geopard[sic] ‘cheetah’ savann ‘savannah’ djur ‘animal’ blomkål ‘cauliflower’ broccoli ‘broccoli’ rosenkål ‘brussels sprout’ usch ‘yuck’ mat ‘food’ fjäril ‘butterfly’ puppa ‘pupa’ --himmel ‘sky’ fågel ‘bird’ klänning ‘dress’ kjol ‘skirt’ kostym ‘suit’ fashion klädnad ‘clothing’ limousin ‘limousine’ taxi ‘taxi’ bil ‘car’ New York --smaragd ‘emerald’ diamant ‘diamond’ --safir ‘sapphire’ --mjölk ‘milk’ grädde ‘cream’ smörgås ‘sandwich’ laktolmjölk [sic] ‘lactose-free milk’ dryck ‘drink’ siden ‘silk’ tyg ‘fabric’ koppar ‘copper’ Indien ‘India’ plagg ‘garment’ 476 F. Mårtensson et al. choklad ‘chocolate’ godis ‘candy’ kola ‘toffee’ gott ‘tasty’ mat ‘food’ nejlika ‘carnation’ blomma ‘flower’ ros ‘rose’ jul ‘Christmas’ --revolver ‘revolver’ gevär ‘rifle’ skott ‘shot’ pistol ‘pistol’ uppskjutningsredskap* ‘~launching device’ järn ‘iron’ guld ‘gold’ koppar ‘copper’ rost ‘rust’ --varg ‘wolf’ björn ‘bear’ räv ‘fox’ Norrland --säckpipa ‘bagpipe’ instrument ‘instrument’ klarinett ‘clarinet’ Skottland ‘Scotland’ blåsinstrument ‘wind instrument’ gurka ‘cucumber’ tomat ‘tomato’ tomat ‘tomato’ grön ‘green’ mat ‘food’ förstärkare ‘amplifier’ högtalare ‘loudspeaker’ radio ‘radio’ högtalare ‘loudspeaker’ höjare* ‘~increaser’ vulkan ‘volcano’ vatten ‘water’ berg ‘mountain’ Etna --ek ‘oak’ träd ‘tree’ björk ‘birch’ golv ‘floor’ träd kamera ‘camera’ lins ‘lens’ foto ‘photo’ digital ‘digital’ tar bilder ‘takes pictures’ Appendix 2d: Word associations for the abstract nouns produced by aphasic participants Abstract nouns Participant 1 Anterior Participant 2 Anterior Participant 3 Anterior Participant 4 Posteriornouns Participant 1 Anterior Participant 2 Anterior Participant 3 Anterior Participant 4 Posterior glädje ‘joy’ surhet* ‘~grumpyness’ sorg ‘sorrow’ glad ‘happy’ sorg ‘sorrow’ tradition ‘tradition’ kontext ‘context’ glädje ‘joy’ stenbock ‘capricorn’ sed ‘custom’ början ‘beginning’ the end slut ‘end’ slut ‘end’ inledning ‘introduction’ ilska ‘anger’ arg ‘angry’ --arg ‘angry’ arg ‘angry’ lojalitet ‘loyalty’ upprättelse ‘redress’ --jag ‘me’ mycket lojal ‘very loyal’ undantag ‘exception’ expression --odd bekräftelse ‘confirmation’ depression ‘depression’ meditation ‘meditation’ fönster ‘window’ MAS sorglighet* ‘~sadness’ moral ‘morale’ etik ‘ethics’ tankeställare ‘eyeopener’ Lund gärningsåsikt* reaktion ‘reaction’ kontext ‘context’ sant ‘true’ stroke ‘stroke’ reagera ‘react’ Neural correlates of word meaning 477 hat ‘hatred’ ilska ‘anger’ ilska ‘anger’ MAS ilska ‘anger’ ärlighet ‘honesty’ människa ‘person’ sant ‘true’ lojal varar längst ‘lasts the longest’ fördel ‘advantage’ jobb ‘work’ sant ‘true’ logoped ‘speech therapist’ mitten ‘the middle’ osäkerhet ‘uncertainty’ skälvningar ‘tremblings’ orädd ‘unafraid’ feg ‘cowardly’ jag ‘me’ samvete ‘conscience’ moral ‘morale’ osant ‘untrue’ vän ‘friend’ förmåga ‘ability’ kris ‘crisis’ glädje ‘joy’ osant ‘untrue’ stress ‘stress’ svårighet ‘difficulty’ passion ‘passion’ NAMN (NAME) opassion* ‘unpassion’ röd ‘red’ önskan ‘wish’ förolämpning ‘insult’ mamma ‘mother’ osant ‘untrue’ mobba ‘bully’ nedsågande* ideal ‘ideal’ NAMN (NAME) sant ‘true’ --bästis ‘~best friend’ entusiasm ‘enthusiasm’ together --NAMN (NAME) entusiastisk ‘enthusiastic’ lydnad ‘obedience’ kris ‘crisis’ olydnad ‘disobedience’ hund ‘dog’ eftergöra* ‘~do as someone says’ kombination ‘combination’ arg och ledsen ‘angry and sad’ --dans ‘dance’ samband ‘relation’ humör ‘mood’ ups and downs ohumör*’unmood’ glad ‘happy’ ilska ‘anger’ prestige ‘prestige’ inte bra ‘not good’ --bra jobb ‘good work’ --omständighet ‘circumstance’ väljer ‘chooses’ --läkare ‘doctor’ orsak ‘cause’ rädsla ‘fear’ glad ‘happy’ orädd ‘unafraid’ inte prata ‘not to speak’ feghet ‘cowardice’ ansvar ‘responsibility’ föräldrarna ‘the parents’ oansvarig ‘irresponsible’ svara ‘answer’ veta ‘know’ framtid ‘future’ purples? orädd ‘unafraid’ arbeta ‘work’ imorgondag* ‘~tomorrowday’ optimism ‘optimism’ glada människor ‘happy people’ orädd ‘unafraid’ jag ‘me’ braåsikt* ‘~goodopinion’ fientlighet ‘hostility’ mamma ‘mother’ --ovänner ‘enemies’ ilska ‘anger’ tendens ‘tendency’ kulturen ‘culture’ otendens* ‘untendency’ målmedveten ‘goaloriented’ framåtskådande* ‘~forward looking’ 478 F. Mårtensson et al. Appendix 2e: Word associations triggered by food-related nouns jordgubbe ‘strawberry’ makaron ‘macaroni’ persika ‘peach’ blomkål ‘cauliflower’ mjölk ‘milk’ choklad ‘chocolate’ gurka ‘cucumber’ A grädde ‘cream’ spaghetti nektarin ‘nectarine’ soppa ‘soup’ kall ‘cold’ sött ‘sweet’ svalt ‘cool’ B hallon ‘raspberry’ spaghetti hallon ‘raspberry’ grönsak ‘vegetable’ mejeriprodukt ‘dairy product’ godis ‘candy’ grön ‘green’ C hallon ‘raspberry’ pasta frukt ‘fruit’ grönsak ‘vegetable’ bomull ‘cotton’ gott ‘tasty’ gott ‘tasty’ D smultron ‘wild strawberry’ spaghetti plommon ‘plum’ grönkål vegetable’ vatten ‘water’ kola ‘toffee’ tomat ‘tomato’ E bär ‘berry’ pasta frukt ‘fruit’ grönsak ‘vegetable’ grädde ‘cream’ godis ‘candy’ gott ‘tasty’ F äta ‘eat’ pasta luddig ‘fuzzy’ huvud ‘head’ kaffe ‘coffee’ dryck ‘drink’ grön ‘green’ G sommar ‘summer’ pasta sharonfrukt ‘persimmon’ grönsak ‘vegetable’ kor ‘cows’ rättvisemärkt ‘fair trade’ NAMN (NAME) H sommar ‘summer’ NAMN (NAME) len ‘smooth’ gott ‘tasty’ kaffe ‘coffee’ sött ‘sweet’ grön ‘green’ I god ‘tasty’ god ‘tasty’ mjuk ‘soft’ stuvning ‘stew’ vitt ‘white’ mörkt ‘dark’ vatten ‘water’ J sommar ‘summer’ pasta luden ‘hairy’ nyttigt ‘healthy’ ko ‘cow’ kexchoklad ‘chocolate wafer’ fräscht ‘fresh’ K sommar ‘summer’ hungrig ‘hungry’ len ‘smooth’ grönsak ‘vegetable’ kossor ‘cows’ kakao ‘cacao’ Västerås (Swedish pickle brand) L gott ‘tasty’ barn ‘child’ lent ‘smooth’ gott ‘tasty’ vitt ‘white’ gott ‘tasty’ tsatsiki 1 mjölk ‘milk’ pasta frukt ‘fruit’ broccoli grädde ‘cream’ godis ‘candy’ tomat ‘tomato’ 2 hallon ‘raspberry’ --apelsin ‘orange’ rosenkål ‘brussels sprout’ smörgås ‘sandwich’ kola ‘toffee’ tomat ‘tomato’ 3 hallon ‘raspberry’ pasta banan ‘banana’ usch ‘yuck’ laktolmjölk [sic] ‘lactose-free milk’ gott ‘t

higher level semantic categories.Empirical evidence will be presented which supports the assumption that frontal and temporoparietal brain areas are crucial for the processing of general semantic frames, whereas occipital areas are relatively more involved in the processing of concrete nouns whose meaning is dependent on sensory semantic (e.g.visual) features.The evidence comes from the results of a word association test involving both healthy and aphasic subjects.The basic assumption behind the study is that damage to occipital regions of the brain would lead to problems associating to concrete words whereas lesions of the 'language areas' around the Sylvian fissure, including inter alia Broca's and Wernicke's areas, would lead to problems associating to abstract words.

neurocognitive models of meaning representation
To get a better understanding of the representation of word meaning in the brain, it is important to relate linguistic models of word semantics to a more general framework of neurocognitive processing.One of the most influential models of neurocognitive information processing is that of Fuster (2009), who represents dif-ferent kinds of meaning in distributed neural networks 1 (Fig. 1).Although semantic networks can be formed between neurons in virtually any parts of the brain, lower-level, perceptual networks have nodes in posterior sensory cortices related to visual, auditory, and tactile processing [dark blue areas (1-3, 17, 41) behind the Rolandic Fissure (RF) in Fig. 1].Similarly, motion activates the primary motor cortex [dark red area (4) directly in front of RF in Fig. 1].The greater the complexity of a perception or memory, the more widely distributed neural activity it is expected to be associated with.Higher-level, executive processing is more dependent on frontal executive networks (white on red in Fig. 1) whereas higher-level semantic conceptualization activates temporoparietal and inferior temporal areas involved in integration of different sensory modalities (whiter areas behind RF in Fig. 1) (Fuster 2009).
Numerous studies have provided evidence that abstract and concrete words 2 are represented and processed differently in the brain.Thus, neuroimaging data support a greater involvement of sensory (e.g.visual) areas in concrete word processing (Martin et al. 1996, Humphreys et al. 1997, Pulvermüller 1999) and a more focal activation of perisylvian 'language' areas 3 for function words as well as abstract nouns (Pulvermüller 1999, Sabsevitz et al. 2005, Noppeney and Price 2004, Binder et al. 2005).
Further support for the idea that a perisylvian network is more involved in abstract word processing comes from studies involving speakers with aphasia.Aphasic speakers with frontal/temporoparietal lesions are generally found to have greater difficulty in processing abstract words (Hagoort 1998, Tyler et al. 1995), whereas posterior 4 damage can be associated with a selective impairment in the naming of visually presented objects and colors (Girkin and Miller 2001, Gainotti 2004, Coslett and Saffran 1992, Forde et al. 1997, Luzatti 2003).
The idea that concrete word processing is more sensory-based has often been interpreted in terms of the 'dual coding theory' (Paivio 2007), which assumes abstract words to mainly activate a language-specific Fig. 1.Hierarchical organization of semantic networks in anterior and posterior cortices (Fuster 2009) (reproduced by kind permission of MIT Press Journals) system, whereas concrete words additionally activate an imagery system5 .This view is compatible with the hierarchical model of Fuster (2009), but the latter adds the dimension that higher-level semantic networks are involved in the processing of abstract information in general, not just language-specific information as in the verbal system described by Paivio.

integrating semantic features and frames in a hierarchical model of word meaning
In this section, it will be shown how modeling of word meaning in terms of semantic features, categories and frames can be related to the hierarchical model of semantic networks described above, with concrete sensory input and motor output at the lowest level, and progressively more abstract frame-related knowledge at higher levels.
In Table I, a proposal is made for how cognitive linguistic information at different levels of abstraction can be related to the hierarchy of semantic networks in Fuster's (2009) model of neurocognition.In what follows, a description of the linguistic information at the different levels in the hierarchy is presented.

Perceived/produced language
This level in Fuster's general model of neurocognition corresponds to the actual input in language perception and the motor output in language production (Table I).

Semantic features
The term 'semantic features' (Weinreich 1966) has frequently been used to describe concrete, sensory-motorbased properties (BANANA: +curved +yellow) and the term is used in this sense in the present study.The processing of basic sensory features such as size, color, shape, movement pattern, sound, taste and smell, activates modality-specific semantic networks at the 'phyletic sensory' level, which are then integrated at the 'polysensory' level in order for events and objects in the environment to be perceived as a seamless whole (Fuster 2009).Following this, sensory-motor semantic features (level 1 in Table II) can be seen as having neural correlates in terms of sensory-motor regions in the brain (level 2 in Table I).A concrete subordinate level word such as STRAWBERRY refers to an object associated with rather specific, sensory-based features and feature constellations, making the lexical form likely to be associated with neural meaning representations at these levels.Semantic categories Rosch andcolleagues (1976, 1978) proposed that categorical semantic relations are hierarchically ordered, with superordinate (FOOD), basic (BERRY) and subordinate (STRAWBERRY) terms.The cognitive categorization of entities and objects has been suggested to be based on certain essential features (Ungerer and Schmid 2006).Sensory-motor based semantic features (e.g.color, shape, smell) are shared to different degrees by words at different hierarchical levels.The higher the level, the less concrete a word is.In other words, BERRY is less concrete than STRAWBERRY, but their referents still share sensory features (e.g.'small', 'sweet').This can be seen as corresponding to the polysensory level in the model of Fuster (2009) (level 3 in Table I) and to be important for word meanings at levels 2 and 3 (sensory-motorbased categories and frames) in Table II.Berries can also be part of categories that are even more abstract and general, such as the superordinate terms FOOD or DESSERT, corresponding to level 5 (semantic/conceptual level) in Table I and level 5 (general semantic frame) in Table II.A mental representation of FOOD may or may not include basic sensory semantic features associated with BERRY.Members of the cognitive category FOOD have the common denominator of 'edibility', but these edible things can have a wide variety of tastes, smells, textures, shapes and colors.In a recent study (Crutch and Warrington 2008), subjects with left middle cerebral artery stroke aphasia were shown to have processing advantages for subordinate level words (e.g.DALMATIAN), in contrast to healthy controls who process basic level terms (e.g.DOG) with greater ease, and semantic dementia patients who are more proficient with superordinate terms (e.g.ANIMAL).One possible explanation for this pattern might be the greater variability of semantic information associated with superordinate categories as compared to the more detailed sensory features defining the referents of subordinate terms.The more diverse and context-dependent the semantic content of a word is, the greater are the demands on controlled access of information.Thus, one suggested reason for this double dissociation is that executive functions are affected by stroke aphasia, whereas detailed semantic representations are degraded in semantic dementia (Jeffries and Lambon Ralph 2006).In particular, superordinate and basic level categories may lack the consistent linkage to sensory semantic features that is present in more specific, subordinate words.Following the models of Fuster (2009) and Paivio (2007), if a subordinate level word, e.g.DALMATIAN is more strongly associated with specific visual features as compared to the basic level word DOG, this would make it easier to access when executive and linguistic processes are impaired but visually related semantic processes are intact.

Semantic frames
The existence of higher level knowledge structures is essential for understanding word meaning, in particular when it comes to abstract words.In this view, more abstract semantic knowledge activates such knowledge structures, which in Fillmore's terms are called semantic frames.For example, the word BACHELOR evokes both abstract semantic properties such as 'male' and 'unmarried' and a semantic frame consisting of encyclopedic knowledge, e.g. the cultural knowledge which makes it possible for us to know that the pope is not a typical bachelor although he is an unmarried man (Fillmore 1982, Lakoff 1987).The less grounded in concrete, sensorimotor experience a word's meaning is, the more integration of diverse, frame-based information the processing of the word can be hypothesized to involve.In neurological terms, accessing more abstract semantic representations would, according to the model outlined by Fuster (2009), be likely to activate higher-level networks in e.g.frontal and temporoparietal areas to a greater extent than concrete word representations.In the pres-ent study, two levels of frame structure will be assumed, a 'personal frame level' and a 'general frame level', further described in the following paragraphs.

Personal semantic frames
This level corresponds to Fuster's 'episodic' level (level 4 in Table I), where autobiographical experiences are represented (Fuster 2009).Although the complexity and richness of memory representations increase from the phyletic sensory information level to personal frames consisting of memories of whole scenarios, these levels of representation share the property that they reflect things that have previously been directly perceived in the environment and experienced with the body.We assume, therefore that specific semantic frames associated with personal experiences (level 4, Table I) are to be distinguished from more general semantic frames (level 5, Table I).Personal frames can be hypothesized to activate a large amount of both sensory and emotional information in the brain, making them easier to access even when areas involved in the processing of general frames are damaged.
Right hemisphere cortical regions (e.g.ventromedial prefrontal cortex) and subcortical areas (e.g. the amygdala) have been found to be more active in the processing of socially related concepts as compared to information involving manipulable objects (Martin and Weisberg 2003).Right hemisphere regions have also been shown to be involved to a larger extent in subjects who respond quickly to an emotional Stroop test, suggesting that the right hemisphere is important for directing attention to task-relevant emotional stimuli (Mincic 2010).On the basis of these results, pragmatic and emotional aspects of word associations could also be thought to involve these brain regions to a higher degree, allowing these types of information to be accessed despite lesions in left hemisphere cortical areas.

General semantic frames
The more abstract and general levels above the personal, episodic level correspond to the level of 'general semantic frames' (level 5, Table I).The more diversely distributed semantic networks corresponding to higher, conceptual levels are "formed to a large extent by the repeated coactivation and instantiation of similar, more concrete (e.g.episodic) memories" (Fuster 2009(Fuster , p. 2062)), which are nested within the larger networks (level 5, Table I).Superordinate categories for concrete words that are not directly associated with sensorymotor semantic features are also assumed to be localized on this level.

Connections between frames and features
Following the connectionistic model of information processing presented in Table I, it is assumed that frame-based representations can be associated with each other as well as directly with sensory features.For example, concrete words such as 'strawberry' can be related directly to features on the lowest phyletic sensory level (e.g. the color 'red') as well as to semantic category related words (e.g.'berry').Further, an association from JORDGUBBE ('strawberry') to GRÄDDE ('cream') involves a connection to a non-categorically related word instead of a more direct sensory-based relation with another category-related word as in the relation between JORDGUBBE ('strawberry') and SMULTRON ('wild strawberry').Words contextually associated to each other but which do not have overlapping sensory features (e.g.JORDGUBBE ('strawberry') and SOMMAR ('summer') are assumed to be related at other, higher levels of common associated frames (Fillmore 1985, Fortescue 2010).

combining features and frames: a word association experiment using concrete and abstract words
In order to obtain experimental support for models of word meaning combining features and frames, word relations in the mental lexicon were investigated using a word association test.Since features and frames are also assumed to be differently involved in the processing of abstract and concrete words, the test compared the processing of these two different lexical categories.In order to obtain information on the distribution of neural networks involved in the processing of feature and frame-based information, the association test was given to healthy speakers as well as to aphasic speakers with perisylvian or occipital lesions.
Word associations can be expected to reflect the different kinds/levels of semantic information processing (see Table II).Associations might be related to both features and frames in different ways: (1) Associations focusing on a single sensory-based feature, e.g.STRAWBERRY-RED, (2) Associations sharing several sensory features with the cue word, e.g.STRAWBERRY-RASPBERRY, (3) Associations related to a frame-related word having sensorimotor features, e.g.STRAWBERRY-CREAM, (4) Associations related to a personal frame, e.g.STRAWBERRY-GRANDMOTHER, (5) Associations involving a more general frame, e.g.STRAWBERRY-SUMMER.These association types are summarized in Table II.
The free word association task makes it possible to investigate the association strength between the different word types and semantic feature-and frame-based information.

hypotheses
On the assumption that semantic features related to lower level semantic networks (phyletic sensory and polysensory), mostly dependent on posterior cortices, would be well-preserved in speakers with perisylvian lesions, it was hypothesized that associations involving shared sensory-based semantic features would be more common in this group than in speakers with occipital lesions, as well as speakers without aphasia.It was further hypothesized that speakers with perisylvian lesions would have difficulty in accessing higher-level, general frame-based associations, both due to the general frames' lack of a direct relation to sensory information and due to the greater demands selecting a word from a general frame may put on executive functions.Conversely, since speakers with occipital lesions could be expected to have problems processing low-level visual features and thus concrete words in general, it was hypothesized that associations involving higherlevel (more abstract) words would more likely be produced by such speakers than by speakers with perisylvian lesions and speakers without brain damage.

methods subjects
Twelve Swedish speakers (ten female) without communication problems in the age range of 23-79 years (M = 47; SD = 19) participated.Contact with four right-handed subjects with left-hemisphere lesions (three female), (Table III) was established through the Neurology Department at Malmö University Hospital and the Aphasia Association in Malmö-Lund.They had all been treated at the Stroke Clinic at Malmö University Hospital.The healthy and aphasic subjects were approximately matched as regards their age and education levels and all subjects gave their informed consent to participate in the study.The subjects with perisylvian lesions all had symptom diagnoses of mild-to-moderate Broca's aphasia, and the fourth subject with occipital lesions6 had mild anomic aphasia including semantic dyslexia.Diagnoses were made based on Swedish equivalents of the Boston Diagnostic Aphasia Examination (Apt 1997), the Boston Naming Test (Kaplan et al. 1983, Apt 1999), and the Token Test (De Renzi and Faglioni 1978).(Coltheart 1981).The words with high concreteness and imageability values can be assumed to be strongly related to the lower sensory-motor levels of processing, whereas the words with low concreteness and imageability values can be assumed to involve higher-level, more abstract conceptual processing.Concreteness and imageability values range from 100-700, with 100 being the least concrete/imageable and 700 the most concrete/imageable.The English nouns from the database were translated to Swedish using Lexin (2007).
The nouns used in the present study had 1-4 syllables (M = 2.55, SD = 0.922), to minimize the risk that word length could affect comprehension negatively for the aphasic participants.The mean number of syllables in concrete nouns (M = 2.43, SD = 0.817) and abstract nouns (M = 2.67, SD = 0.922) did not differ significantly t 58 =−1.037,P=0.304.
Written word frequencies from the Stockholm Umeå Corpus (SUC) (Ejerhed et al. 1992) were obtained for the Swedish translations of the nouns.The abstract nouns had a higher mean frequency M = 58.93,SD = 76.368than the concrete nouns M = 11.83,SD = 25.551,t 58 =−3.204,P<0.01.Familiarity values 9 , also obtained from the MRC database (Coltheart 1981), did not differ significantly for abstract nouns M = 526, SD = 38 and concrete nouns M = 503, SD = 56, t 58 =−1.872,P=0.066, although there was a trend for higher familiarity of the abstract nouns. 10procedure In the word association test, participants were instructed to say the first word that the word uttered by the test leader made them think of, and they were informed that any answer would be correct as long as it was the first word that came to mind.Prior to the test, two practice words, KATT ('cat') and STOL ('chair'), were presented in order to verify that the participants had understood the instructions correctly.Thirty concrete and 30 abstract nouns were then presented orally by the test leader.In the association task, every second noun presented was abstract.The same order of presentation was used for all participants. 11Pauses and repetitions of the nouns were made when necessary.The speakers' associations were recorded using a Marantz PMD660 Portable Solid State Recorder and an IMG Stage Line ECM-302 B Boundary Microphone.

Coding of semantic relations
The five types of word associations listed in Table II, corresponding to different kinds of semantic relationships (1-5) were labeled.These 5 types of 9 Familiarity values are based on subjects' ratings of how familiar, i.e. frequently experienced, words are. 10 Spoken word frequencies from the Brown Corpus for the English translations of the words were also looked up in the MRC database.Although relatively few of the words were found in the corpus (approximately 30% of the concrete words and 60% of the abstract words), nevertheless, judging from this small sample, abstract words also seemed to be more frequent in spoken word production.These differences in written and spoken word frequencies as well as subjectively judged familiarity with the abstract nouns being more common is probably due to the fact that they express more general concepts which can be used in a variety of contexts, whereas the concrete words refer to relatively more specific concepts. 11The reason for not presenting the stimuli in randomized order was that it was deemed that if not all aphasic participants could complete the whole test, there would still be a data set consisting of the same words and an equal number of concrete and abstract trials left for analysis.
semantic relations can also be seen as associated with five different levels of semantic representations, with level 1 being the least abstract and level 5 being the most abstract.In addition, labeling of perseverations, derivations, absent and unidentifiable associations (6-9), were coded in the material.These latter categories can all be considered to be "unacceptable" associations in the sense that they contain no new semantic information related to the cue words.Coding was carried out by the first author.All labeling was discussed with, and agreed on by the second and third authors.

Statistical analysis
Statistical analyses were performed using Crawford Statistics; more specifically, with the software Singlims 12 .In the first analysis, mean scores were calculated based on the levels (categories 1-5) of the semantically acceptable associations.Mean scores were calculated for all groups (perisylvian, occipital, control) for the responses to concrete and abstract cue words, respectively.
In the second analysis, the number of acceptable associations (max: 30 for concrete cue words and 30 for abstract cue words) produced by occipital and perisylvian subjects were compared with those of the healthy controls.However, this analysis could only be carried out for the associations to the abstract cue words, since a requirement in Crawford statistics is that there be some variability in the material.Contrary to this requrement, the control group produced 100% acceptable associations for concrete cue words, leading to a standard deviation of 0.

resuLts distribution of semantic relations between cue words and associated words
The data, coded according to the previously described labels, was analyzed in order to determine the distribution of the different kinds of relations found between test words and associated words produced by the test participants.Some examples of associations and their relation to cue words are illustrated in Table IV.
In the healthy subjects, concrete cue words gave rise to a large number of sensory-motor based asso-

Absent
The subject is unable to produce any association.

Unknown
Association which could not be categorized as any of the types 1-6.
ciations, most of which were categorically related based on sensory semantic features or belonged to a related sensory-based semantic frame (54%).However, a large proportion of the associations produced as responses to concrete cue words by the healthy subjects were general semantic frame based (27%, the next largest category) (Table V).In the perisylvian aphasic group, sensory-motor based categorical and sensory-motor frame based associations constituted an even larger proportion of the associations (78%), whereas general semantic frame associations were fewer than in controls (only 12%).
The subject with occipital lesions produced fewer sensory-motor based associations (17%), whereas the largest proportion (50%) of his associations were general semantic frame-based.In general, he had difficulties with concrete word associations, and was even unable to produce associations for 23% of the concrete cue words.Abstract cue words naturally never led to associations based on sensory-motor features or categories as defined in the present study 13 (Table VI).Thus, for the 13 i.e. associated words involving perceptually observable semantic features shared by the cue words.abstract cue words, all associations were considered to be semantic frame-based.In the healthy participants, 83% of the associations for abstract cue words fell into the general semantic frame category.In the perisylvian aphasic group, a dramatically lower proportion of the associations (39%) were based on general semantic frames.Of the remaining associations to test words, many were personal frame associations (22%), perseverations (14%), or even absent (13%), patterns which did not occur for the concrete cue words in the perisylvian group and not for any words in the healthy subjects.The subject with occipital lesions, in a manner similar to the controls, produced mainly general semantic frame associations for abstract words (73%).

number of acceptable associations for abstract and concrete cue words between groups
The analysis of the number of acceptable (semantically related) associations showed that for the abstract cue words, both clinical groups produced fewer acceptable associations than controls (M controls = 28.75;SD = 0.2) The perisylvian group's test score (M = 19.67)was significantly lower than that of the controls (t=−43.619,P<0.001) as was the score of the

mean category level of the responses for abstract and concrete cue words between groups
Figure 2 shows the mean category level of the responses for concrete and abstract cue words in the different groups (controls, perisylvian, occipital).As can be seen in Fig. 2, the subject with occipital lesions produced associations on the highest level of abstraction, both when cue words were concrete (M = 4.35) and abstract (M = 4.96).Controls were intermediate in the overall level of association with associations being at low levels for concrete cue words (M = 3.07) and at high levels for abstract cue words (M = 4.76).The perisylvian subjects were similar to controls but at an overall lower level of abstraction for both concrete cue words (M = 2.79) and abstract cue words (M = 4.53).In the responses for concrete cue words, the mean category level of the associations was significantly higher for the occipital subject (M = 4.35) compared to controls (M = 3.07, SD = 0.073) (t=16.846,P<0.001) and in contrast, significantly lower in the perisylvian group (M = 2.79) (t=−3.685,P<0.05).In the responses for abstract cue words, the mean category level of the occipital subject (M = 4.96) did not differ significantly from that of the controls (M = 4.76.SD = 0.627) (t=0.306,P>0.05) and neither did the mean category level of the perisylvian group (M = 4.53) (t=−0.352,P>0.05).

discussion
The qualitative analysis of the word association data revealed differences between the participant groups regarding the distribution of the semantic relationships between cue words and associations.In the healthy subjects, a very large proportion of the associations given as responses to abstract nouns, as well as a considerable number of the concrete word associations were words belonging to general semantic frames, e.g.STRAWBERRY-SUMMER, i.e. associations which were not directly associated with any sensory-motor-based features related to the cue word.In the perisylvian aphasic group, abstract nouns triggered repeated associations to words related to personal semantic frames, e.g.REACTION-STROKE whereas concrete nouns mostly gave rise to sensory-based word associations, in particular concrete words sharing sensory-based semantic features with the cue words, e.g.CHAIR-TABLE .The subject with occipital lesions showed the opposite pattern, with mainly general semantic frame-based word associations for abstract as well as concrete cue words and very few sensory/visual-based associations.The results obtained support the hypothesis that concrete and abstract words activate partly distinct neural networks with nodes in different parts of the brain.The results also support a model of the mental lexicon in which semantic features and semantic frames are integrated in word meaning processing and show how such a model fits into a general account of neurocognitive information processing (Fuster 2009).As regards the statistical analysis, some of the results reached significance and some did not.Surprisingly, all aphasic subjects produced significantly fewer acceptable associations for abstract cue words although this was only Fig. 2. Mean semantic category levels produced as responses by healthy controls, perisylvian aphasic, and occipital aphasic subjects for concrete and abstract cue words.The degree of abstraction increases from 1 to 5, corresponding to the levels in Table I.
expected for the perisylvian aphasics.A possible explanation is that the occipital subject produced a number of derivations, e.g.LOYALTY-LOYAL; REACTION-REACT, (category 7, see Table VI and  Appendix 2d), which were considered to be unacceptable associations, as well as a number of associations where the semantic relation to the cue word was unknown e.g.TWEEZERS-HEAD, (Appendix 2c).In the perisylvian group, the failure to associate was more often due to perseveration or omitted associations.As regards concrete cue words, the number of semantically acceptable associations for the perisylvian group (28.3/30) was higher than that for the occipital subject (20/30).
The clearest statistical difference was seen in the mean category levels of the associations for concrete cue words, where a significant double dissociation was seen between occipital and perisylvian subjects as compared to controls, with the occipital subject scoring higher and the perisylvian subjects lower.As regards the associations for abstract cue words, the differences were not significant in any direction, indicating that when the subjects were able to produce acceptable responses for abstract words, they were generally rather high-level associations.

neural correlates of concrete word meaning
Since the occipital aphasic subject's lesions affects areas involved in primary visual [in the terms of Fuster (2009) 'phyletic sensory'] processing, his difficulties with concrete nouns can be thought to be caused by a failure to access semantic networks involving low-level visual feature representations.Word access from visual input, e.g.naming objects and colors from visual presentation, has previously been shown to be selectively impaired in persons with aphasia due to left occipital lesions14 .In the present study, occipital lesions were associated with specific problems accessing concrete nouns from verbal presentation, supporting the assumption that posterior areas involved in processing of visual information, e.g.colors and shapes, are intimately involved in the processing of concrete nouns, not only when they are activated by visually presented stimuli.This indicates that lowlevel semantic networks involved in the perception of concrete objects and entities are crucial for the processing of the concrete words denoting them, even in the absence of visual presentation or explicit mental imagery.Thus, vision-related and other sensory representations are a likely contributing factor to the processing advantage normally present for concrete as compared to abstract nouns, as suggested by e.g.Paivio (2007) and Pulvermüller (1999).The involvement of phyletic sensory, visual semantic networks in the occipital cortex also makes concrete noun processing likely to be relatively well preserved in aphasia caused by left perisylvian lesions.In the present study, this is manifested as the production of a larger proportion of sensory-based associations, in particular words denoting objects sharing sensory-based features with the cue word (Table V; Fig. 2, for examples see Appendix 2c).It has been suggested that concrete and abstract words also differ in their involvement of the two hemispheres, with concrete words and imagery showing a right hemisphere advantage (e.g.Paivio 2010).Binder and others (2005) found that processing of concrete words to a greater extent activated bilateral association cortices, whereas abstract word processing was more localized to the left hemisphere.Even if it is the case that concrete word processing normally activates bilateral areas, the fact that the subject with left occipital lesions in the present study is impaired in concrete word processing suggests that the spared representations in the right hemisphere are not sufficient to allow normal word associations for concrete words when access to low-level visual features is disrupted.
It has been suggested that some more abstract discourse-related aspects of word meaning could be processed in the right hemisphere.In the present study, the speakers with left perisylvian lesions had more problems with abstract words.A possible explanation for this could be that the right hemisphere has difficulty accessing associations to decontextualized abstract cue words.It could, however, also be the case that whereas the left hemisphere is more involved in the processing of lexical meaning, the right hemisphere is involved in association to pragmatically related aspects of word meaning which perhaps cannot be interpreted without access to information in the left hemisphere (Fortescue 2010).

neural correlates of abstract word meaning
The fact that aphasic subjects with perisylvian lesions showed more atypical patterns of abstract word processing is in line with Fuster's (2009) assumption that the distribution of neuronal activity is more cen-tered in frontal and temporoparietal semantic networks in higher, more abstract levels of processing.Accessing pragmatic and linguistic context (semantic frames) puts greater demands on executive and linguistic processes mediated by perisylvian brain areas.Although the test nouns elicited both categorical and contextual associations in both aphasic and control subjects, the pattern of consistently staying within an episodic memory context was only observed in the perisylvian aphasic group (Table VI, Fig. 2, Appendix 2d).The explanation for this may be that when access to general semantic frames is impaired by lesions in cortical networks important for higher-level processing, personal semantic frames can still be accessed due to their perceptual and emotional content, processed by networks in partly different areas of the brain, e.g.cortical sensory regions and emotional (subcortical) areas.Another aspect of the repeated associations to the same autobiographical context is that it may be considered as a form of perseveration behavior.It is possible that impairments in executive processing in the subjects with lesions in the perisylvian network may have made the personal contexts difficult to inhibit once they were activated.Accessing semantic frame-based word associations, rather than perceptually similar concepts, puts greater demands on selecting the contextually appropriate information and keeping it in working memory while at the same time inhibiting irrelevant information.There is neurobiological support for the idea that selection processes are mediated by anterior, e.g.prefrontal brain areas (Scheler 1999).

hierarchical categorical relationships and abstraction: general word association patterns in brain-damaged speakers
The subjects with aphasia due to perisylvian lesions mainly produced categorically related words which were on the same, subordinate, visually detailed level as the concrete cue words, whereas the subject with aphasia resulting from an occipital lesion mainly produced categorically related associations which were on a lexically superordinate, more abstract conceptual level and were not related to any specific sensorybased semantic features.For example, several of the food items elicited the unspecific association 'food' by the speaker with occipital lesions, while the other participants' associations were words describing the type of food, its properties or other specifically associated concepts (Appendix 2e).Producing a superordinate instead of a basic level term as response to the subordinate level concrete cue words was very rare in any of the other participants, including the controls.
This contrasting pattern between subjects with perisylvian and occipital stroke aphasia is similar to findings in previous studies comparing subjects with stroke aphasia and semantic dementia (Crutch and Warrington 2008), with occipital lesions being associated with a pattern similar to that found in semantic dementia.Meaning representations dependent on an amodal semantic hub in the anterior temporal lobe have been suggested to be degraded in semantic dementia, leading to the loss of detailed semantic information, whereas executive impairments are the cause of semantic impairments in stroke aphasia (Jeffries andLambon Ralph 2006, Patterson et al. 2007).In the present study, differences were found between speakers with different types of stroke aphasia.Given the fact that the occipital lobe processes modality-specific (visual) information and that occipital lesions are unlikely to affect executive function, the semantic impairment in occipital aphasia may be a result of a failure to access modality-specific visual meaning representations.

concLusions
The present study presents evidence supporting lexical semantic models integrating different kinds of meaning representation.It shows how this kind of linguistic modeling can be related to a general model of information processing assuming different levels of semantic networks.The results from the word association test are consistent with the assumption that different kinds of associated semantic networks are important for the processing of concrete as compared to abstract nouns -more specifically, sensory-based, e.g.visual semantic features for concrete nouns and general semantic frames for abstract nouns.Access to, and coordination of, the more diverse meaning networks related to higher-level semantic frames is probably more dependent on executive function.When access to higher, more abstract levels of representation is hampered by anterior lesions, word processing becomes more dependent on sensory-based information, reflected in the present study as retrieval of representations of perceptually similar objects when associating to concrete nouns.In the case of abstract words, which are not clearly associated with sensory experiences, personal semantic frames based on episodic memories, which are more concrete, are accessed when abstract nouns are given as input and general semantic frames are unavailable.Conversely, damage to the visual cortex may leave higher-level conceptual levels of representation for abstract words well preserved, but cause problems accessing representations at the lower, sensory-based levels and thus impair concrete word access.

Table I
Fuster (2009)tions between different levels of cognitive linguistic information (third column) and the levels in the general neurocognitive model byFuster (2009)(see also Figure1)

Table III
Aphasic participants with left hemisphere lesions extending cranially dorsally over the Sylvian fissure at the height of the cella media of the lateral ventricles.Low attenuation also included the insula area and the major part of the external capsule.MR: Lesions in large parts of the irrigation areas of the middle cerebral artery both frontally and temporally.

Table IV
Labeling of the relations between cue words and associations with examples from the data

Table V
Distribution of association types for concrete cue words produced by healthy, perisylvian aphasic, and occipital aphasic subjects

Table VI
Distribution of association types for abstract cue words produced by healthy, perisylvian aphasic, and occipital aphasic subjects Appendix 2a: Word associations for the concrete nouns produced by healthy participants