<<

Running head: MULTIPLEX LEXICAL NETWORK 1

1Quantifying the interplay of semantics and phonology during failures of word retrieval by people

2 with using a multiplex lexical network

3

4 Nichol Castro*1,2, Massimo Stella*3,4, and Cynthia S. Q. Siew5

5

6 1Georgia Institute of Technology, 2University of Washington, 3University of Southampton,

7 4Complex Science Consulting, 5National University of Singapore

8

9

10

11 Author Note

12 Nichol Castro, Department of Speech and Hearing Sciences, University of Washington;

13Massimo Stella, Complex Science Consulting, Italy; Cynthia S. Q. Siew, Department of

14Psychology, National University of Singapore. *These authors contributed equally.

15 This research was supported in part by a Ruth L. Kirschstein NRSA Institutional

16Research Training Grant from the National Institute on Aging (T32AG000175) and a Ruth L.

17Kirschstein NRSA Individual Fellowship Grant from the National Institute on Deafness and

18Other Communication Disorders (F32DC017074) to Nichol Castro, an EPSRC Doctoral

19Training Centre grant (EP/G03690X/1) to Massimo Stella, and an Overseas Postdoctoral

20Fellowship from the National University of Singapore to Cynthia S. Q. Siew.

21 Correspondence concerning this article should be addressed to Nichol Castro,

22Department of Speech and Hearing Sciences, University of Washington, [email protected] (or)

23Massimo Stella, Complex Science Consulting, Italy, [email protected]

24 Abstract MULTIPLEX LEXICAL NETWORK 2

25Investigating instances where lexical selection fails can lead to deeper insights into the cognitive

26machinery and architecture supporting successful word retrieval and speech production. In this

27paper, we utilized a multiplex lexical network approach that combines semantic and

28phonological similarities among words to model the structure of the mental . Network

29measures at different levels of analysis (degree, network distance, and closeness centrality) were

30used to investigate the influence of network structure on picture naming accuracy and errors by

31people with Anomic, Broca’s, Conduction, and Wernicke’s aphasia. Our results reveal that word

32retrieval is influenced by the multiplex lexical network structure in at least two ways – (i) the

33accuracy of production and error type on incorrect productions were influenced by the degree

34and closeness centrality of the target word, and (ii) error type also varied in terms of network

35distance between the target word and produced error word. Taken together, the analyses

36demonstrate that network science techniques, particularly the use of the multiplex lexical

37network to simultaneously represent semantic and phonological relationships among words,

38reveal how the structure of the mental lexicon influences language processes beyond traditionally

39examined psycholinguistic variables. We propose a framework for how the multiplex network

40approach allows for understanding the influence of mental lexicon structure on word retrieval

41processes, with an eye toward a better understanding the nature of clinical impairments like

42aphasia.

43

44

45 MULTIPLEX LEXICAL NETWORK 3

461. Introduction

47 Word retrieval failures and speech errors do not occur at random. A close examination of

48such retrieval failures is vital for understanding normal and impaired speech and language

49processes (Baars, 1992). Indeed, word retrieval failures and speech errors are generally viewed,

50in part, as by-products of disruptions in the spreading activation process in the mental lexicon

51(e.g., Dell, Schwartz, Martin, Saffran, & Gagnon, 1997; Foygel & Dell, 2000; Levelt, Roelofs, &

52Meyer, 1999). Although errors may be due to disruptions in the process of spreading activation,

53it is important to also consider how the structure of the mental lexicon influences that process.

54We argue that gaps in our understanding of word retrieval, and particularly instances of failures,

55will continue to persist if investigations of spreading activation processes do not include a careful

56examination of the mental lexicon structure. This view is in line with several researchers who

57have pointed out that current limitations in fully understanding the mental lexicon and language

58processes are in part due to our inability to clearly define the structure of the mental lexicon

59(Jones, Willits, & Dennis, 2015; Sousa & Gabriel, 2015; Steyvers & Tenenbaum, 2005).

60Therefore, this paper adds to the growing body of research in cognitive network science (see

61reviews Baronchelli, Ferrer-i-Cancho, Pastor-Satorras, Chater, & Christiansen, 2013; Siew,

62Wulff, Beckage, & Kenett, 2019) by explicitly quantifying the large-scale structure of the mental

63lexicon as a multiplex language network (e.g., Castro & Stella, 2019; Siew & Vitevitch, 2019;

64Stella & Brede, 2016) and testing the influence of that network structure on word retrieval.

65Given that word retrieval failures are infrequent in typical adults, we focus on picture naming

66accuracy and errors made on incorrect trials by people with aphasia who often have significant

67word retrieval impairments to test the limits of this network model approach.

681.1. Models of typical and impaired word retrieval MULTIPLEX LEXICAL NETWORK 4

69 Most models of word retrieval describe three levels of word representation: a conceptual/

70semantic level, a lemma (or word) level, and a phonological level (see review Levelt, 1999). A

71network-like architecture is used to represent information in each level via nodes, and these

72nodes are connected to each other across levels. For example, several conceptual/semantic nodes

73have converging connections to a specific lemma node, which in turn has diverging connections

74to individual phoneme nodes (see Figure 1, left).

75

76Figure 1. The architecture of the interactive activation model (left) and a simplified depiction of 77the multiplex lexical network (right). In the interactive activation model (Dell et al., 1997), 78semantic nodes connect to word nodes and word nodes connect to phoneme nodes. The blue 79shaded nodes reflect shared semantic features between dog, cat, and rat. Note that there is no 80label as to what those shared features are in the interactive activation model. In the (simplified) 81multiplex lexical network, two networks are represented: a semantic network in blue and a 82phonological network in red, with semantic and phonological word form nodes connected 83between the separate networks as indicated by the dotted lines. Red phonological links represent 84a difference of one phoneme between connected pairs of words. Note that two shades of blue 85semantic links are used to highlight two different types of semantic relations existing between 86the words dog, cat, and rat. For example, the darker blue lines reflect shared features, like ‘is an 87animal’, ‘has a tail’, etc., while the lighter blue lines reflect free association cue-response pairs 88(i.e., dog-cat and cat-rat may be commonly provided free association pairs, but not dog-rat). 89

90 In order to retrieve a word, current models enact a process of spreading activation to

91bring semantic, lemma, and phonological information to conscious awareness. Spreading MULTIPLEX LEXICAL NETWORK 5

92activation processes have been articulated by several investigators (e.g., Anderson, 1983; Collins

93& Loftus, 1975), all hypothesizing three fundamental aspects: (i) activation spreads from one

94node to other related nodes (e.g., from a concept node to a lemma node), (ii) an entity must

95accumulate enough activation to either reach a particular threshold and/or hold the highest

96activation level in order to be “activated”, and (iii) activation diminishes in strength over time

97and over successive processing steps. Word retrieval failures and speech errors can occur at any

98one of these points during the spread of activation (e.g., a semantically related word receives

99more activation than a target word, resulting in a “semantic” error).

100 Despite these fundamental aspects of spreading activation processes, there are two types

101of models that differ in how activation spreads in the aforementioned cognitive architecture (see

102review Levelt, 1999): serial models and interactive activation models. As an example, consider a

103confrontation (i.e., picture) naming task where an image activates conceptual/semantic nodes.

104Serial models posit that the conceptual/semantic nodes transmit activation to relevant lemma

105nodes and it is only after lemma selection that activation is then transmitted to that lemma’s

106phonological nodes; in other words, activation spreads in a single direction with no feedback of

107activation from the phonological level to the lemma level (Levelt et al., 1999). On the other

108hand, interactive activation models begin with the transmission of activation from the

109conceptual/semantic level to the lemma level and further to the phonological level, but allows for

110the “upward” influence of phonological information on lemma selection (Dell & O’Seaghdha,

1111992). Unlike serial models, lemmas in interactive activation models can receive activation from

112both semantic and phonological levels (c.f. Rapp & Goldrick, 2000). Given that interactive

113activation models have been most prominently investigated in the literature with respect to the

114analysis of word retrieval failures by people with aphasia, whereas serial models have been MULTIPLEX LEXICAL NETWORK 6

115prominently used in the chronometric analysis of word retrieval (Levelt et al., 1991), the

116remainder of this paper focuses on the comparison between the class of interactive activation

117models and the multiplex network approach.

118 Aphasia is a disorder of language, often as the result of a left hemisphere stroke.

119Although there are several sub-types of aphasia related to deficits in language production and/or

120comprehension, word retrieval failures are a common feature. Thus, the study of word retrieval

121failures and limits of current models has benefited from research conducted with this unique

122participant sample. The leading hypothesis for word retrieval failures in aphasia is a primary

123breakdown in access-related deficits (Mirman & Britt, 2014); specifically, there is an impairment

124in the maintenance and/or transmission of spreading activation in the lexicon that underlies word

125retrieval (Martin & Dell, 2019). In particular, interactive activation models (Abel, Huber, & Dell,

1262009; Dell & O’Seaghdha, 1992; Dell et al., 1997; Foygel & Dell, 2000; Martin & Dell, 2019)

127have been influential in the study of word retrieval by focusing on the distribution of error types

128(e.g., semantically or phonologically related errors to a target word) that an individual with

129aphasia might produce in a confrontation naming or word repetition task.

130 For example, Foygel and Dell (2000) showed that by modifying certain parameters of

131their interactive activation model (e.g., “lesioning” links between the conceptual/semantic and

132the lemma levels to disrupt the spread of activation in the system), they were able to not only

133simulate the error distributions of individuals with aphasia, but also predict the locus of their

134word retrieval impairment (i.e., impairment in the semantic system or the phonological system).

135In comparing three types of aphasia, Conduction, Anomic, and Wernicke’s aphasia, Foygel and

136Dell were able to correlate the locus of impairment predicted by their model and hallmarks of

137these classic aphasia sub-types. Specifically, people with , who are often deemed MULTIPLEX LEXICAL NETWORK 7

138to be more mildly impaired, had relatively intact semantic and phonological systems compared to

139the other two types of aphasia. In contrast to people with Anomic aphasia, people with

140Conduction aphasia had greater phonological impairment aligning with classification requiring

141phonological deficits in repetition and production tasks, whereas people with Wernicke’s aphasia

142had greater semantic impairment aligning with classification requiring greater comprehension

143difficulties in repetition and production tasks (Foygel & Dell, 2000).

1441.2. Multiplex network representation of the mental lexicon

145 Although the interactive activation model has been influential in better understanding the

146processes underlying word retrieval (i.e., spreading activation), we propose that a complete

147understanding of word retrieval requires accounting for the structural influence of word

148representations in the mental lexicon. Indeed, Foygel and Dell (2000)’s work highlight the

149importance of considering the connections between words in the mental lexicon. Therefore, this

150paper draws on techniques from network science to quantify the similarity structure of word

151representations in the mental lexicon using a multiplex network framework.

152 In the multiplex lexical network studied by Stella and colleagues (Castro & Stella, 2019;

153Stella, 2018; Stella, Beckage, Brede, & De Domenico, 2018; Stella & De Domenico, 2018),

154nodes represent words (or lemmas) and are connected to each other on four layers, where each

155layer represents a different type of word-word similarity link. Three layers represent semantic

156relationships through free association (e.g., DOG-CAT, where the cue DOG elicited the response

157CAT in a free associate task), synonym sets (e.g., COUCH-SOFA, where COUCH and SOFA

158have similar meaning), and generalizations (e.g., FURNITURE-CHAIR, where CHAIR is a type

159of FURNITURE), and one layer represents phonological relationships through a one-phoneme

160edit distance rule (e.g., MAT-RAT). Note that the same set of nodes are represented in each MULTIPLEX LEXICAL NETWORK 8

161layer. Such an architecture allows for within-layer connectivity based on specific word-word

162similarity relations and between-layer connectivity reflecting the multiple linguistic features

163associated with a given word. Importantly, the multiplex lexical network conceptually resembles

164long-standing verbal descriptions of the mental lexicon (e.g., Collins & Loftus, 1975), as well as

165more modern computational models of word retrieval (Dell & O’Seaghdha, 1992; Levelt et al.,

1661999), both of which describe separate but connected semantic and phonological systems. In

167what follows, we discuss how the multiplex lexical network is structurally compatible with the

168interactive activation model and provides new ways of investigating instances of word retrieval

169failures (see Figure 1, right).

170 One difference between the architecture of the multiplex lexical network and the

171interactive activation model is the explicit definition of semantic relationships between words. In

172the interactive activation model, the conceptual/semantic nodes are not explicitly defined.

173Instead, each word is connected to 10 arbitrary semantic nodes (see Dell et al., 1997, Table 2, p.

174808). In contrast, the multiplex lexical network explicitly defines the number and type of

175semantic relations among words, in line with the previous literature on semantic memory

176representations in typical adults and people with aphasia (e.g., De Deyne & Storms, 2008;

177Erdeljac & Sekuliç, 2008; Sigman & Cecchi, 2002; Steyvers & Tenenbaum, 2005). Although

178other types of semantic relations may be important (e.g., features; McRae & Cree, 2005), the

179current formulation of the multiplex lexical network provides an explicit way of defining

180semantic relations.

181 An even more significant difference between the multiplex lexical network and the

182interactive activation model is simply the size of lexicon that is represented. Thousands of words,

183as well as the semantic and phonological relationships between all of these words, can be readily MULTIPLEX LEXICAL NETWORK 9

184represented in the multiplex lexical network, whereas existing models typically consider a very

185small set of lemmas. Specifically, the interactive activation model (Dell et al., 1997), and the

186class of models that followed (e.g., Abel et al., 2009; Foygel & Dell, 2000; Martin & Dell,

1872019), only has 6 lemmas. Recall that this model has been used to explain the distribution of

188speech errors made by persons with aphasia; therefore, the lemmas included in the model are

189related to the target word, as well as related and unrelated “error” words (e.g., 1 target lemma,

190plus 1 semantically related lemma, 1 phonologically related lemma, 1 “mixed” lemma that is

191both semantically and phonologically related, and 2 unrelated lemmas). Given that Dell and

192colleagues focus on investigating the parameters of spreading activation processes (i.e.,

193“lesioning” connections in the model to disrupt the spread of activation), their analyses have

194focused on the level of the individual; specifically, the overall picture naming behavior of an

195individual person.

196 Thus, one key limitation of models implementing a small lexicon is the inability to (i)

197investigate specific properties of the error word itself (i.e., an item-level analysis rather than a

198person-level analysis), and (ii) uncover the relationship between the error word and the intended

199target word. In this paper, we leverage on the fact that the multiplex lexical network

200representation naturally encodes word-word similarities among thousands of words, allowing

201researchers to conduct a more fine-grained analysis into the nature of the errors themselves. The

202next section will highlight different local and global network structure properties shown to

203influence word retrieval in typical adults and people with aphasia.

2041.3. Network structure influences word retrieval

205 The application of network science to the study of language representation and process is

206not novel (see reviews Baronchelli et al., 2013; Beckage & Colunga, 2015; Siew et al., 2019; MULTIPLEX LEXICAL NETWORK 10

207Vitevitch, Goldstein, Siew, & Castro, 2014). However, much of this work has focused almost

208exclusively on “single-layer” networks (i.e., a network with only one type of link; semantic only

209or phonological only); although network measures can be readily extrapolated from single-layer

210to multiplex networks. In this section, we describe three structural features of a network (i.e.,

211degree, network distance, closeness centrality; see Figure 2) and briefly discuss prior literature

212showing that these structural features influence various aspects of word retrieval when assuming

213a process of spreading activation in the networks.

214 First, we briefly discuss the body of psycholinguistic research focused on investigating a

215local network measure called degree (also called neighborhood density in the psycholinguistic

216literature; Luce & Pisoni, 1998; see review Vitevitch & Luce, 2016). Degree is simply the

217number of immediately connected “neighbors” of a particular node (i.e., a word), In Figure 2, the

218node for book has seven immediate connections, and thus has a degree of seven. The degree of

219words in semantic and phonological networks influences several language processes (e.g., De

220Deyne, Navarro, & Storms, 2013; Vitevitch & Luce, 2016), including how quickly and

221accurately typical adults (Vitevitch, 2002) and people with aphasia (Gordon, 2002; Vitevitch &

222Castro, 2015) produce words. In the context of spreading activation, words with higher degree

223(i.e., many immediate connections) accumulate activation more quickly from their many

224neighbors than words with lower degree (i.e., few immediate connections), leading to less

225errorful production. MULTIPLEX LEXICAL NETWORK 11

226

227Figure 2. A portion of the free association network depicting network measures of degree, 228closeness centrality, and shortest path length. Degree is the number of immediate connections, 229(e.g., the node for “book” has a degree of seven). Closeness centrality captures how close a node 230is to all other nodes in the network and is represented by the size of a node. Larger nodes have 231greater closeness centrality than smaller nodes (e.g., the node for “write” has high closeness 232centrality whereas the node for “meaning” has low closeness centrality). Path length refers to the 233number of edges between a pair of nodes, specifically the shortest path length between a pair of 234nodes. For example, the length of the shortest path length between the node for “letter” and the 235node for “sentence” is two (but note that there are other possible paths between the two nodes 236that are longer). 237

238 Another network measure shown to influence word retrieval is network distance, which

239quantifies the relationship between a specific pair of nodes in the network by computing the

240length of the shortest path connecting two nodes. For example, Figure 2 depicts the shortest path

241in red from the node for LETTER to the node for SENTENCE, which has a network distance of

242two. Note that it is possible to traverse from LETTER to SENTENCE using other paths;

243however, the shortest path would be the most efficient path for spreading activation (Collins &

244Loftus, 1975). Recently, Kenett, Levi, Anaki, and Faust (2017) showed that semantic relatedness MULTIPLEX LEXICAL NETWORK 12

245judgements were influenced by the network distance between two words in a semantic network

246of free associations, giving ecological validity to this network measure. Pairs of words separated

247by few links in the semantic network (i.e., pairs of words that were relatively close in the

248network) were rated as more semantically similar to each other than pairs of words separated by

249many links (i.e., pairs of words that were relatively far in the network). Theoretically, this

250measure has important implications for the spread and decay of activation: Words that are close

251to an activated word would receive more activation than words that are farther from an activated

252word, since activation decays over successive steps.

253 Lastly, the global network measure of closeness centrality captures the average distance

254of a node to all other nodes in the network, by considering the length of the shortest paths

255between a target node and all other nodes in the network. Note that this measure is an extension

256of network distance, which only considers the length of the (single) shortest path between a

257specific pair of nodes. Figure 2 depicts the closeness centrality of words through the size of

258nodes. For example, the node for WRITE has high closeness centrality, whereas the node for

259MEANING has low closeness centrality. Recent reports show that closeness centrality of words

260in the phonological network influences lexical decision in typical adults (Goldstein & Vitevitch,

2612017) and picture naming in typical adults and people with aphasia (Vitevitch & Castro, 2015).

262Nodes with high closeness centrality tend to reside on “short cut” paths in the network (i.e., short

263paths that connect distant nodes), whereas nodes with low closeness centrality tend to reside on

264the periphery of a network and are connected to other nodes via long paths (Holme, 2005). Thus,

265in terms of spreading activation, high closeness centrality words would receive more activation

266than low closeness centrality words because their location on short-cut paths allows for more

267activation to accumulate. On the other hand, words with low multiplex closeness centrality MULTIPLEX LEXICAL NETWORK 13

268would not receive enough activation because they are too distant from other words in the

269network and activation would have diminished over long network paths.

2701.4. Specific research aims

271 Although some psycholinguistic models are able to account for the influence of local

272structural effects on word retrieval (i.e., the influence of degree as in Dell & Gordon, 2003), it is

273not yet clear how those models in their present form would account for the influence of global

274structural effects on word retrieval (i.e., closeness centrality; e.g., Castro & Stella, 2019).

275Additionally, as noted previously, current models do not consider the relationship between target

276words and produced errors, whereas the measure of network distance may prove informative in

277this respect. Thus, we propose that network science methods enable language researchers to

278model the structure of the mental lexicon in a qualitatively intuitive and quantitatively precise

279manner, permitting continued investigations into how the structure of word representations in the

280mental lexicon influences word retrieval.

281 In this paper, we sought to extend upon the growing body of research examining the role

282of mental lexicon structure on language processes as described above by considering the

283influence of a lexical multiplex network structure on confrontation naming performance

284(accuracy and errors on incorrect trials) by people with aphasia. Our approach differs

285significantly from typical investigations into word retrieval failures, which focus on person-level

286analysis of error distributions (Dell et al., 1997; Foygel & Dell, 2000). Instead, we capitalize on

287the fact that the large-scale multiplex lexical network contains representations of thousands of

288words, as well as the semantic and phonological relationships among those words, to conduct a

289more fine-grained analysis of item-level confrontation naming data. MULTIPLEX LEXICAL NETWORK 14

290 To better understand the influence of the multiplex lexical network structure on word

291retrieval, we conducted a series of archival analyses on existing confrontation naming data by

292people with aphasia. In Section 3, we replicate and extend upon the findings of Castro and Stella

293(2019), who found that the global multiplex network measure of closeness centrality influenced

294successful word retrieval by people with Broca’s and Wernicke’s aphasia. In Sections 4 and 5,

295we consider a novel item-level analysis of errors on incorrect naming trials to consider the

296influence of the target word’s network structure (Section 4) and the influence of the connectivity

297in the network between a target word and an error word (Section 5). In all of these analyses, we

298compare the relative contributions of single-layer and multiplex variants of degree, network

299distance, and closeness centrality to assess the relative influence of the entire multiplex lexical

300network in relation to the additive contribution of individual network layers on word retrieval

301and speech errors. Finally, in Section 6 we propose a theoretical account for the influence of a

302multiplex network structure on the accuracy and error types associated with word retrieval and

303provide concluding remarks in Section 7.

3042. General Method

3052.1. Materials

3062.1.1. Aphasic Confrontation Naming Database

307 We acquired confrontation naming data by people with aphasia from the Moss Aphasia

308Psycholinguistic Project Database (MAPPD; Mirman et al., 2010). Classification of aphasia sub-

309type was provided in the database. We included data from 131 people with Anomic aphasia, 64

310people with Broca's aphasia, 53 people with Conduction aphasia, and 38 people with Wernicke's

311aphasia, which includes the aphasia sub-types analyzed in Foygel and Dell (2000) and Castro

312and Stella (2019). Furthermore, the inclusion of several different types of aphasia varying in type MULTIPLEX LEXICAL NETWORK 15

313and severity of deficits was to instantiate the notion that aphasia primarily impairs access to the

314mental lexicon, not its structure. (Mirman & Britt, 2014). Even though these sub-types of

315aphasia are hallmarked by differential impairment in comprehension and/or production abilities,

316based on the results of Castro and Stella (2019) we hypothesized that the underlying multiplex

317lexical network structure would have the same direction of influence on word retrieval,

318regardless of the person’s aphasia classification.

319 We focused our analyses on the Philadelphia Naming Test (PNT; Roach, Schwartz,

320Martin, Grewal, & Brecher, 1996), which is a 175-item picture naming test. Note that we only

321considered 139 of the target words due to the multiplex lexical network construction outlined

322below (see Supplementary Materials). The database contained codes for accuracy (‘correct’ or

323‘incorrect’) and error type according to the “model” response coding scheme (Mirman, Strauss,

324et al., 2010), which categorizes errors into five types based on relatedness to the target word:

325semantic, formal (or phonological), mixed (i.e., both semantically and phonologically related),

326nonword, and unrelated. Note that responses not falling into these categories (e.g., no response

327on a trial) are not considered in this coding scheme (Mirman, Strauss, et al., 2010). We followed

328this narrower coding scheme to remain consistent with what has been used in previous reports

329analyzing the interactive model of word production (Dell & O’Seaghdha, 1991; Dell et al., 1997;

330Foygel & Dell, 2000). Table 1 provides the percentage of trials for each scoring code.

331

332 MULTIPLEX LEXICAL NETWORK 16

333Table 1.

334Percentage of Correct and Errors for each Sub-type of Aphasia considering only trials for the 335139 PNT target words in the multiplex lexical network. 336 % of “Model” Errors % of % of Aphasia Unclassified Correct Semantic Formal Mixed Nonword Unrelated Sub-Type Errors Anomic 83.3% 3.1% 1.6% 1.7% 2.7% 0.7% 6.8% Broca’s 56.7% 5.1% 4.3% 2.5% 8.7% 4.9% 18.0% Conduction 65.8% 2.6% 7.5% 1.7% 10.5% 2.5% 9.5% Wernicke’s 45.0% 4.6% 8.6% 2.5% 14.0% 8.3% 16.9% 337* Unclassified Errors refer to the catch-all for all remaining incorrect responses not accounted for 338by the 5 error types used in the “model” coding scheme (e.g., no response; Mirman et al., 2010). 339

340 We briefly define error code criteria as obtained from

341https://mrri.org/wp-content/uploads/2016/01/PNT-Glossary-of-Terms_scoring-guide.pdf, which

342provides additional details on error coding. A semantic error code is given when the error

343response is related to the target word through synonym, category coordinate, superordinate,

344subordinate, associated, or diminutive relations. A formal error code is given when the error

345response is phonologically similar to the target word through sharing of: a) stressed vowel,

346initial, or final phoneme, b) two or more phonemes in any position (except unstressed vowels), or

347c) one or more phonemes at a corresponding syllable and word position. A mixed error code is

348given when the error response meets the criteria for a formal error (as outlined above) AND has

349an obvious semantic relation to the target (as outlined above). A nonword error code is given

350when the nonword is phonologically similar to the target (as outlined above). And, an unrelated

351error code is given when the response does not meet the criteria as semantically or

352phonologically related (as outlined above), including responses that have only a visual relation to

353the target and responses that are perseverations of previous naming trials.

3542.1.2. Multiplex Network Construction MULTIPLEX LEXICAL NETWORK 17

355 We represented the semantic and phonological similarities among words in the mental

356lexicon of the average adult native English speaker as a multiplex lexical network, as in previous

357studies (Castro & Stella, 2019; Stella, 2018; Stella et al., 2018; Stella & De Domenico, 2018).

358Figure 3 provides a depiction of the multiplex lexical network, where words are connected

359differently on each layer, but a particular node is connected to itself across each layer. Note that

360we do not make assumptions about the directionality of traversing the layers, nor the hierarchy of

361layers, in contrast to what one might expect in models that restrict the movement of spreading

362activation processes (e.g., Levelt et al., 1999; Rapp & Goldrick, 2000).

363 MULTIPLEX LEXICAL NETWORK 18

364Figure 3. The multiplex lexical network. There are four layers in the multiplex lexical network 365where nodes represent words in the mental lexicon. Words are connected within each layer 366according to free associations, synonyms, generalizations, or phonological similarities (i.e., intra- 367layer edges). Note that a word is connected to itself across all the layers (i.e., inter-layer edges). 368

369 The multiplex lexical network includes 8,531 words/nodes. Given that the multiplex

370lexical network captures several different types of word-word similarity relations, we rely upon

371the union of words across existing databases. Specifically, we obtained data from the Edinburgh

372Associative Thesaurus (Coltheart, 1981) and Wolfram (2018)’s Language WordData. Nodes in

373the multiplex lexical network were connected on at least one of four layers: free association,

374synonym, generalizations, and phonological similarity. (Wolfram, 2018). Free associations were

375obtained from the Edinburgh Associative Thesaurus (Coltheart, 1981), whereas synonyms,

376taxonomic dependencies, and phonological transcriptions were obtained from Wolfram’s

377Language WordData (Wolfram, 2018). The free association layer connected cue words based on

378participant responses during a free association task (e.g., cloud – rain). We did not consider the

379weighting (i.e., frequency of cue-response pairs) or directionality (i.e., asymmetric cue-response

380pairs; e.g., DOG leads to BONE, but BONE does not lead to DOG) of these edges following

381previous studies (e.g., Kenett, Anaki, & Faust, 2014; Kenett et al., 2017). The synonym layer

382connected synonymous words (e.g., character – nature). Note that synonym relations in

383WordData are obtained from WordNet 3.0 (Miller, 1995), which determines synsets (or groups

384of synonymous words) based on overlap in definitions and conceptual relations. The

385generalization layer connected words that exhibit a category-exemplar relationship (e.g., blue –

386color); again we did not consider weighting or directionality related to the relative position in the

387categorical hierarchy between the pair of words. Finally, the phonological layer connected words

388differing by only one phoneme through addition, substitution, or deletion (Luce & Pisoni, 1998;

389e.g., cat-cab) following the commonly studied phonological network of Vitevitch (2008). MULTIPLEX LEXICAL NETWORK 19

3902.1.3. Psycholinguistic Properties of Target Words

391 We obtained three psycholinguistic variables known to influence word retrieval for the

392139 PNT target words: frequency, length, and age of acquisition. Log word frequency was

393obtained from Opensubtitles dataset (Brysbaert & New, 2009), which contains a large corpus of

394subtitles from English movies and TV series. The target words had a mean log word frequency

395of 8.30 (SD = 1.44; ranging from 5.19 to 13.59). Word length was measured as the number of

396phonemes and obtained from Wolfram’s Language WordData (Wolfram, 2018). The target

397words had a mean length of 4.90 (SD = 1.47; ranging from 3 to 10). Age of acquisition was

398obtained from Kuperman, Stadthagen-Gonzalez, and Brysbaert (2012), who conducted a large-

399scale study gathering self-reported age of acquisition ratings for English words by adult native

400speakers via Amazon Mechanical Turk. The target words had a mean age of acquisition of 4.77

401(SD = 1.25; ranging from 2.50 to 8.85).

4023. Predicting Picture Naming Accuracy

403 We first analyzed accuracy of confrontation naming. This analysis extends upon that of

404Castro and Stella (2019), who found that multiplex closeness centrality was a stronger predictor

405of picture naming accuracy than the combination of single-layer closeness centrality predictors

406(i.e., association, synonym, generalization, and phonological closeness centrality) in individuals

407with Broca’s and Wernicke’s aphasia. In the present analysis, we assess the influence of both

408local network structure (i.e., degree) and global network structure (i.e., closeness centrality)

409measure, and consider additional sub-types of aphasia (i.e., Anomic, Broca’s, Conduction, and

410Wernicke’s aphasia).

4113.1. Method MULTIPLEX LEXICAL NETWORK 20

412 For the 139 PNT target words analyzed, we obtained two network measures, degree and

413closeness centrality, on each individual network layer (i.e., single-layer network measures) and

414computed a multiplex variant of each network measure (i.e., multiplex network measure). This

415allowed us to consider the individual influence of each network layer, as well as the combined,

416interactive influence of the entire multiplex lexical network. Table 2 provides the mean, standard

417deviation, and range for each single-layer and multiplex variant of degree and closeness

418centrality.

419

420Table 2.

421Mean, standard deviation, and range of single-layer and multiplex variants of degree and 422closeness centrality for the 139 PNT target items. 423 Network Measure Mean (SD) Range Association Degree 4.77 (31.94) 0.00 – 242.00 Synonym Degree 3.59 (4.54) 0.00 – 26.00 Generalization Degree 11.49 (10.70) 0.00 – 59.00 Phonological Degree 8.89 (8.67) 0.00 – 35.00 Multidegree 51.68 (40.86) 2.00 – 301.00 Association Closeness Centrality 0.29 (0.06) 0.00 – 0.37 Synonym Closeness Centrality 0.21 (0.25) 0.00 – 1.00 Generalization Closeness Centrality 0.25 (0.06) 0.00 – 0.32 Phonological Closeness Centrality 0.15 (0.12) 0.00 – 1.00 Multiplex Closeness Centrality 0.34 (0.02) 0.25 – 0.41 424

425 Degree is a local network measure, which is simply determined by counting the number

426of immediate edges for a given node. We computed degree of a word on each layer. For the

427multiplex variant, we computed multidegree, which is simply the sum of degree on all layers

428(Battiston, Nicosia, & Latora, 2017; Bianconi, 2018). MULTIPLEX LEXICAL NETWORK 21

429 Closeness centrality is a global network measure that captures how a node is connected to

430the rest of the network (Borgatti, 2005). Closeness centrality ci for node i in a given network with

431N connected nodes is defined as:

N c = i , (1) ∑ dij j 432

433where dij is the shortest path connecting nodes i and j. We calculated closeness centrality of a

434word on each individual layer by only considering the shortest path using edges for a particular

435layer. For the multiplex variant, we computed the average shortest path length by considering all

436edges in all layers at the same time. Closeness centrality ranges from 0 to 1, with higher values

437indicating greater centrality within the network.

4383.2. Statistical Analysis

439 We used a binomial logistic crossed-effects regression model to predict the probability of

440correct picture naming (i.e., correct vs incorrect) on a given trial for a given participant using the

441“lme4” package in R (Bates, Mächler, Bolker, & Walker, 2015). We tested several models in

442three stages by using a hierarchical model building procedure using Chi-Square tests (see

443Appendix 3) to determine the appropriate random effect structure and the constellation of fixed

444effects most critical for predicting picture naming accuracy. Fixed effects included single-layer

445and multiplex variants of degree and closeness centrality, while controlling for the expected

446influences of psycholinguistic variables. All predictors were z-scored prior to conducting the

447analyses.

4483.3. Results

449 The final model for predicting the probability of correct picture naming (i.e., correct vs

450incorrect) included fixed effects of word frequency, word length, age of acquisition, MULTIPLEX LEXICAL NETWORK 22

451generalization degree, multidegree, phonological closeness centrality, and multiplex closeness

452centrality, with random intercepts of diagnosis, subjects, and words. Table 3 provides the model

453output. Relevant to consideration of the multiplex lexical network, we found that both

454multidegree and multiplex closeness centrality were significant predictors of picture naming

455accuracy. Specifically, the odds of correct picture naming increased by 1.26 as multiplex

456closeness centrality increased by one standard deviation and 1.22 as multidegree increased by

457one standard deviation, holding all other variables constant. That is, pictures whose names were

458more central in the multiplex lexical network or had more immediate connections (regardless of

459edge type) were more likely to be named correctly than pictures whose names were less central

460in the multiplex network or had fewer immediate connections.

461

462Table 3.

463Final logistic crossed-effects regression model predicting correct picture naming from 464psycholinguistic variables, diagnosis, and single-layer and multiplex variants of degree and 465closeness centrality. 466 Fixed Effects Est. (log odds) St. Err. p-Value Intercept 0.633 0.473 0.180 Word Frequency 0.036 0.066 0.589 Word Length -0.114 0.055 0.040 Age of Acquisition -0.287 0.049 < .0001 Generalization Degree -0.265 0.073 0.0002 Multidegree 0.204 0.095 0.032 Phonological Closeness Centrality 0.075 0.039 0.056 Multiplex Closeness Centrality 0.231 0.099 0.020 467

Random Effects Variance St. Dev. Subject: Intercept 2.009 1.417 Diagnosis: Intercept 0.852 0.923 Word: Intercept 0.254 0.504 468

469 MULTIPLEX LEXICAL NETWORK 23

4703.4. Discussion

471 These results replicate and extend upon the findings of Castro and Stella (2019). First, we

472replicate the finding that multiplex closeness centrality has a facilitative influence on picture

473naming across several different types of aphasia. Recall that closeness centrality captures how

474central a word is in the network relative to all other words, which will influence the amount of

475activation that a word is likely to receive, with multiplex closeness centrality considering

476simultaneously the semantic and phonological connectivity of a word in the network. Our results

477reflect one of the fundamental aspects of spreading activation: namely, activation diminishes in

478strength over time and successive processing steps. That is, low closeness centrality words are at

479a disadvantage as such words do not receive enough activation needed for successful word

480retrieval, particularly for individuals with aphasia who may have deficits in the transmission of

481spreading activation (Martin & Dell, 2019).

482 Additionally, we provide novel evidence that multidegree facilitates picture naming

483accuracy. Recall that degree captures the number of immediate “neighbors,” or connections of a

484given word, with multidegree considering simultaneously semantic and phonological

485connectivity of a word in the network. There are two ways by which high multidegree may

486facilitate word retrieval: 1) having more neighbors leads to a greater number of pathways

487available for accumulating activation, and/or 2) converging activation from semantic and

488phonological neighbors of the target word reduces the likelihood of selecting an incorrect word.

489 Together, the results of this analysis predicting confrontation naming accuracy add to the

490growing body of evidence that the structure of a multiplex lexical network, at both local and

491global levels of analysis, contributes to word retrieval processes. However, understanding word

492retrieval processes also requires consideration of what happens when the process is disrupted. MULTIPLEX LEXICAL NETWORK 24

493Therefore, the remainder of this paper will focus on how the multiplex network structure

494influences words that are selected as “errors” on incorrect naming trials.

4954. Predicting Error Type on Incorrect Productions

496 Recall that the interactive activation model can help explain error distribution patterns for

497a specific person by determining the locus (i.e., within the semantic or phonological system) of

498their impairment (i.e., a person-level analysis; Foygel & Dell, 2000). This type of classification

499has significant implications for assigning certain individuals to appropriate treatment paradigms

500(e.g., a semantically impaired person would receive a semantic-based treatment, whereas a

501phonologically impaired person would receive a phonologically-based treatment). However, the

502interactive activation model cannot further our understanding in how the large-scale structure of

503word representations influences the type of error produced on incorrect trials (i.e., an item-level

504analysis). In the following analysis, we consider whether local and global network measures

505(degree and closeness centrality) of the target word influence the type of errors people with

506aphasia make for a target word.

5074.1. Method

508 Recall that the dataset from MAPPD includes five types of error codes: semantic, formal,

509mixed, nonword, and unrelated. As an initial attempt at understanding how network structure

510influences picture naming errors, we chose to focus on semantic, (e.g., camera for the target

511picture BINOCULARS), formal (e.g., zero for the target picture ZEBRA), and mixed (e.g., cattle

512for the target picture CAMEL) error types since they most closely correspond to the semantic

513and phonological layers considered in our multiplex lexical network. We did not analyze

514nonwords because they are not represented in the multiplex lexical network (i.e., nodes represent

515word lemmas). Although unrelated words may include single-word responses, we did not MULTIPLEX LEXICAL NETWORK 25

516analyze them because it is not always clear what an unrelated error reflects. Specifically, coding

517an unrelated error includes perseverations (i.e., producing a response given earlier in testing), as

518well as responses that have some visual relation to the target (e.g., producing the word ‘tie’ in

519response to the picture ‘waterfall’; https://mrri.org/wp-content/uploads/2016/01/PNT-Glossary-

520of-Terms_scoring-guide.pdf).

5214.2. Statistical Analysis

522 We used a multinomial regression model to predict the probability of error type

523(semantic, formal, or mixed) on a given incorrect picture naming trial of a given participant

524using “mlogit” in R (Croissant, 2012). (NB: We do not report a crossed-effects multinomial

525regression model given the complexity of the random effects structure and issues of data

526sparsity, even when attempted within a Bayesian framework; Bates, Kliegl, Vasishth, & Baayen,

5272015; Eager & Roy, 2017). We again tested several models in three stages by using a

528hierarchical model building procedure using Chi-Square tests (see Appendix 3) to determine

529which constellation of predictors were most critical for predicting error type on incorrect trials.

530Predictors considered included single-layer and multiplex variants of degree and closeness

531centrality, while controlling for the effects of psycholinguistic variables and diagnosis. All

532predictors were z-scored prior to conducting the analyses.

5334.3. Results

534 Our final model for predicting the error type on incorrect trials included word frequency,

535word length, age of acquisition, diagnosis (with Anomic as the reference group), association

536degree, synonym degree, phonological degree, multidegree, association closeness centrality,

537synonym closeness centrality, generalization closeness centrality, and multiplex closeness

538centrality. Table 4 provides the final model output. MULTIPLEX LEXICAL NETWORK 26

539 Of the network measures, association degree, multidegree, and multiplex closeness

540centrality had the largest regression coefficients across error type comparisons, holding all other

541variables constant. Specifically, the results indicated that as association degree increased by one

542standard deviation, the odds of semantic errors increased relative to mixed errors by 7.92 and the

543odds of formal errors increased relative to mixed errors by 4.13. On the other hand, as

544multidegree increased by one standard deviation, the odds of mixed errors increased relative to

545semantic errors by 17.15 and increased relative to formal errors by 8.23. Finally, as multiplex

546closeness centrality increased by one standard deviation, the odds of semantic errors increased

547relative to mixed errors by 2.25 and the odds of formal errors increased relative to mixed errors

548by 2.54.

5494.4. Discussion

550 Our results provide novel evidence that the structure of target words in the multiplex

551lexical network influences the types of errors people with aphasia produce for that given target

552word. These findings indicate that picture names with higher association degree and higher

553multiplex closeness centrality elicited more semantic and formal errors as compared to mixed

554errors, but picture names with higher multidegree elicited more mixed errors as compared to

555semantic and formal errors. These results suggest that semantic and formal errors are more

556influenced by network properties that capture separate layers of information, whereas mixed

557errors are more influenced by network properties that capture the overlap across layers of

558information. For example, although association degree reflects primarily semantic neighbors, this

559also includes phonologically related neighbors (N.B.: There was no explicit instruction during

560the free association task for participants to only provide semantically related responses to cue

561words). Thus, in the case of higher association degree, an error is likely to reflect either a MULTIPLEX LEXICAL NETWORK 27

562semantically or phonologically related neighbor. In contrast, multidegree captures neighbors that

563are both semantically and phonologically related. Thus, mixed errors are more likely to emerge

564as multidegree increases given the selection of neighbors that share both semantic and

565phonological relations.

566 Additionally, although multiplex closeness centrality is a network measure that

567simultaneously considers semantic and phonological layers in its computation, this measure is a

568global network measure that likely captures the initial search process of the network across

569semantic and phonological paths. Although a target word with high multiplex closeness

570centrality may benefit during spreading activation resulting in fewer errors made (as shown in

571Section 3), in the situation where an error is made, residing along many short-cut paths across

572layers increases the likelihood of producing either a semantically or phonologically related error

573(depending on which layer the search process terminates on) rather than a mixed error.

574

575 MULTIPLEX LEXICAL NETWORK 28

576 MULTIPLEX LEXICAL NETWORK 29

577Table 4.

578Final multinomial regression model predicting error type on incorrect picture naming trials from psycholinguistic variables, 579diagnosis, and single-layer and multiplex network measures of degree, and closeness centrality. Semantic v. Mixed Formal v. Mixed Est. (log odds) Std. Err. p-Value Est. (log odds) Std. Err. p-Value Intercept 0.720 0.076 < .0001 0.074 0.085 0.386 Word Frequency -0.136 0.064 0.035 -0.265 0.063 < .0001 Word Length -0.288 0.063 < .0001 0.005 0.062 0.925 Age of Acquisition -0.217 0.048 < .0001 -0.053 0.045 0.239 Broca (v. Anomic) 0.200 0.115 0.083 0.637 0.123 < .0001 Conduction (v. Anomic) -0.250 0.126 0.048 1.519 0.122 < .0001 Wernicke (v. Anomic 0.040 0.115 0.724 1.221 0.117 < .0001 Association Degree 2.072 0.234 < .0001 1.425 0.220 < .0001 Synonym Degree 0.227 0.073 0.001 0.227 0.068 0.0008 Phonological Degree 0.401 0.084 < .0001 0.529 0.078 < .0001 Multidegree -2.843 0.279 < .0001 -2.108 0.263 < .0001 Association Closeness Centrality -0.169 0.070 0.015 -0.189 0.069 0.006 Synonym Closeness Centrality -0.115 0.046 0.012 0.049 0.043 0.255 Generalization Closeness Centrality 0.004 0.057 0.933 0.199 0.061 0.001 Multiplex Closeness Centrality 0.810 0.124 < .0001 0.932 0.125 < .0001 MULTIPLEX LEXICAL NETWORK 30

5805. Predicting Error Type on Incorrect Productions using Network Distance

581 It is important to note that the prior two analyses (Sections 3 and 4) focused on

582investigating how the network properties of the target word affected the probability of correct

583naming and the likelihood of error types on incorrect trials. Particularly concerning is the small

584McFadden’s pseudo-R2 of our final multinomial model in Section 4 predicting error types

585(0.094), which indicates that a substantial proportion of variance in the data is unaccounted for.

586However, we may gain a more complete picture of word retrieval mechanisms by considering

587how the target word is connected to the produced error. A recent paper by Kenett et al. (2017)

588suggests that network distance, a metric quantifying the shortest path between two specific nodes

589in a network, could be a viable measure to consider given that network distance was able to

590reliably capture the magnitude of conceptual relatedness between two words. That is, participants

591judged the similarity of two words, and these behavioral results correlated with the network

592distance measure as calculated from a semantic network. We sought to investigate network

593distance in our data by first following best practices in statistics (Grimmett & Stirzaker, 2001) by

594examining the validity of network distance in our data. We compared the network distance

595between empirically generated target-error pairs to a random null model (see Appendix 2). We

596found that network distance between empirical target-error pairs was not random, indicating that

597this measure of network distance could be included as an additional predictor in the previous

598multinomial regression model predicting the type of error on incorrect trials produced by people

599with aphasia.

6005.1. Method

601 In this analysis, we had to further restrict the data to consider trials in which both the

602target and error words were located in the multiplex lexical network because we cannot calculate MULTIPLEX LEXICAL NETWORK 31

603network distance between a target word and a produced error if one or both of the words are not

604located in the multiplex lexical network. Given that this analysis is an extension of Section 4,

605Table 5 provides a depiction of the available data analyzed here in comparison to Section 4, with

606an approximate loss of 22% of the error analysis data.

607

608Table 5.

609Number and Percentage of Empirical Target-Error Pairs available for analysis depending on 610whether the target and error words were located in the multiplex lexical network. 611 Aphasia Anomic Broca’s Conduction Wernicke’s Sub-Type Error Type S F M S F M S F M S F M # Trials (Target + 573 299 293 476 381 200 261 789 171 437 732 209 Error in Network) # Trials (Target Only 789 401 420 596 512 289 346 991 219 529 984 285 in Network) % of data analyzed in 73 70% 75% 70% 80% 75% 69% 80% 80% 78% 80% 74% comparison % to Section 4 612*S = semantic error, F = formal error, M = mixed error 613

614 In network science, network distance is determined by considering the shortest path

615between nodes in the network. However, two nodes can be disconnected, resulting in an infinite

616value. In order to avoid this issue, we defined network distance as the inverse of the shortest path

617between two nodes. In this way, network distance can range from 0 (i.e., disconnected nodes) to

6181 (i.e., adjacently connected nodes). In our data, we obtained the network distance between a

619target word and the produced error word considering edges in each individual layer (i.e., single-

620layer network distance), as well as considering all edges simultaneously (i.e., multiplex network MULTIPLEX LEXICAL NETWORK 32

621distance. Table 6 provides the mean, standard deviation, and range of network distance between

622target-error pairs on each layer and on the multiplex network.

623

624Table 6.

625Mean, standard deviation, and range of single-layer and multiplex variants of network distance 626between target-error pairs. 627 Network Measure Mean (SD) Range Association Network Distance 0.31 (0.16) 0.00 – 1.00 Synonym Network Distance 0.09 (0.11) 0.00 – 1.00 Generalization Network Distance 0.24 (0.11) 0.00 – 1.00 Phonological Network Distance 0.17 (0.15) 0.00 – 1.00 Multiplex Network Distance 0.38 (0.11) 0.17 – 1.00 628

6295.2. Statistical Analysis

630 We extended upon the prior multinomial regression model in Section 4, by including an

631additional stage of model testing (see Appendix 2). Specifically, we tested whether the inclusion

632of network distance (single-layer and multiplex) between target-error pairs significantly

633contributed to the multinomial regression model predicting the probability of error type

634(semantic, formal, or mixed) on incorrect picture naming trials, while controlling for

635psycholinguistic variables, diagnosis, and network measures of degree and closeness centrality.

6365.3. Results

637 Our final model for predicting the type of error made on incorrect trials included word

638frequency, word length, age of acquisition, diagnosis (with Anomic as the reference group),

639association degree, synonym degree, phonological degree, multidegree, association closeness

640centrality, synonym closeness centrality, generalization closeness centrality, multiplex closeness

641centrality, association network distance, synonym network distance, generalization network

642distance, phonological network distance, and multiplex network distance. Table 7 provides the MULTIPLEX LEXICAL NETWORK 33

643model output. Critically, we found that inclusion of all network distance measures explained

644significantly more variance in predicting error types than the final regression model of Study 2,

645with an R2 gain of 0.235.

646 Network distance on all layers and the multiplex lexical network significantly influenced

647error type on incorrect picture naming trials. As distance between target-error pairs decreased on

648the association layer by 1 standard deviation, the odds of mixed errors increased relative to

649semantic errors by 1.34 and increased relative to formal errors by 1.56, keeping all other

650variables constant. As distance between target-error pairs decreased on the synonym layer by 1

651standard deviation, the odds of semantic and mixed errors were not different, but the odds of

652formal errors increased relative to mixed errors by 1.25, keeping all other variables constant. As

653distance between target-error pairs decreased on the generalization layer by 1 standard deviation,

654the odds of semantic errors increased relative to mixed errors by 1.15, while the odds of mixed

655errors increased relative to formal errors by 1.17, keeping all other variables constant. As

656distance between target-error pairs decreased on the phonological layer by 1 standard deviation,

657the odds of mixed errors increased relative to sematic errors by 1.20, while the odds of formal

658errors increased relative to mixed errors by 5.44, keeping all other variables constant. Finally, as

659distance between target-error pairs decreased on the multiplex network by 1 standard deviation,

660the odds of semantic and mixed errors were not different, but the odds of mixed errors increased

661relative to formal errors by 5.67, keeping all other variables constant.

6625.4. Discussion

663 These results demonstrate the importance of considering not just the lexical and structural

664properties of the target word, but also the way in which the target word and produced error are

665connected in the mental lexicon. One way of interpreting these results is to focus on how mixed MULTIPLEX LEXICAL NETWORK 34

666errors differ from semantic and formal errors, given that mixed errors reflect the combination of

667both. Mixed errors were more likely than formal errors when target-error pairs were close on the

668association layer or given the entire multiplex network, whereas formal errors were more likely

669than mixed errors when target-error pairs were close on the phonological and synonym layers. In

670regard to semantic errors, mixed errors were more likely when target-error pairs were close on

671the association and phonological layers, whereas semantic errors were more likely than mixed

672errors when target-error pairs were close on the generalization layer. Notably, the interactive

673activation model, and other models of word retrieval, cannot account for these results given their

674inability to model connections between specific pairs of words given their small lexicon sizes.

675Table 7.

676Final multinomial regression model predicting error type on incorrect picture naming trials 677from psycholinguistic variables, diagnosis, and single-layer and multiplex network measures of 678degree, closeness centrality, and network distance. Semantic v. Mixed Formal v. Mixed Est. (log odds) Std. Err. p-Value Est. (log odds) Std. Err. p-Value Intercept 0.723 0.085 < .0001 -0.068 0.109 0.535 Word Frequency -0.206 0.068 0.002 -0.256 0.074 0.0006 Word Length -0.302 0.069 < .0001 0.560 0.077 < .0001 Age of Acquisition -0.197 0.048 < .0001 -0.109 0.056 0.053 Broca’s (v. Anomic) 0.230 0.117 0.049 0.559 0.150 0.0001 Conduction (v. Anomic) -0.220 0.128 0.086 1.252 0.148 < .0001 Wernicke (v. Anomic) 0.053 0.117 0.647 0.992 0.142 < .0001 Association Degree 2.079 0.237 < .0001 1.259 0.272 < .0001 Synonym Degree 0.255 0.075 0.0007 0.115 0.085 0.175 Phonological Degree 0.428 0.084 < .0001 0.235 0.099 0.018* Multidegree -2.939 0.284 < .0001 -2.137 0.321 < .0001 Association Closeness Centrality -0.075 0.084 0.371 0.050 0.086 0.557 Synonym Closeness Centrality -0.121 0.047 0.010 -0.112 0.052 0.032 Generalization Closeness Centrality -0.023 0.063 0.708 0.437 0.079 < .0001 Multiplex Closeness Centrality 0.981 0.132 < .0001 1.418 0.146 < .0001 Association Network Distance -0.293 0.097 0.002 -0.448 0.102 < .0001 Synonym Network Distance -0.070 0.044 0.115 0.227 0.063 0.0003 Generalization Network Distance 0.145 0.054 0.007 -0.159 0.075 0.033 Phonological Network Distance -0.188 0.070 0.007 1.694 0.083 < .0001 Multiplex Network Distance 0.090 0.097 0.350 -1.736 0.111 < .0001 679 MULTIPLEX LEXICAL NETWORK 35

680 6816. General Discussion

682 Quantifying the structure of the mental lexicon and its influence on language processes

683remains an open challenge for cognitive science (Baronchelli et al., 2013; Beckage & Colunga,

6842015; Siew et al., 2019). In this paper, we applied computational and mathematical tools from

685network science to quantify the mental lexicon structure as a multiplex lexical network, which

686simultaneously represents semantic and phonological relationships between words, and

687examined the influence of this structure on confrontation naming performance (accuracy and

688errors) by people with aphasia. Our analyses revealed a complex pattern of results; although in

689the remainder of this paper we focus our discussion on the most significant patterns that were

690obtained from the measures computed from the multiplex lexical network and discuss how our

691results contribute to psycholinguistic theory regarding word retrieval processes.

6926.1. A framework for the influence of multiplex lexical network structure on picture naming

693performance.

694 We propose a framework that maintains the three fundamental aspects of spreading

695activation, following the interactive activation model. That is, spreading activation is

696implemented within the multiplex lexical network as the main mechanism driving word retrieval,

697specifically: (i) activation spreads across edges in the network from one node to other related

698nodes, (ii) a node must accumulate enough activation to either reach a particular threshold and/or

699hold the highest activation level in order to be “activated,” and (iii) activation diminishes in

700strength over time and successive processing steps in the network.

701 Although we align with these fundamental aspects of spreading activation, there is a

702major difference in the implementation of spreading activation within our framework as

703compared to the interactive activation model. Consider word retrieval in response to a picture MULTIPLEX LEXICAL NETWORK 36

704(i.e., confrontation naming). The class of models put forth by Dell and colleagues adopts a two-

705step process of retrieval. The stimulus (i.e., picture) introduces a jolt of activation to all the

706relevant semantic nodes in the semantic layer, which then spreads to lemma nodes and onward to

707phoneme nodes in an interactive fashion. A second jolt of activation is assigned to the lemma

708node that has the highest activation level when lexical selection occurs, and again activation is

709allowed to spread interactively across the layers. In contrast, we suggest a simpler process

710whereby during word retrieval, the stimulus (i.e., a picture) introduces a jolt of activation in both

711the semantic and phonological systems, initiating spreading activation separately but also

712simultaneously in the semantic and phonological layers of the multiplex lexical network.

713Importantly, this activation is not merely spread interactively (as in a simple diffusion process)

714between and within the layers; rather activation spreads in parallel in the various layers of the

715multiplex lexical network and activation patterns in any layer has the ability to inform activation

716patterns in another layer in a rapid, dynamic fashion until activation patterns settle on a given

717word. Such a process is consistent with the findings of the extensive body of research on parallel

718distributed processing models (Nadeau, 2012).

719 Additionally, we hypothesize that our implementation of spreading activation in a

720multiplex network could address one limitation of the interactive activation model, namely the

721small, consistent underprediction of mixed errors as reported by Foygel and Dell (2000, p. 211).

722The authors suggested that this might be due to the composition of the neighborhood of 6

723lexemes, which does not allow sufficient opportunities for the occurrence of mixed errors.

724However, another possible explanation may be the underlying structure of the model itself where

725lexemes essentially function as the “hidden layer” between semantic and phonological layers,

726and when combined with the 2-step spreading activation process, might not be sufficient for MULTIPLEX LEXICAL NETWORK 37

727semantic and phonological information to be truly interactive within the system. The alternative

728that we have outlined above has interactivity directly built into the structure of the multiplex

729lexical network, and when combined with a spreading activation process that can occur in

730parallel across layers, could be one way of naturally introducing more opportunities for mixed

731errors to occur.

732 Critically, our proposed implementation of spreading activation in the multiplex lexical

733network emphasizes the importance of network structure in influencing where activation spreads

734and accumulates, which in turn influences the likelihood of correct picture naming and the type

735of errors made on incorrect trials. Given this process of spreading activation, we propose that

736word retrieval first results in a very broad search of the mental lexicon as activation spreads

737within and between layers of the multiplex lexical network. As activation continues to spread in

738the network, the structure constrains where activation spreads and accumulates, resulting in a

739pool of possible lemma candidates that emerge (i.e., have increasingly greater activation levels)

740until a single word is selected for production. Our results considering accuracy and errors

741indicate that the global multiplex network measure of closeness centrality may be particularly

742important for the initial search of the network, whereas the local multiplex network measure of

743degree may be particularly important for narrowing the pool of lemma candidates at a later stage.

744 First, consider the role of multiplex closeness centrality in the word retrieval process.

745Recall that closeness centrality captures how central a word is in the network relative to all other

746words. Thus, words high in multiplex closeness centrality are close to and easily accessible from

747many other words, taking into consideration simultaneously semantic and phonological

748connectivity, due to their location on short-cut paths across multiple layers. Given our finding

749that high multiplex closeness centrality words were more likely to be named accurately than low MULTIPLEX LEXICAL NETWORK 38

750multiplex closeness centrality words (Section 3), we propose that centrality in the multiplex

751lexical network is an important component in guiding the initial search process to the general

752vicinity of the correct lexical representation within the network. Consider the following analogy:

753high multiplex closeness centrality words are like lighthouses. A lighthouse “pops out” on the

754horizon, drawing the attention of ships. In the same way, a word with high multiplex closeness

755centrality also “pops out” in the landscape of the network, drawing activation toward it to allow

756for greater accumulation of activation that is necessary for successful word retrieval.

757 However, the presence of multiple network paths passing through targets with high

758multiplex closeness centrality could act as a double-edged sword. Our results considering the

759error type of incorrect trials (Section 4) indicated that when errors do occur, they were less likely

760to be mixed errors (as compared to semantic and formal errors) when the target words had high

761multiplex closeness centrality than low multiplex closeness centrality. While a target word

762residing along many short-cut paths would accumulate more activation, the presence of many

763short-cut paths could also increase the likelihood that people with aphasia inadvertently wind up

764on an incorrect path while searching for the correct representation, thus leading to the production

765of either a semantic or formal error.

766 Now consider the role of multidegree on the word retrieval process. Recall that degree

767captures the number of immediate “neighbors", with multidegree considering simultaneously

768semantic and phonological connectivity in the network. Our results on picture naming accuracy

769(Section 3) indicated that words with high multidegree were more likely to be named accurately

770than words with low multidegree. This result suggests that not only does an increasing number of

771neighbors provide more paths for activation to accumulate, but also speaks to the convergence of

772information across multiple linguistic relations that boosts the likelihood of selecting the correct MULTIPLEX LEXICAL NETWORK 39

773target word over other related “error” words. In other words, the consideration of both semantic

774and phonological layers simultaneously helps to more efficiently narrow down the pool of

775candidates available for lexical selection, leading to greater naming accuracy for high

776multidegree words than low multidegree words. Furthermore, this result aligns with interactive

777activation models that indicate that activation in the phonological system can have an “upward”

778facilitative influence on word retrieval, in combination with activation in the semantic system.

779 Considering semantic and phonological neighbors simultaneously can also impact the

780type of errors likely to be made on incorrect trials. Specifically, our results (Section 4) indicated

781that when errors do occur, they were more likely to be mixed errors (as compared to semantic

782and formal errors) when the target word had high multidegree than low multidegree. This could

783be accounted for within the same framework outlined above. Although multidegree captures the

784fact that many semantic and phonological neighbors will be partially activated, the activation of

785phonological neighbors at the same time as semantic neighbors can help narrow down the pool

786of possible candidates (e.g., lead to the exclusion of semantic competitors that do not sound like

787the correct word). This process, however, also leads to the closest-competitor being a word that

788is both a semantic and phonological neighbor of the target (i.e., a mixed error). A word with high

789multidegree will have potentially more opportunities for these mixed errors (i.e., more neighbors

790that are both semantically and phonologically related) than a word with low multidegree. Thus,

791the structural overlap of the semantic and phonological layers of the mental lexicon in these

792instances could reinforce the activation of the incorrect word and lead to the production of a

793mixed error.

794 In sum, our results indicate that both the micro-level and macro-level structure of the

795mental lexicon, quantified by network measures of multidegree and multiplex closeness MULTIPLEX LEXICAL NETWORK 40

796centrality respectively, influence word retrieval. We have outlined a framework that provides a

797tentative account, but acknowledge that analysis of chronometric data and conducting

798computational simulations will be essential in testing our framework.

7996.2. Network distance as a new metric to quantify structure between two words

800 Although Sections 3 and 4 demonstrated that the inclusion of network measures derived

801from the multiplex lexical network were important for predicting picture naming accuracy and

802error types, it is important to note that these analyses focused on the structural properties of the

803target word only. However, other structural aspects of the lexicon might play a role in word

804retrieval processes, particularly in determining the type of errors produced on incorrect trials.

805Therefore, in light of recent work showing that network distance, a metric that quantifies the

806shortest path between two nodes in a network, was a significant predictor of performance on a

807semantic judgment task (Kenett et al., 2017), we examined if network distance might be

808important for predicting error types in picture naming.

809 As described earlier in the paper, network distance offers an intuitive way to measure the

810structure between specific pairs of words in a multiplex lexical network. Importantly, in a

811multiplex lexical network, it is possible for the shortest path between two nodes to traverse

812multiple layers of the network. The most striking result from the final set of multinomial

813regression models (Section 5) was that as multiplex network distance increased, the proportion of

814formal errors decreased relative to semantic and mixed errors. In other words, formal errors were

815most likely when the target and produced error were far apart on the multiplex network. This

816necessarily implies that there were no paths of short length to connect the two words, and a long,

817inefficient path through the multiplex lexical network must have ensued for the error to be

818produced. MULTIPLEX LEXICAL NETWORK 41

819 To understand this result, we must consider the structure of the multiplex lexical network

820and its individual layers, and how one would traverse the network given this structure. Previous

821work has shown that the global structure of semantic and phonological layers in the multiplex

822lexical network differs (Stella et al., 2018). Specifically, semantic layers tend to be more

823decentralized and are almost fully connected, whereas the phonological layer has a fragmented

824structure with several connected components (i.e., lexical islands) and disconnected "hermit"

825words (i.e., isolates). Therefore, spreading activation processes may be more efficient within

826semantic layers than on the phonological layer. For example, if one is navigating within a

827semantic layer, the structure of that layer might allow for better exploitation of the rest of the

828multiplex lexical network structure to find the correct representation, averting the production of a

829potential semantic error. In contrast, if one is navigating within the phonological layer, there is a

830higher likelihood of being trapped within a lexical island, leading to an inability to exploit the

831structure of the entire multiplex network and a higher probability of producing a formal error.

832 Furthermore, the process of traversing a multiplex lexical network may be especially

833problematic for a person with aphasia given that aphasia could lead to faster decay of spreading

834activation (i.e., impairment in maintenance) than expected of a typical adult (Dell et al., 1997;

835Martin & Dell, 2019). In other words, the number of edges that could be traversed in the network

836by a person with aphasia may be significantly less than a typical adult. Taken together with the

837unique structure of the multiplex lexical network, one could hypothesize that if a typical adult

838were “trapped” within the phonological layer, they could potentially re-trace their path back to

839the semantic layers and continue the search process. On the other hand, a person with aphasia

840might be unable to re-trace their path due to a lack of activation remaining in the system, leading

841to an early termination of the search process itself and production of the formal error. MULTIPLEX LEXICAL NETWORK 42

8426.3. Addressing long-standing research questions in cognitive science

843 Cognitive scientists have long been interested in the representation and processes of the

844human mind, including how we represent words in the mental lexicon and the processes we use

845to retrieve those words from the lexicon. Notable theories that describe how the mental lexicon is

846represented include a single mental lexicon representation (Fay & Cutler, 1977) that can account

847for both fixed and fuzzy meanings of words (Aitchison, 2012), a representation that includes

848both semantic and phonological levels/layers (Collins & Loftus, 1975; Dell et al., 1997; Levelt et

849al., 1999; Rapp & Goldrick, 2000), and the ability for processing to occur in semantics and

850phonology simultaneously (Collins & Loftus, 1975; Nadeau, 2012). Somewhat surprisingly,

851many models of semantic memory and the mental lexicon employ network-like structures, if not

852the explicit use of a network representation (e.g., Collins & Loftus, 1975; De Deyne, Navarro,

853Perfors, Brysbaert, & Storms, 2018; Landauer & Dumais, 1997; Miller, 1995; Quillian, 1967;

854Smith, Shoben, & Rips, 1974), suggesting that these types of representations are intuitive for

855capturing the structure of word-word similarity relationships. However, to date there is little

856consensus on the exact nature of the underlying representation of the mental lexicon (e.g., what

857linguistic relations to include) in part because the ability to quantify such a large, complex

858system (i.e., tens of thousands of words) has remained somewhat elusive (Jones, Willits, et al.,

8592015; Sousa & Gabriel, 2015; Steyvers & Tenenbaum, 2005).

860 In the present paper, we demonstrate that emerging techniques from network science

861provide a means to explicitly quantify the large-scale, complex structure of the mental lexicon,

862which includes consideration of semantic and phonological similarity among words, and allows

863us to begin explicitly testing the role of mental lexicon structure on language processes. One key

864advantage of the network science framework lies in the computational and mathematical power MULTIPLEX LEXICAL NETWORK 43

865of network science tools to formally represent the structure of the mental lexicon at multiple

866levels of analysis (e.g., the local and global levels) across language. Our findings pose significant

867challenges to existing models of lexical access, which, when taken to their limit, do not explicitly

868predict that a global structural measure, such as closeness centrality, would influence lexical

869processes. Furthermore, it is unclear how such a measure would even be inferred or derived from

870current models. Although one might naturally presume that a straightforward fix would be to

871simply “scale up” existing models by including more nodes at the lemma level, it remains to be

872explicitly implemented and tested if global network structure effects would naturally emerge

873from such a modification.

874 It is important to acknowledge previous work that has attempted to capture the influence

875of “distance” between two words in semantic memory, or the relative amount of similarity (low

876distance) or dissimilarity (high distance) between words. Distributional semantic space models,

877such as Latent Semantic Analysis (Landauer, Foltz, & Laham, 1998) and Hyperspace Analogue

878to Language (HAL; Lund & Burgess, 1996), are commonly used methods for defining semantic

879distance, where the distance between two words involves computing a spatial distance (e.g.,

880cosine distance) between two semantic vectors. For instance, work by Mirman and colleagues

881(e.g., Chen & Mirman, 2012; Mirman, Kittredge, & Dell, 2010; Mirman & Magnuson, 2008)

882showed that whether a related neighbor to a target word has a facilitatory or inhibitory effect on

883processing depends on the distance of that related neighbor (i.e., a “near” neighbor versus a “far”

884neighbor) in distributional semantic space. In contrast, the network science approach defines

885distance by computing the shortest path between two nodes in the network. Indeed, recent

886evidence indicates that the network approach better captures behavioral data than the

887distributional semantic approach (De Deyne, Perfors, & Navarro, 2016; De Deyne, Verheyen, & MULTIPLEX LEXICAL NETWORK 44

888Storms, 2016), including the work by Kenett et al. (2017) who used semantic network distance to

889predict semantic similarity judgments of human participants. These findings, as well as those of

890the present paper, suggest that in addition to semantic distributional models, the network

891approach represents a viable method to quantify the similarity structure of words in the mental

892lexicon.

893 In addition to theoretical questions related to the representation of the mental lexicon

894structure, there remains a long-standing debate in the cognitive sciences with regard to the

895challenges in dissociating the influence of cognitive processes and the structure of the cognitive

896systems on behavior. For example, understanding how semantic search operates in memory

897among a clinical population (frequently studied using a semantic fluency task where people list

898as many members of a given category as possible) requires a close consideration of how the type

899of representation that the modeler specifies (Abbott, Austerweil, & Griffiths, 2015; Jones, Hills,

900& Todd, 2015) interacts with disruptions to the process or damage to that structure (Gotts &

901Plaut, 2002). Furthermore, although many models of speech production consider how semantic

902and phonological information interact to influence the process of lexical selection (Dell et al.,

9031997; Rapp & Goldrick, 2000), they do not explicitly consider how the structure of semantic and

904phonological relationships among words influences that interactivity during lexical retrieval.

905 This theoretical debate of disentangling the influences of structure and process is

906particularly relevant in the present context if we are to advance our understanding of the nature

907of language impairment among clinical populations. The network science framework allows us

908to explicitly and systematically test whether language impairment in aphasia is largely due to

909deficits in the structure of the mental lexicon, deficits in the processes operating on the mental

910lexicon, or some combination of both. First, one could conduct simulations of a spreading MULTIPLEX LEXICAL NETWORK 45

911activation process on the multiplex lexical network in accordance with recent work by Siew

912(2019) who demonstrated how the implementation of spreading activation in a network

913representation could account for a variety of different lexical tasks, including spoken word

914recognition, false memory, and semantic priming. To simulate a deviant process, extra noise

915could be added to the spreading activation process by allowing activation to decay more rapidly

916(as done by Dell and colleagues; e.g., Dell et al., 1997; Foygel & Dell, 2000; Martin & Dell,

9172019), or by increasing the probability of jumps between layers of the multiplex network.

918Second, given that the underlying structure of the mental lexicon changes as a result of the

919aphasia through the weakening of connections between nodes (Dell et al., 1997; Foygel & Dell,

9202000), one could model “damaged” multiplex lexical networks with different structures based on

921aphasia type by deleting nodes from the representation and/or adjusting edge weights between

922and within layers to simulate a deficient structure. These simulations could systematically vary

923the network structure and the processes operating on that structure to examine which

924combinations of structure and process accounts for the overall pattern of picture naming

925performance by persons with aphasia.

926 In, sum, we argue that the application of network science will prove fruitful in addressing

927long-standing questions in the cognitive sciences by providing the means to explicitly quantify

928the vast and massively complex mental lexicon, and carefully incorporating the influence of

929structure on process. Although our proposed tentative framework remains to be formally tested

930and compared against existing models of word retrieval, this framework aligns with neural

931network models of word retrieval and naturally allows for incorporation of chronometric and

932error analyses to assess the dynamics of word retrieval—critical issues that have remained in the

933study of word retrieval (Levelt, 1999). MULTIPLEX LEXICAL NETWORK 46

9347. Conclusion

935 To recapitulate, we found that multiplex lexical network measures of degree, closeness

936centrality, and network distance influence picture naming accuracy and errors by people with

937aphasia. Our results add to the growing body of evidence showing the importance of adopting a

938multiplex approach to study the influence of mental lexicon representation on language

939processes. Specifically, we proposed a novel framework for the role of local and global multiplex

940lexical network measures on word retrieval based on the present pattern of results of picture

941naming accuracy and errors on incorrect naming trials: multiplex closeness centrality influences

942the initial search of the mental lexicon, whereas multidegree influences the narrowing of lemma

943candidates until lexical selection. In conclusion, this work shows how the application of network

944science can address long-standing questions in cognitive science, namely the quantification of

945mental lexicon representation and the ability to capture the influence of mental lexicon

946representation on language processes.

947

948

949

950 MULTIPLEX LEXICAL NETWORK 47

951 References

952Abbott, J. T., Austerweil, J. L., & Griffiths, T. L. (2015). Psychological Review Random Walks

953 on Semantic Networks Can Resemble Optimal Foraging. Psychological Review, 122(3),

954 558–569.

955Abel, S., Huber, W., & Dell, G. S. (2009). Connectionist diagnosis of lexical disorders in

956 aphasia. Aphasiology, 23(11), 1353–1378.

957Aitchison, J. (2012). Words in the mind : an introduction to the mental lexicon (4th ed.). Oxford:

958 John-Wiley & Sons.

959Anderson, J. (1983). A spreading activation theory of memory. Journal of Verbal Learning and

960 Verbal Behavior, 22(3), 261–295.

961Baars, B. (Ed.). (1992). Experimental slips and human error: Exploring the architecture of

962 volition. New York: New York: Plenum Press.

963Baronchelli, A., Ferrer-i-Cancho, R., Pastor-Satorras, R., Chater, N., & Christiansen, M. H.

964 (2013). Networks in Cognitive Science. Trends in Cognitive Sciences, 17(7), 348–360.

965Bates, D., Kliegl, R., Vasishth, S., & Baayen, H. (2015). Parsimonious Mixed Models. Retrieved

966 from http://arxiv.org/abs/1506.04967

967Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting Linear Mixed-Effects Models

968 Using lme4. Journal of Statistical Software, 67(1), 1–48.

969Battiston, F., Nicosia, V., & Latora, V. (2017). The new challenges of multiplex networks:

970 Measures and models. The European Physical Journal Special Topics, 226(3), 401–416.

971Beckage, N. M., & Colunga, E. (2015). Language networks as models of cognition:

972 Understanding cognition through language. In A. Mehler, A. Lucking, S. Banisch, P.

973 Blanchard, & B. Job (Eds.), Towards a Theoretical Framework for Analyzing Complex MULTIPLEX LEXICAL NETWORK 48

974 Linguistic Networks (pp. 3–28). Berlin: Springer.

975Bianconi, G. (2018). Multilayer Networks: Structure and Function. Oxford: Oxford University

976 Press.

977Borgatti, S. (2005). Centrality and network flow. Social Networks, 27, 55–71.

978Brysbaert, M., & New, B. (2009). Moving beyond Kučera and Francis: A critical evaluation of

979 current word frequency norms and the introduction of a new and improved word frequency

980 measure for American English. Behavior Research Methods, 41(4), 977–990.

981Castro, N., & Stella, M. (2019). The multiplex structure of the mental lexicon influences picture

982 naming in people with aphasia. Journal of Complex Networks, 7(6), 913-931.

983Chen, Q., & Mirman, D. (2012). Competition and cooperation among similar representations:

984 toward a unified account of facilitative and inhibitory effects of lexical neighbors.

985 Psychological Review, 119(2), 417–430.

986Collins, A., & Loftus, E. (1975). A spreading-activation theory of semantic processing.

987 Psychological Review, 82(6), 407–428.

988Coltheart, M. (1981). The MRC Psycholinguistic Database. The Quarterly Journal of

989 Experimental Psychology Section A, 33(4), 497–505.

990Croissant, Y. (2012). Estimation of multinomial logit models in R : The mlogit Packages.

991De Deyne, S., Navarro, D., Perfors, A., Brysbaert, M., & Storms, G. (2018). The “Small World

992 of Words” English word association norms for over 12,000 cue words. Behavior Research

993 Methods, 1–20.

994De Deyne, S., Navarro, D., & Storms, G. (2013). Better explanations of lexical and semantic

995 cognition using networks derived from continued rather than single-word associations.

996 Behavior Research Methods, 45(2), 480–498. MULTIPLEX LEXICAL NETWORK 49

997De Deyne, S., Perfors, A., & Navarro, D. (2016). Predicting human similarity judgments with

998 distributional models: The value of word associations. Proceedings of COLING 2016, the

999 26th International Conference on Computational : Technical Papers, 1861–

1000 1870.

1001De Deyne, S., & Storms, G. (2008). Word associations: network and semantic properties.

1002 Behavior Research Methods, 40(1), 213–231.

1003De Deyne, S., Verheyen, S., & Storms, G. (2016). Structure and Organization of the Mental

1004 Lexicon: A Network Approach Derived from Syntactic Dependency Relations and Word

1005 Associations. In A. Mehler, A. Lucking, S. Banisch, P. Blanchard, & B. Job (Eds.),

1006 Towards a Theoretical Framework for Analyzing Complex Linguistic Networks.

1007 Understanding Complex Systems (pp. 47–79). Berlin: Springer.

1008Dell, G., & Gordon, J. (2003). Neighbors in the lexicon: Friends of foes? In N. Schiller & A.

1009 Meyer (Eds.), Phonetics and phonology in langauge comprehension and production:

1010 Differences and similarities (Vol. 6) (pp. 9–37). Berlin: Mouton de Gruyter.

1011Dell, G., & O’Seaghdha, P. (1991). Mediated and convergent lexical priming in language

1012 production: A comment on Levelt et al. (1991). Psychological Review, 98(4), 604–614.

1013Dell, G., & O’Seaghdha, P. (1992). Stages of Lexical Access in Language Production.

1014 Cognition, 42, 287–314.

1015Dell, G., Schwartz, M., Martin, N., Saffran, E., & Gagnon, D. (1997). Lexical Access in Aphasic

1016 and Nonaphasic Speakers Lexical Access in Aphasic and Nonaphasic Speakers.

1017 Psychological Review, 104(4), 801–838.

1018Eager, C., & Roy, J. (2017). Mixed Effects Models are Sometimes Terrible. Retrieved from

1019 http://arxiv.org/abs/1701.04858 MULTIPLEX LEXICAL NETWORK 50

1020Erdeljac, V., & Sekuliç, M. (2008). Syntactic-semantic relationships in the mental lexicon of

1021 aphasic patients. Clinical Linguistics and Phonetics, 22(10–11), 795–803.

1022Fay, D., & Cutler, A. (1977). Malapropisms and the structure of the mental lexicon. Linguistic

1023 Inquiry, 8(3), 505–520.

1024Foygel, D., & Dell, G. (2000). Models of Impaired Lexical Access in Speech Production.

1025 Journal of Memory and Language, 43(2), 182–216.

1026Goldstein, R., & Vitevitch, M. (2017). The Influence of Closeness Centrality on Lexical

1027 Processing. Frontiers in Psychology, 8, 1683.

1028Gordon, J. (2002). Phonological neighborhood effects in aphasic speech errors: Spontaneous and

1029 structured contexts. Brain and Language, 82(2), 113–145.

1030Gotts, S. J., & Plaut, D. C. (2002). The impact of synaptic depression following brain damage: A

1031 connectionist account of "access/refractory" and "degraded-store" semantic impairments.

1032 Cognitive, Affective, & Behavioral Neuroscience, 2(3), 187–213.

1033Grimmett, G., & Stirzaker, D. (2001). Probability and random processes (3rd ed.). Oxford:

1034 Oxford Unviersity Press.

1035Holme, P. (2005). Core-periphery organization of complex networks. Physical Review E, 72(4),

1036 04611.

1037Jones, M., Hills, T., & Todd, P. (2015). Hidden processes in structural representations: A reply to

1038 Abbott, Austerweil, and Griffiths (2015). Psychological Review, 122(3), 570–574.

1039Jones, M., Willits, J., & Dennis, S. (2015). Models of semantic memory. In J. R. Busemeyer, Z.

1040 Wang, J. T. Townsend, & A. Eidels (Eds.), The Oxford Handbook of Computational and

1041 Mathematical Psychology (pp. 232–254). Oxford: Oxford University Press.

1042Kenett, Y., Anaki, D., & Faust, M. (2014). Investigating the structure of semantic networks in MULTIPLEX LEXICAL NETWORK 51

1043 low and high creative persons. Frontiers in Human Neuroscience, 8, 407.

1044Kenett, Y., Levi, E., Anaki, D., & Faust, M. (2017). The semantic distance task: Quantifying

1045 semantic distance with semantic network path length. 2Journal of Experimental

1046 Psychology: Learning, Memory, and Cognition, 43(9), 1470–1489.

1047Kuperman, V., Stadthagen-Gonzalez, H., & Brysbaert, M. (2012). Age-of-acquisition ratings for

1048 30,000 English words. Behavior Research Methods, 44(4), 978–990.

1049Landauer, T., & Dumais, S. (1997). A solution to Plato’s problem: The latent semantic analysis

1050 theory of acquisition, induction, and representation of knowledge. Psychological Review,

1051 104(2), 211–240.

1052Landauer, T., Foltz, P., & Laham, D. (1998). An introduction to latent semantic analysis.

1053 Discourse Processes, 25(2–3), 259–284.

1054Levelt, W. (1999). Models of word production. Trends in Cognitive Sciences, 3(6), 223–232.

1055Levelt, W., Roelofs, A., & Meyer, A. (1999). A theory of lexical access in speech production.

1056 Behavioral and Brain Sciences, Vol. 22, pp. 1–75.

1057Levelt, W., Schriefers, H., Vorberg, D., Meyer, A., Pechmann, T., & Havinga, J. (1991). The

1058 Time Course of Lexical Access in Speech Production : A Study of Picture Naming.

1059 Psychological Review, 98(1), 122–142.

1060Luce, P., & Pisoni, D. (1998). Recognizing spoken words: The neighborhood activation model.

1061 Ear and Hearing, 19(1), 1–36.

1062Lund, K., & Burgess, C. (1996). Producing high-dimensional semantic spaces from lexical co-

1063 occurrence. Behavior Reseach Methods, Instruments, & Computers, 28(2), 203–208.

1064Martin, N., & Dell, G. (2019). Maintenance versus transmission deficits: The effect of delay on

1065 naming performance in aphasia. Frontiers in Human Neuroscience, 406. MULTIPLEX LEXICAL NETWORK 52

1066Matuschek, H., Kliegl, R., Vasishth, S., Baayen, H., & Bates, D. (2017). Balancing Type I Error

1067 and Power in Linear Mixed Models. Journal of Memory and Language, 94, 305–315.

1068McRae, K., & Cree, G. (2005). Semantic feature production norms for a large set of living and

1069 nonliving things. Behavior Research Methods, 37(4), 547-559.

1070Miller, G. (1995). WordNet: a lexical database for English. Communications of the ACM, 38(11),

1071 39–41.

1072Mirman, D., & Britt, A. (2014). What we talk about when we talk about access deficits.

1073 Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1634),

1074 20120388.

1075Mirman, D., Kittredge, A., & Dell, G. (2010). Effects of near and distant phonological neighbors

1076 on picture naming. In Proceedings of the 32nd Annual Cognitive Science Society Meeting,

1077 32(32), 1447–1452.

1078Mirman, D., & Magnuson, J. (2008). Attractor dynamics and semantic neighborhood density:

1079 processing is slowed by near neighbors and speeded by distant neighbors. Journal of

1080 Experimental Psychology. Learning, Memory, and Cognition, 34(1), 65–79.

1081Mirman, D., Strauss, T., Brecher, A., Walker, G., Sobel, P., Dell, G., & Schwartz, M. (2010). A

1082 large, searchable, web-based database of aphasic performance on picture naming and other

1083 tests of cognitive function. Cognitive Neuropsychology, 27(6), 495–504.

1084Nadeau, S. (2012). The neural architecture of grammar. Cambridge, MA: The MIT Press.

1085Quillian, R. (1967). Word concepts: A theory and simulation of some basic semantic capabilities.

1086 Behavioral Science, 12(5), 410–430.

1087Rapp, B., & Goldrick, M. (2000). Discreteness and interactivity in spoken word production.

1088 Psychological Review, 107(3), 460–499. MULTIPLEX LEXICAL NETWORK 53

1089Roach, A., Schwartz, M., Martin, N., Grewal, R., & Brecher, A. (1996). The Philadelphia

1090 naming test: Scoring and rationale. Clinical Aphasiology, 24, 121–133.

1091Siew, C. (2019). spreadr: A R package to simulate spreading activation in a network. Behavior

1092 Research Methods, 51(2), 910–929.

1093Siew, C., & Vitevitch, M. (2019). The phonographic language network: Using network science

1094 to investigate the phonological and orthographic similarity structure of language. Journal of

1095 Experimental Psychology: General, 148(3), 475-500.

1096Siew, C., Wulff, D., Beckage, N., & Kenett, Y. (2019). Cognitive Network Science: A review of

1097 research on cognition through the lens of network representations, processes, and dynamics.

1098 Complexity, 2019, 2108423.

1099Sigman, M., & Cecchi, G. (2002). Global organization of the Wordnet lexicon. Proceedings of

1100 the National Academy of Sciences, 99(3), 1742–1747.

1101Smith, E., Shoben, E., & Rips, L. (1974). Structure and process in semantic memory: A featural

1102 model for semantic decisions. Psychological Review, 81(3), 214–241.

1103Sousa, L. B. de, & Gabriel, R. (2015). Does the mental lexicon exist? Revista de Estudos Da

1104 Linguagem, 23(2), 335-361.

1105Stella, M. (2018). Cohort and rhyme priming emerge from the multiplex network structure of the

1106 mental lexicon. Complexity, 2018, 6438702.

1107Stella, M., Beckage, N., Brede, M., & De Domenico, M. (2018). Multiplex model of mental

1108 lexicon reveals explosive learning in humans. Scientific Reports, 8, 2259.

1109Stella, M., & Brede, M. (2016). Mental lexicon growth modelling reveals the multiplexity of the

1110 English language. In H. Cherifi, B. Goncalves, R. Menezes, & R. Sinatra (Eds.), Complex

1111 Networks VII: Studies in Computational Intelligence, Vol. 644. (pp. 267-279). Cham, MULTIPLEX LEXICAL NETWORK 54

1112 Switzerland: Springer.

1113Stella, M., & De Domenico, M. (2018). Distance entropy cartography characterises centrality in

1114 complex networks. Entropy, 20(4), 268.

1115Steyvers, M., & Tenenbaum, J. B. (2005). The large-scale structure of semantic networks:

1116 statistical analyses and a model of semantic growth. Cognitive Science, 29(1), 41–78.

1117Vitevitch, M. (2002). The Influence of Phonological Similarity Neighborhoods on Speech

1118 Production. Journal of Experimental Psychology: Learning, Memory, and Cognition, 28(4),

1119 735–747.

1120Vitevitch, M. (2008). What Can Graph Theory Tell Us About Word Learning and Lexical

1121 Retrieval? Journal of Speech, Language, and Hearing Research, 51, 408–422.

1122Vitevitch, M., & Castro, N. (2015). Using network science in the language sciences and clinic.

1123 International Journal of Speech-Language Pathology, 17(1), 13–25.

1124Vitevitch, M., Goldstein, R., Siew, C., & Castro, N. (2014). Using complex networks to

1125 understand the mental lexicon. Yearbook of the Poznań Linguistic Meeting, 1, 119–138.

1126Vitevitch, M., & Luce, P. (2016). Phonological Neighborhood Effects in Spoken Word

1127 Perception and Production. Annual Review of Linguistics, 2, 75–94.

1128Wolfram. (2018). Wolfram Language WordData.

1129

1130

1131 MULTIPLEX LEXICAL NETWORK 55

1132 Supplementary Materials 1133 1134S1. The 139 Philadelphia Naming Test items analyzed. anchor apple baby ball balloon banana basket bat beard bed bell belt bench bone book boot bottle bowl bread bridge broom bus butterfly cake camera can candle cane cannon carrot cat celery chair chimney church clock closet clown comb corn cow cowboy cross crown dive dog door dragon drum duck ear eye fan fireman fireplace fish flower foot football fork frog garage ghost glass glove goat hair hammer hand harp hat heart helicopter horse hose house iron key king knife leaf letter lion man map monkey mountain nail nose nurse octopus owl pear pen piano pig pillow pineapple pipe pirate plant queen rake ring rope ruler saddle sailor saw scale scarf scissors seal shoe snail snake sock spider spoon star suit sun table tent toilet top train tree turkey van vest volcano wagon waterfall well whistle wig window zipper 1135 1136 MULTIPLEX LEXICAL NETWORK 56

1137S2. Correlation Matrix of Psycholinguistic Variables, Degree, and Closeness Centrality. 1138Correlations that were not statistically significant using a Kendall Tau test at a 0.01 confidence 1139level were grayed out. Red indicates strongly positive correlation while blue indicates strongly 1140negative correlation. 1141

1142

1143 MULTIPLEX LEXICAL NETWORK 57

1144Appendix 1. Regression Model Building Procedures for Sections 3-5.

1145 All regression model building used a data-driven approach to determine the constellation

1146of predictors best suited for each analysis. We tested several models over multiple stages using

1147forward and backward hierarchical model building procedures and Chi-Square tests. We outline

1148below the model building stages for each analysis.

1149 Predicting Confrontation Naming Accuracy

1150 As stated in Section 3, we used a binomial logistic crossed-effects regression model to

1151predict the probability of correct picture naming (i.e., correct vs incorrect) using the “lme4”

1152package in R (Bates, Mächler, Bolker, & Walker, 2015). Three stages of model testing ensued

1153(see Table A1). The first stage identified the appropriate random effects structure for our data to

1154account for variability among subjects and words. The second stage tested the inclusion of

1155psycholinguistic variables (word frequency, word length, and age of acquisition) as fixed effects.

1156Finally, the third stage tested the inclusion of single-layer and multiplex network measures of

1157degree and closeness centrality as fixed effects, while controlling for psycholinguistic variables

1158and diagnosis.

1159 The appropriate random intercept structure for our data included the nested effect of

1160diagnosis and subject, where each subject was assigned to only one type of diagnosis, and the

1161crossed random intercept of word (as each subject responded to each word). We found that

1162random slopes did not contribute to the model without issues of convergence, and follow best

1163practices outlined by (Matuschek, Kliegl, Vasishth, Baayen, & Bates, 2017) to model the

1164simplest random effects structure possible. Stage 2 testing resulted in retaining all fixed effects

1165of psycholinguistic variables. Stage 3 testing results in retaining some of the network measures, MULTIPLEX LEXICAL NETWORK 58

1166specifically generalization degree, multidegree, phonological closeness centrality, and multiplex

1167closeness centrality.

1168

1169Table A1.

1170Model building procedure to determine the appropriate fixed and random effects to predict the 1171likelihood of correct picture naming. 1172 R2 AIC Chi-Square Test Stage 1: Random Effect Structure A. Int + (1|Subject) + (1|Word) .519 53882 B. Int + (1|Diagnosis) + (1|Word) .521 53615 A-B: Χ2 (1) = 268.98, p < .0001 C. Int + (1|Diagnosis/Subject) + (1|Word) .514 53565 A-C: Χ2 (1) = 319.24, p < .0001 B-C: Χ2 (0) = 50.26, p < .0001 Stage 2: Fixed Effects of Psycholinguistic Variables D. Int + Freq + Len + AoA + (1|Diagnosis/ .513 53479 C-D: Χ2 (3) = 91.63, p < .0001 Subject) + (1|Word) Stage 3: Fixed Effects of Network Measures E. Int + Freq + Len + AoA + Gene Deg + .513 53462 D-E: Χ2 (4) = 25.12, p < .0001 Multidegree + Phon CC + Multiplex CC + (1|Diagnosis/Subject) + (1|Word) 1173Int = Intercept, Freq = Log Word Frequency, Len = Word Length, AoA = Age of Acquisition, 1174Asso = Association Layer, Syno = Synonym Later, Gene = Generalization Layer, Phon = 1175Phonological Layer, Deg = Degree, and CC = Closeness Centrality. We report the conditional 1176theoretical R2, which accounts for fixed and random effects, according to Nakagawa, et al. 1177(2017). 1178

1179 Predicting Error Type on Incorrect Confrontation Naming Trials

1180 As stated, in Sections 4 and 5, we used a multinomial regression model to predict the

1181probability of error type (semantic, formal, or mixed) on incorrect picture naming trials using

1182“mlogit” in R (Croissant, 2018). Four stages of model testing ensued (see Table A2), where

1183Section 4 focuses on stages 1-3 and Section 5 focused on stage 4. The first stage provided a null,

1184intercept-only model, since we could not test the random effects of subject and word in this

1185approach. The second stage tested the inclusion of psycholinguistic variables (word frequency,

1186word length, and age of acquisition) and diagnosis as fixed effects. The third stage tested the MULTIPLEX LEXICAL NETWORK 59

1187inclusion of single-layer and multiplex network measures of degree and closeness centrality as

1188fixed effects, while controlling for psycholinguistic variables and diagnosis. Finally, the fourth

1189stage tested the inclusion of single-layer and multiplex measures of network distance.

1190 Stage 2 testing resulted in retaining all psycholinguistic variables and diagnosis. Stage 3

1191testing resulted in retaining association degree, synonym degree, phonological degree,

1192multidegree, association closeness centrality, synonym closeness centrality, generalization

1193closeness centrality, and multiplex closeness centrality. Stage 4 testing resulted in retaining all

1194variables of Stage 3, plus all network distance variables.

1195

1196Table 2A.

1197Model building procedure to determine the appropriate model to predict the error type on 1198incorrect picture naming trials. 1199 R2 LogLik Chi-Square Test Stage 1: Intercept Only Null Model A. Int < .01 -4990.9 Stage 2: Psycholinguistic Variables and Diagnosis B. Int + Freq + Len + AoA + Diagnosis .060 -4687.2 A-B: Χ2 (12) = 567.72, p < .0001 Stage 3: Degree and Closeness Centrality C. Int + Freq + Len + AoA + Diagnosis + .094 -4520.8 B-C: Χ2 (16) = 314.74, p Asso Deg + Syno Deg + Phon Deg + < .0001 Multidegree + Asso CC + Syno CC + Gene CC + Multiplex CC Stage 4: Network Distance D. Model C + Asso Dist + Syno Dist + 0.329 -3347.9 C-D: Χ2 (10) = 1167.1, p Gene Dist + Phon Dist + Multiplex Dist < .0001 1200 1201Int = Intercept, Freq = Log Word Frequency, Len = Word Length, AoA = Age of Acquisition, 1202Asso = Association Layer, Syno = Synonym Later, Gene = Generalization Layer, Phon = 1203Phonological Layer, Deg = Degree, CC = Closeness Centrality, and Dist = Network Distance. 1204We report McFadden’s pseudo-R2. 1205 Appendix 2. Null model testing the viability of network distance used in Section 5. MULTIPLEX LEXICAL NETWORK 60

1206 In order to assess whether network distance between a target word and a produced error

1207is important for predicting error type, we first aimed to determine the dependency of error types

1208over network distance. In other words, we asked whether there was a quantifiable relationship

1209between empirical target-error word pairs that was not random. Here, we examined the network

1210distance between a target word and produced error, and compared these target word-error pairs

1211to a random expectation baseline (i.e., randomly selected errors). If meaningful differences were

1212obtained, this would motivate the inclusion of network distance as a new predictor to be included

1213in the multinomial regression model tested in Section 4 predicting error types on incorrect

1214picture naming trials. On the other hand, if network distance between empirical target-error pairs

1215is not different from network distance between random target-error pairs, then this would suggest

1216that network distance might not be a viable metric to consider for predicting error types.

1217 To assess a null model for network distance between empirical target words and

1218produced errors, we first needed to obtain a set of random errors. We conducted random error

1219sampling for every empirical target-error pair in our dataset. Specifically, for a given empirical

1220target-error pair (e.g., CAT-DOG), we sampled 50 random “errors,” or words from the multiplex

1221lexical network (e.g., SNOW). We connected each of the obtained random errors to the given

1222target word (CAT-SNOW), resulting in 50 random target-error pairs to compare to that given

1223empirical target-error pair (CAT-DOG). Choosing to do the randomization for every empirical

1224target-error pairs allowed us to maintain the distribution of error types across aphasia sub-types

1225(e.g., Foygel & Dell, 2000). To illustrate, there were a total of 1,165 empirical target-error pairs

1226produced by the Anomic aphasia group, with 573 coded as semantic errors, 299 coded as formal

1227errors, and 293 coded as mixed errors (see Table 5). Given our random sampling procedure, for

1228the Anomic aphasia group, we had 28,650 random target-error pairs to compare to the empirical MULTIPLEX LEXICAL NETWORK 61

1229target-(semantic) error pairs, 14,950 random target error pairs to compare to the empirical target-

1230(formal) error pairs, and 14,650 random target-error pairs to compare to the empirical target-

1231(mixed) error pairs.

1232 We sampled random errors with replacement to allow for the repetition of empirical

1233target-error pairs naturally found in the data (e.g., the same “error” could be selected more than

1234one time for a given target word). Additionally, although it is possible for a random error to be

1235the same as the empirical error, we did not attempt to control for any similarity between random

1236errors and empirical targets and errors. Thus, this randomization results in a useful null model for

1237detecting and quantifying similarities between the empirical target-error pairs generated by

1238people with aphasia.

1239 For both empirical target-error pairs and random target-error pairs, we calculated the

1240inverse of network distance between a target word and the error using the shortest path length.

1241Network distance ranged from 0 (disconnected nodes) to 1 (adjacent nodes) and can be computed

1242using edges on a specific layer of the multiplex lexical network or all layers of the multiplex

1243lexical network.

1244 Non-parametric testing using the Kruskal-Wallis test was used to compare the relative

(α) 1245difference, Δ , between the network distance of empirical target-error pairs (eij) to the network

1246distance of random target-error pairs (rij) on a specific network layer or the entire multiplex

1247lexical network structure (α). Specifically, we used the following formula:

(α) (α) (α) E −R (2) Δ = , E(α)

1248where E(α) was the median network distance between empirical target-error pairs and R(α) was the

1249median network distance between random target-error pairs. Large, positive values indicate that

1250the network distance between the empirical target-error pairs were closer (i.e., smaller distance) MULTIPLEX LEXICAL NETWORK 62

1251than expected when compared to the random baseline (i.e., target-random error pairs), whereas

1252large, negative values indicate that the network distance between the empirical target-error pairs

1253were farther (i.e., greater distance) than expected when compared to the random baseline. Values

1254close to 0 indicate that the distance between the empirical target-error pairs were not different

1255from the random baseline. For example, a relative difference of +30% would indicate that the

1256empirical target-error pairs were on average 30% closer than the random target-error pairs.

1257 Table 6 shows the relative difference in network distance between empirical target-error

1258pairs and random target-error pairs computed to match the error type distributions of each

1259aphasia sub-type. That is, we include error type in Table 6 to capture the differences in the

1260number of trials for each error type, not to instantiate any type of similarity between random

1261errors and target words.

1262 Across all error types, empirical target-error pairs were closer than the random target-

1263error pairs on the multiplex lexical network structure, which considers all semantic and

1264phonological layers simultaneously. Note that these patterns of relative differences in network

1265distance between empirical target-error pairs and random target-error pairs were consistent

1266across all types of aphasia. The results indicate that network distance between a target and

1267produced error could be considered in order to better understand what aspects of the network

1268structure influence lexical retrieval processes in people with aphasia.

1269

1270 MULTIPLEX LEXICAL NETWORK 63

1271 1272Table 1. 1273 1274The relative distance of empirical target-error pairs as compared to random target-error pairs 1275for each diagnosis, each error type, and each network layer/structure. 1276 Aphasia Empirical Association Synonym Generalization Phonological Multiplex Sub-type Error Type Layer Layer Layer Layer Network Anomic Semantic 46% **** 28% **** 26% **** 4% 40% **** Mixed 46% **** 22% *** 28% **** 34% **** 40% **** Formal 10% * 4% 3% 50% **** 30% **** Broca’s Semantic 45% **** 29% **** 23% *** 4% 39% **** Mixed 46% **** 26% **** 27% **** 23% *** 38% **** Formal 10% * 1% 10% * 37% **** 20% *** Conduction Semantic 45% **** 21% *** 28% **** 11% * 39% **** Mixed 41% **** 23% **** 23% **** 29% **** 37% **** Formal 13% ** 2% 9% * 48% **** 25% **** Wernicke Semantic 43% **** 22% *** 23% **** 4% 37% **** Mixed 43% **** 22% *** 22% *** 26% **** 39% **** Formal 14% ** 3% 7% * 40% **** 22% *** 1277* < .05, ** < .01, *** < .001, **** < .0001 MULTIPLEX LEXICAL NETWORK 64

1278