Robustness of the mental lexicon

Multiplex networks quantify robustness of the mental lexicon tocatastrophic concept failures, aphasic degradation and ageing

Massimo Stella1, a)

Complex Science Consulting, Via Amilcare Foscarini 2, 73100 Lecce, Italy

(Dated: 9 January 2020)

Concepts and their mental associations influence how language is processed and used. Networks represent powerfulmodels for exploring such cognitive system, known as mental lexicon. This study investigates lexicon robustness toprogressive word failure with multiplex network attacks. The average lexicon of an adult English speaker is built byconsidering 16000 words connected through semantic free associations and phonological sound similarities. Progres-sive structural degradation is modelled as random and targeted attacks. Words with higher psycholinguistic features(e.g. frequency, length, age of acquisition, polysemy) or network centrality (e.g. closeness, PageRank, betweennessand degree) are targeted first. Aphasia-inspired attacks are introduced here and target first words named correctly,more or less frequently, by patients with anomic aphasia, a pathology disrupting word finding. Robustness is measuredas connectedness, fundamental for activation spreading and lexical retrieval, and viability, a multi-layer connectivityidentifying language kernels. The lexicon is resilient to random, aphasia-inspired and psycholinguistic attacks. Catas-trophic phase transitions happen when phonological and semantic degrees are combined, making the lexicon fragileto multidegree attacks. The viable kernel is fragile to multi-PageRank and to aphasia-inspired attacks. Consequently,connectedness in the lexicon is mediated by hubs, whereas viability enables a lexical semantic/phonological interplayand corresponds to a facilitative naming effect in aphasia. These effects persist also through ageing, in different net-work representations of younger and older lexicons. This study indicates the need to prevent failure of high multidegreeand viable words in the mental lexicon when pursuing the design of effective language restoration strategies againstcognitive impairing.

I. INTRODUCTION

Complex networks constitute a powerful tool1,2 for investi-gating how conceptual associations, as represented in the hu-man mind, influence cognitive computing for language acqui-sition and use3–5. This mental lexicon of concepts and their as-sociations has been successfully investigated by several small-scale network models in psycholinguistics3,5–8. The recentadvent of network science9 further promoted the developmentof large-scale network models, enabling new understanding ofhow large ensembles of concepts are structured and influencelanguage processing in the semantic/meaning-driven10–14 andphonological/sound-driven4,15–18 aspects of the lexicon. Oneof the strongest appeal of such network models is to builda given layout of conceptual associations and then observehow it degrades when some damage is done19,20, e.g. somelinks or nodes are made unavailable to the system. Thesequantitative experiments are called network attacks and areused for investigating the robustness of networked systemsto random or targeted damage9,21–25. Although it is appeal-ing to investigate robustness of single aspects of the mentallexicon to degradation, this approach would neglect the ex-tensive evidence indicating how failed lexical retrieval cru-cially depends on the interplay between semantic and phono-logical conceptual similarities3,6,8,26,27. In order to embracemultiple aspects of the mental lexicon in a single networkrepresentation, this work suggests studying lexicon robustnessthrough a minimal representation of semantic and phonolog-ical associations for 16000 English words as a multiplex lex-

a)Electronic mail: massimo.stella@inbox.com

ical network26,28–31. Such representation is used as a quan-titative framework for exploring the large-scale robustness ofthe mental lexicon under progressive word failure through theformalism of network-based attacks20,22,23,25. Finding moreor less efficient ways of disrupting connectivity within wordsin the mental lexicon can offer valuable insights about thecognitive organisation of words in the mind and, more impor-tantly, also provide useful information aimed at understandingand then hampering or preventing cognitive decline in clinicalpathologies where language is progressively disrupted, likeaphasia8,32 or Alzheimer’s disease14,33.

Contrary to physical systems, that can be reproduced andtested directly in the laboratory, the mental lexicon remainsa cognitive construct, i.e. a system that cannot be accessedor tested straight way3. Once hypothesised from the struc-ture of language, conceptual associations in the mental lex-icon can only be indirectly tested through psycholinguistictasks, whose outcomes can provide indirect evidence aboutthe importance of the original associations for language pro-cessing and acquisition4. In the last forty years, overwhelm-ing empirical evidence at the interface of computer scienceand psycholinguistics has reported the relevance of seman-tic relationships such as free associations34–36 or sound sim-ilarities like phonological edit distances15,16,18 for predictingand understanding a variety of processes like lexical retrievalin recall experiments and identification responsiveness4,11,language acquisition2,12, writing styles37, knowledge struc-turing and search35,38 and even creativity levels34,39. Allthese approaches underline the importance of semantic andphonological associations between concepts for better un-derstanding how language is perceived and processed in thehuman mind. Single-layer complex networks have been avaluable asset in opening the way to large-scale, quantita-

arX

iv:2

001.

0262

1v1

[ph

ysic

s.so

c-ph

] 8

Jan

202

0

Robustness of the mental lexicon 2

tive investigations about the global layout and organisationof the mental lexicon1,2,12. Recently, thanks to advance-ments in the field of multiplex networks40,41, a novel gener-ation of large-scale models known as multiplex lexical net-works and aimed at achieving multiplex lexical representa-tions have been suggested and successfully tested for bet-ter understanding phenomena like language processing29, cre-ativity and fluid intelligence30 and early word acquisition28,31.

This work adopts a large-scale multiplex representation ofwords in the mental lexicon by means of the two most power-ful single-layer networks used in previous studies, free asso-ciations and phonological similarities, with the aim of study-ing how combining semantics and phonology in the mentallexicon affects robustness to progressive word failure, a phe-nomenon where words and their associations become increas-ingly inactive20.

By drawing inspiration from statistical physics and networkscience22,23, the robustness of the mental lexicon is testedby both random attacks (where nodes are made inactive uni-formly at random) and targeted attacks (where most promi-nent/central nodes are made inactive first as indicated by acertain attack strategy). Robustness is quantified as indicatedin previous approaches22,42 to network robustness under noderemoval, which is in terms of connectedness. Given the mul-tiplex nature of the current system, connectedness has to beclearly defined in this approach. In agreement with previ-ous literature41, connectedness here is tested by the existenceof any sequence or path of semantic and/or phonological as-sociations connecting words. Hence, connectedness on themultiplex network assumes the possibility to freely transi-tion between semantic and phonological aspects of the mentallexicon. This minimal assumption, while simplistic, has theadvantage of avoiding the quantification, if any, of a cogni-tive cost41 for transitioning between semantics and phonology,which is also an open question for the relevant psycholinguis-tic literature3 and for most empirical multiplex networks40,41.Importantly, the combination of semantic and phonologicallayers of conceptual knowledge opens new ways for connect-ing concepts that could be absent in the starting individualnetworks when kept as separate. An example is the workby Stella29, where the multiplex combination of semantic orphonological layers highlighted connectivity patterns relatedto phonological priming3,4 and absent in the individually con-sidered networks. The size of the largest connected compo-nent of the attacked multiplex network is not only a com-monly used metric of robustness in network science22,23 butit also has a clear cognitive relevance for lexical processing.According to the revised spreading activation model by Bockand Levelt43, a powerful model of psycholinguistics for ex-plaining positive semantic and phonological priming in lexicalretrieval, words are identified and recalled through an activa-tion signal that propagates over the mental lexicon along con-ceptual associations encapsulating both semantic and phono-logical aspects of words. While decaying over time, activationcan bounce back between conceptually associated words andalso concentrate on the desired concept, that can be subse-quently identified and recalled. Activation cannot spread out-side of conceptual connections. While inhibitory connections

can also be present1,5,43, connectedness represents a first, fun-damental constraint driving lexical processing. For this rea-son, this study will adopt multiplex connectedness as a metricfor robustness of the mental lexicon under attack.

Notice that the above notion of connectedness is not theonly one possible in a multiplex network. Rather than usingpaths of conceptual associations being made of either seman-tic or phonological network, a more restrictive "AND" logiccould be used, instead. Baxter and colleagues44 identified vi-able connectivity in terms of any two nodes being connectedwithin each and every layer, e.g. paths connecting the sametwo words through free associations only and, at the sametime, through phonological associations only. The largest vi-able component of the mental lexicon was identified26 as anetwork core for the mental lexicon, thus facilitating networknavigation, and emerging through an explosive phase tran-sition in children around age 7-8 yrs. Subsequent work30

identified the largest viable cluster as a relevant componentof the mental lexicon, made of more frequent, easier to iden-tify, earlier acquired words, of relevance also for knowledgeexploration across groups with different levels of creativity.Building upon such evidence for the cognitive relevance ofthe largest viable component in the mental lexicon, also itssize will be adopted as a metric for robustness under attack.

In addition to attacking the mental lexicon throughpsycholinguistics-driven strategies (e.g. attack first wordswith more meanings) and network-driven strategies (e.g. at-tack first words of higher closeness centrality), this work testsalso strategies based on clinical data from the PhiladelphiaNaming Test32 (PNT). By training a logistic regression modelover a dataset of picture naming by people with anomic apha-sia, the probability for correct naming of a given word in thewhole multiplex lexical network was predicted based on psy-cholinguistic and network metrics of relevance, reminiscentlyof previous network approaches in people with aphasia27,45.The training dataset included 31000 picture naming attemptsof 173 words. Although relatively small, the adopted datasetrepresented valuable knowledge for approximating the diffi-culty people with aphasia have in retrieving and producing agiven word. Testing also aphasia-like attack strategies basedon these picture naming probabilities represented an interest-ing large-scale and immediate test for identifying how theconnectivity in the mental lexicon degrades and potentiallydesign training approaches aimed at hampering such decline.

Identifying the robustness of the mental lexicon to all theabove attacks can open new ways for understanding, withinthe due assumptions and approximations, how language de-clines in specific pathologies and therefore come up with ef-ficient countermeasures46. Notice that the main assumptionof this analysis is that words fail progressively and the cor-roboration of such working hypothesis opens new challengesfor clinical fieldwork. In addition to the clinical implica-tions, robustness results represent also a valuable source ofinformation20 for exploring and understanding how the humanmental lexicon is structured and behaves under degradation.

This study is structured as follows. The Methods sectionreports on the data and network metrics used in the attack ex-periments. The Results section is divided in three parts: The

Robustness of the mental lexicon 3

first one outlines the robustness profile of the mental lexiconin terms of global connectedness; the second part focuses onthe viable component; the third part tests lexicon robustnessthrough ageing when semantic connections get weakened. Afinal Discussion compares and interprets results and identifiesconcrete future research directions opened by the current ap-proach.

II. METHODS

A. Cognitive data

Four psycholinguistic features of words were used in thisstudy: (i) frequency of words in language as computed inSUBTLEX47 (a dataset including 2.2 billions subtitles), (ii)word length, (iii) mean age of acquisition of a concept in En-glish native speakers48 and (iv) number of meanings of a word(polysemy score) computed from the dictionary implementedin WordData[] by WolframResearch and based on WordNet3.049. All these measures were reported as being predic-tive of a variety of language processing tasks such as lexicalidentification3,47,48, recall17,47 and confusability4,47,48. Basedon such evidence, the above psycholinguistic features wereused here as indicators for designing targeted attack strategiesto the multiplex lexical network of relevance from a cognitiveperspective.

B. A logistic regression model for estimating correct namingin people with anomic aphasia

With the aim of having a data-driven way of simulatingprogressive decline of the mental lexicon as affected by aclinical pathology, this study used cognitive data from thePhiladelphia Naming Test (PNT)32. The PNT dataset, avail-able through a free registration on the Moss Aphasia Psy-cholinguistics Project Database website, reports on the perfor-mance of people with aphasia and healthy controls in a nam-ing task including 175 items/words50. People with a diagnosisfor aphasia were shown a picture and asked to name it. Eachpicture represented a concept, which was correctly named byalmost all healthy subjects (with a probability above 0.99).People with aphasia had degraded linguistic skills, so that tothem picture naming was more difficult51.

Considering almost 31600 utterances produced by peoplewith anomic aphasia, the probability for correct picture nam-ing in people with aphasia was 0.806. In general, peoplewith anomic aphasia are characterised by an overall fluentspeech complicated by occasional difficulties in word finding,happening with higher frequency in comparison to healthysubjects51–53.

By aggregating together all naming experiments relativeto each single item/word, empirical probabilities pw for cor-rect naming were obtained for all the 173 words from PNTpresent also in the multiplex lexical network. These prob-abilities were then fed to a logistic regression model basedon predictors available in the multiplex network, namely the

TABLE I. Results of the logistic regression of empirical produc-tion probabilities pw by using psycholinguistic and network predic-tors. The more stars, the lower the p-value. The multi-step proce-dure discarded network metrics and identified the model minimisingAkaike’s Information Criterion (28709) in 5 Fisher steps.

Coefficient Estim. St.Err. z-score p-value(Intercept) -0.466 0.120 -3.87 0.0001 ***

Freq. -0.073 0.022 -3.28 0.0010 **Length -0.326 0.026 -12.55 < 10−8 ***

Age of Acq. -0.721 0.037 -19.22 < 10−7 ***Polysemy 0.069 0.017 3.91 < 10−5 ***

Asso. Clos. 1.100 0.198 5.53 < 10−7 ***Mult. Clos. -0.221 0.060 -3.68 0.0002 ***

Asso. PageR. 0.709 0.142 4.97 < 10−7 ***Phon. PageR. 0.059 0.023 2.57 0.0101 *Asso. Degree -0.492 0.141 -3.47 0.0005 ***Asso. Betw. -0.239 0.038 -6.17 < 10−7 ***

above mentioned psycholinguistic features and also networkmetrics9 like single-layer and multiplex degree, PageRank,closeness and betweenness centrality. The selection of thesenetwork metrics was based on previous results from27, whichused a similar approach for predicting correct naming in peo-ple with aphasia. The regression analysis was fitted through amulti-step procedure, as implemented in R, in order to discardredundant predictor variables and minimises Akaike’s Infor-mation Criterion.

The fitted logistic model reported a pseudo-R2 of 0.096(Cragg and Uhler’s). All psycholinguistic variables were iden-tified as having an influence over correct naming predictions.Multiplex closeness was the only multiplex measure to displaypredictive power over correct naming performance. For de-gree, PageRank and betweenness, single-layer measures wereselected to be within the model. Multiplex degree indicatethe sum of degrees across layers40. Multiplex PageRank wasbased on versatility41,54. Multiplex betweenness and central-ity exploited multilayer paths, like in previous works26,29,55.

The resulting logistic regression model, fitted over empir-ical data of word production in people with anomic aphasia,was then fed with vectorial inputs for all words in the mul-tiplex lexical network. For the 173 words in the PNT andin the multiplex network, logistic predictions and empiricalvalues of the probabilities for correct word naming correlatedstrongly (Spearman Rho 0.726, p-value < 10−7), indicating agood consistency between regression predictions and empiri-cal data. Through the trained regressor, for each word a prob-ability of correct naming pc(w) was estimated. These prob-abilities were then used for designing aphasia-inspired attackstrategies (cfr. Attack strategies in the Methods).

C. Multiplex lexical network building

This work adopts a minimal representation of semanticsand phonology in the human mental lexicon as a multiplexlexical network26–29, where concepts are represented by nodesand are linked on two network layers by:

Robustness of the mental lexicon 4

• Free associations10,13,36, indicating recall of conceptselicited by participants in a cognitive task (e.g., read-ing "bed" elicited the concept "sleep" x times). Thislayer was based on the Small World of Words datasetby De Deyne and colleagues36. Associations recalledmore than x ≥ 11 times were considered, in order forthe whole layer to feature as many links as the phono-logical one. Also for consistency with the phono-logical layer, this layer was treated as undirected andunweighted, as done in previous cognitive computingapproaches27,30,35.

• Phonological similarities15,16, linking words sound-ing similar (e.g. "cab" and "cat"). Sound analo-gies were quantified by checking that phonologi-cal IPA transcriptions of words differed for the ad-dition/substitution/deletion of one phoneme only, asin Vitevitch’s original approach15. Transcriptionswere obtained from the WordData[] repository curatedby WolframResearch, available through Mathematica11.3.

The resulting multiplex lexical network was minimal in thesense that, differently from other approaches29,30, it consid-ered only one layer of semantic relationships coupled with aphonological one. Indeed, free associations identify recallsfrom semantic memory11 that might be related not only to se-mantics but also to sound similarity or visual features. De-spite a small overlap with semantic features and phonologicalrelationships, as highlighted in28,31, a layer similarity analysisperformed in a previous work26 identified free associations asstructurally dissimilar from the phonological layer and moresimilar, for its connectivity layout, to other semantic layerslike conceptual generalisations or synonyms. This quantita-tive result identified the layer of free associations as beingmainly semantic, which justifies the main approach of thisstudy.

The whole multiplex lexical network included 15886 En-glish concepts and almost 75410 conceptual associations. Ta-ble II reports network statistics about link density and connec-tivity for the individual network layers, the resulting multiplexlexical network and random null models (i.e. configurationmodels, where links are randomised while keeping fixed theexact degree distribution of original networks56). The mul-tiplex network features a largest viable component (LVC) of4183 words and it is fully connected, i.e. its largest connectedcomponent (LCC) includes all 15886 words in the multiplexnetwork. Like in past approaches26–28,30, two words are con-nected if a multiplex path exists between them, i.e. a networkpath combining free associations and phonological similari-ties connected the words.

As reported in Figure 1, viability requires connectedness onall the individual layers, i.e. viable words have a path made ofsemantic associations only and also a path made of phonolog-ical similarities only both connecting them. Viable connec-tivity is more restrictive than connectedness when layers arecombined in the multiplex structure. Instead, viability reducesto connectedness when layers are considered individually.

The free association data from the Small World of Words

FIG. 1. Example visualisation of a multiplex lexical network wherethe LVC and LCC are highlighted. Every word is represented as tworeplica nodes, one representing the semantic lexical representation ofa word, linked through free associations, and another one represent-ing the phonological word form, being connected through phonolog-ical sound similarities. Jumps across layers are possible at no cost interms of distance or transition probability.

TABLE II. Network statistics for the layers of free associations (FreeAsso.), of phonological similarities (Phon. Sim.) and the wholemultiplex lexical network (Multiplex N.). LCC Size is the size of(or nodes in) the largest connected component (LCC). Mean Lengthstands for the average path length between any two nodes. LVC Sizeis the size of the large viable component (LVC). In single-layer net-works, the LVC coincides with the LCC. All layers included 15886words. For comparison, the mean values of 10 configuration modelsof each layer are provided too, where the empirical degree distribu-tion is fixed but links are randomised.

Network Links LCC Size Mean Length LVC SizeFree Asso. 36383 11512 4.8 11512

Free A. CM 36383 11515 4.3 11512Phon. Sim. 35803 8029 7.5 8029Phon.S. CM 36383 10897 4.6 10897Multiplex N. 75410 15886 4.6 4183

project36 included also the age of participants who provided agiven free association. Hence, the dataset enabled the pos-sibility to build free association networks for different agegroups. Age groups were chosen in order to produce layersof free associations with the same link density as the aggre-gated one (for which associations produced by at least 11 in-dividuals were selected). The resulting age groups consideredweaker free associations, produced by at least 3 individuals ofthe following ages: (i) younger than 22 years old, (ii) between22 and 28 years old, (iii) between 28 and 36 years old, (iv)between 36 and 50 years old and (v) older than 53 years old.Free associations map patterns of memory recall10,35, which issupposed to change with ageing57,58. In this way, considering

Robustness of the mental lexicon 5

more fine-grained layers of semantic associations filtered byage can provide temporal snapshots about how the robustnessof the mental lexicon evolves with ageing.

D. Viability identifies key words in the mental lexicon

The largest viable cluster, or LVC, identified a set of wordspossessing different psycholinguistic features compared toothers outside of it. Location equivalence tests indicated thatwords within the LVC: (i) had a higher median frequency,(ii) were shorter in length, (iii) were learned earlier on duringlanguage acquisition and (iv) possessed more meanings thanwords outside (cfr. Table III). These results indicate that theLVC contains words key to language processing and in thisway it represents a language kernel59, i.e. a set of general,frequent and easier to acquire words of crucial relevance forefficient communication.

Also the empirical probability of correct picture naming pwwas tested between words within and outside the LVC (seeMethods). For people with anomic aphasia, pictures repre-senting words in the LVC were easier to name compared topictures representing concepts outside of the LVC. This dif-ference might indicate a beneficial effect of the LVC in pro-moting lexical identification and word production in a clinicalpopulation but this conjecture will be further tested with at-tacks in the Results section.

The above results indicate that the LVC is of relevance forlanguage representation and processing and its connectivitywill therefore be used as a metric for robustness in the attackexperiments.

TABLE III. Location equivalence tests for psycholinguistic featuresof words outside and within the LVC (Mann-Whitney test). Lengthwas measured in terms of letters. Polysemy identified the number ofmeanings of a word. A log correction was applied to frequency asperformed in most psycholinguistic regression analyses.

Psychol. Feature In LVC Out LVC TestWord Length 5 7 7.7 ·106, p < 10−7

Log Frequency 7.05 1.31 3.6 ·106, p < 10−7

Age of Acquisition 7.06 yrs 9.57 yrs 1.4 ·106, p < 10−7

Polysemy Score 4 2 4.2 ·106, p < 10−7

Correct Naming Prob. 0.87 0.74 5131, p < 10−7

E. Attack strategies

In order to mimic progressive cognitive decline inthe mental lexicon, network attacks were performedprogressively22,23, i.e. by removing increasing fractionsof nodes over time from the network topology. An at-tacked/removed node did not contribute to connectivity andcould not be traversed in order to connect other nodes. At-tacked nodes did not recover over time, i.e. when more andmore nodes where attacked. The number of nodes (size) in thelargest connected component and in the largest viable compo-nent of the multiplex lexical network were measured over time

for each and every attack strategy. Nodes were ranked accord-ing to the initial layout of the multiplex lexical network. Noonline, intra-layer update24 was performed in this case sincethe robustness experiments were focused on detecting wordcentralities in the original, fully working networked mentallexicon.

Attack strategies were either random or targeted. In ran-dom attacks, words were selected uniformly at random andattacked/removed. Results were averaged over 100 instancesof random attacks. Targeted attacks removed first nodes cen-tral/relevant for the mental lexicon. Words were ranked andremoved at uniform steps (i.e., removing top 100 nodes, top200 nodes, top 300 nodes, etc.). Centrality was estimated interms of either psycholinguistic norms, network location oraphasia-inspired probabilities pc(w). In psycholinguistic at-tacks, words of either higher frequency or smaller length orlower mean age of acquisition or higher polysemy (i.e. num-ber of different meanings) were attacked first. For networkattacks, degree, PageRank, betweenness and closeness wereused as metrics of word centrality. Degree counted the num-ber of associations of a given word at a microscopic level inthe conceptual structure of the mental lexicon. Phonologicaldegree is generally also known as phonological neighbour-hood density in psycholinguistics. Multidegree40 counted allphonological and semantic associations together. In both se-mantic and phonological networks, words with high degreewere found to be easier to learn12,31 and retain2,17 in mem-ory tasks. PageRank measured the probability for a randomwalker to visit a given node, indicating how words are cen-tral for random navigation of conceptual associations9. Onthe multiplex network, the definition of PageRank versatilitywas adopted40,41,54. In semantic networks, words with highPageRank were found to require a smaller effort in terms ofmemory search60. Betweenness counted the number of short-est paths going through a given node either on individuallayers or on the whole multiplex network. In this way, be-tweenness could capture how relevant concepts are as "bottle-necks" for the diffusion of spreading activation signals12.Both betweenness and PageRank represented meso-scale net-work measures9, crucially depending on how clusters of con-cepts are associated with each other. Closeness indicated theinverse of the average number of conceptual associations con-necting a word to every other concept9. In case of a uniformdiffusion of spreading activation across all concepts, closenesswould capture the potential for a concept to attract and storeactivation signals at a global network level2. On multiplexnetworks, high-closeness concepts were found to be acquiredearlier during language learning28,31 and were also identifiedas easier to be correctly named by people with aphasia27.

For aphasia-inspired attacks, two strategies were deployed.In a "small-to-big" attack strategy, words with a lower prob-ability for correct naming pc(w) were attacked first, simulat-ing a more realistic progressive decline were more difficult-to-produce words became inactive in the mental lexicon first.A less realistic reverse strategy was simulated as well. In a"big-to-small" attack, words with a higher probability for cor-rect naming pc(w) were attacked first, simulating a worst-casescenario were the mental lexicon started failing mainly with

Robustness of the mental lexicon 6

words better processed by a population of people with anomicaphasia. This strategy is a "worst-case" scenario because ittargets first words whose cognitive processing is still relativelyefficient even in a clinical population with a diagnosed clinicallanguage impairment. These two attacks represent a big dataapproach to testing how the large scale structure of the mentallexicon behaves under modelled cognitive decline specificallybased on clinical data.

III. RESULTS

This section is divided in three different parts. The firstone outlines the robustness of the global connectivity of themultiplex lexical network. The second part focuses on thespecific connectedness of the largest viable component23,41.Both these network measures aim at assessing the overallrobustness of the mental lexicon to progressive word fail-ures/attacks/removals. Given the richness of the LVC44 interms of words of relevance for language use, acquisition andprocessing (cfr. Methods), the results on its connectednessprovide detailed additional information on how associationsbetween specifically "core words" in the mental lexicon de-cline over progressive attacks. The third and last part testshow fragility patterns evolve over time in the mental lexiconwhen semantic connections age and weaken.

A. Robustness of the networked mental lexicon in terms oflargest connected component

The size of the LCC identifies the number of words con-nected by either semantic or phonological conceptual asso-ciations. Its decline depends on the specific way of remov-ing/attacking words. As reported in the Methods section, in alltarget attacks node removal was linear (i.e., removing top 100nodes, top 200 nodes, top 300 nodes, etc.). Hence, a lineardecline in LCC is considered to be the lowest damage pos-sible to LCC size. Such linear decline is called also grace-ful decline9,23 and it represents evidence for robustness to therelative attack strategy. Other behaviours are possible25, e.g.non-linear fall in LCC size or phase transitions. The trend ofthe LCC size over the number of removed nodes is also calledrobustness profile22.

Figure 2 reports the robustness profiles of the networkedmultiplex representation of the mental lexicon, adopted here,under different attacks. For visual clarity, profiles are clus-tered in two groups and reported in two subpanels (Fig.2, topleft and right).

The networked mental lexicon is found to be resilient torandom node attacks, since its LCC displays a graceful de-cline in size when nodes are increasingly removed (cfr. Fig.2,top right). In other words, random failures of words are not ca-pable of crippling the connectivity of conceptual associationsbetween words by erasing or drastically reducing the size ofthe largest connected component.

A lower resilience of multiplex, networked mental lexi-con is found across all psycholinguistics-driven attacks (cfr.

Fig.2, top left). Removing words of shorter length, higherfrequency, acquired earlier or with more meanings does notcompletely disrupt the LCC when 50% of the concepts havebeen attacked. However, these attacks can significantly impairconnectedness. For instance, removing up to top 50% wordsof lower age of acquisition reduces the LCC up to 25% ofits original size. Age of acquisition is found to be the mostdisruptive psycholinguistic strategy, indicating a structural or-ganisation of concepts in the mental lexicon significantly re-lying over age of acquisition rather than on frequency or othermetrics. Alternatively put, the robustness profile indicates thatwords acquired earlier are more relevant than words of higherfrequency, shorter length or more meaning in order to connectdifferent concepts together across semantic and phonologicalaspects of language. This finding is in agreement with previ-ous small-scale studying showing the predominance of age ofacquisition over frequency in driving lexical retrieval61.

Network-driven attacks can dismantle the networked repre-sentation of the mental lexicon significantly more than psy-cholinguistic strategies (cfr. Fig.2, top left and right). Acatastrophic62 phase transition over the size of the LCC isfound in the multiplex degree-based attack strategy (Fig.2, topright). Whereas removing words with the most free associa-tions (phonological similarities) cripples the LCC to a limitedextent, analogous of psycholinguistic attacks (Fig.2, left), in-stead removing nodes based on the multiplex combination ofsemantics and phonology leads to a disruptive attack strategy,where the LCC vanishes completely through a catastrophicphase transition after only 33% of words are removed fromthe multiplex network. Such phase transition could not beobserved in case semantic and phonological layers were ob-served individually and it is therefore a phenomenon emergingfrom the multiplex structure of the mental lexicon. The com-plete disruption of a largest connected component containinga large fraction of nodes when only a part of nodes is attackedindicates a lack of robustness or rather a network fragility23,25.In other words, the modelled representation of the mental lex-icon is found to be fragile to local attacks mediated by mul-tiplex degree. The same lexical structure is more resilient toattacks based on single-layer network degree, indicating therelevance of considering semantics and phonology together asa multiplex network when investigating the robustness of themental lexicon to progressive cognitive decline.

Other multiplex metrics like PageRank versatility and mul-tiplex betweenness could dismantle the LCC completely likemultiplex degree, though by using considerably more attacks.Single-layer PageRank, betweenness and closeness attacks re-ported analogous results to single-layer degrees and were notincluded in Fig. 2 for brevity. In general, in large-scale single-layer networks all these measures are strongly positively cor-related, so that analogous results are expected. Interestingly,the multiplex lexical network resulted being resilient to at-tacks based on multiplex closeness, which were unable to dis-mantle the whole LCC as quickly as other multiplex attacksdid. This resilience indicates that the layout of paths connect-ing concepts is not localised only in shortest paths, which arecaptured by closeness, but also in longer, more winded, pathsof conceptual associations. In other words, connectivity of

Robustness of the mental lexicon 7

FIG. 2. Top: Robustness profiles of the largest connected component (LCC) for a multiplex lexical network representing semantic andphonological associations in the mental lexicon of an adult English speaker. Bottom: Robustness profiles for random configuration modelswith the same degree distribution of above. All results are averaged over 20 configuration models. In all subpanels, in targeted attacks, wordswere ranked and removed at uniform steps (i.e., removing top 100 nodes, top 200 nodes, top 300 nodes, etc.). Progressively attacking/removingnodes decreases the size of the LCC in different ways, depending on attack strategies. The dashed line represents the expected decline in LCCsize due only to the decrease in total number of nodes available in the network, without considering network links.

concepts in the mental lexicon does not rely solely on shortestpaths but it is rather more de-localised, underlining the needfor developing additional measures of path-based centralitybeyond closeness itself55.

Aphasia-inspired attacks reported two distinct behaviours.Removing first words with smaller probability of correct nam-ing pc(w) produced even less damage than random attacksup until when 50% of words were removed. This patternrepresents evidence that patients with anomic aphasia do notlose the ability to name correctly words uniformly at ran-dom but rather tend to fail in finding words peripheral in themultiplex lexical network and for language use and acquisi-tion, extending previous small-scale results8,27,45,52. Remov-ing these peripheral words leaves the mental lexicon relativelyunscathed in terms of conceptual connectedness, i.e. the LCCsize falls gracefully under removal of large portions of periph-eral words. The overall multiplex structure of the mental lexi-con is resilient to progressive failing of words as indicated bybehavioural data in anomic aphasics. Assuming that languagedegradation in people with aphasia corresponded completelyto a degraded lexical representation8,19, ruling out functionalsearch deficits52 on such structure and considering only in-creasingly more and more failing words, this result indicates

that the structure of the mental lexicon itself would degradeand fall in a way as to minimally impact conceptual connec-tivity, a graceful decline indicating a strong resilience to cog-nitive degradation in anomia.

If the mental lexicon is robust to removing words whoselexical retrievals fails first in people with anomic aphasia andsuch words are peripheral in the multiplex structure of themental lexicon, then what about robustness to removal of corewords, instead? Fig.2 (top right) reports the robustness pro-file of the mental lexicon when words easier to name foranomic aphasics are removed first, i.e. "big-to-small" attacks.The LCC does not vanish as quickly as in the case of mul-tiplex attacks, but it is still reduced to 16% of its originalsize when only 50% of words in the whole network are at-tacked/removed. Hence, these words with high pc(w) are in-deed core words in the global lexicon structure. This resultindicates that aphasia-inspired attacks can capture relevanceof words in the mental lexicon and produce sensible disrup-tions of conceptual connectivity. The effects of such attacksdrastically depend on the type of words targeted by them, e.g.peripheral vs core.

Resilience to random attacks but fragility to degree-basedattacks is a feature displayed by scale-free networks9, where

Robustness of the mental lexicon 8

hubs are fundamental for connecting nodes since they detain alarge fraction of links. As reported also in a previous work30,both the free associations and phonological layers reportedhere have a heavy-tailed degree distribution featuring hubs.In order to understand to what extent presence of hubs deter-mines the robustness/fragility of the empirical multiplex net-work, a null model made of configuration layers is consid-ered. Configuration models randomise links by keeping fixedthe degree of every node9. In this case, configuration mod-els disrupt the meaning of conceptual associations26 but keepfixed the phonological neighbourhood density and the seman-tic richness (i.e. the degrees) of every word.

Figure 2 reports the average robustness profiles estimatedover a sample of 20 configuration models of the original se-mantic and phonological layers. Psycholinguistic norms werekept but conceptual associations were randomised. The nullmodels display different robustness features in comparison tothe empirical multiplex network.

Psycholinguistic norms are less effective in disrupting therandom multiplex network, see also Fig. 2 (bottom left). Themost drastic difference is observed over attacks based on theage of acquisition. This quantitative result indicates the pres-ence of correlations between psycholinguistic norms and con-ceptual links beyond the local scale/node degree (which is pre-served by configuration models), further underlining the needfor developing cognitive network metrics in conjunction withpsycholinguistic norms beyond the scale of word degree, rich-ness and neighbourhood density.

The random multiplex network is still fragile to multiplexdegree but more robust to single-layer degree attacks. Fig. 2(bottom right) highlights a continuous phase transition in thesize of the LCC when more and more nodes are attacked basedon multiplex degree. However, notice that this transition hap-pens when 1000 more words are removed in comparison tothe empirical case. Hence, the empirical multiplex network ismore fragile than random expectation to multiplex degree at-tacks. Degree, which is preserved in configuration models, isnot enough for explaining this fragility by itself. Such differ-ence indicates that words with many semantic and phonologi-cal associations are highly relevant in the empirical multiplexnetwork not only because of their degree but also because ofthe overall meso-scale and global connectivity of the mentallexicon. These robustness profiles represent a strong quantita-tive evidence that in the mental lexicon degree correlates alsowith large-scale features of conceptual associations, which to-gether enable the identification of central words and whoseremoval dismantles catastrophically62 (i.e. through a phasetransition) connectivity among concepts.

Another difference with the empirical case is that in thenull model multiplex betweenness, PageRank and multidegreegive rise to an analogous phase transition whereas in the em-pirical network these attack strategies provided strongly dif-ferent results. This represents additional evidence that heavy-tailed degree distributions are not enough for explaining therobustness and fragility patterns observed in the empiricalmultiplex lexical network. The empirical multiplex networkis more robust than random expectation to meso-scale attacks,as captured by betweenness and Page-Rank. This indicates

the presence of empirical conceptual associations in the hu-man mental lexicon providing robustness to degradation anddisrupted by the random rewiring of configuration models.

B. Robustness of the networked mental lexicon in terms oflargest viable cluster

Figure 3 reports the robustness profiles for the empiricalLVC (left) and its randomised counterpart through configu-ration models (right). The empirical LVC is fragile to mul-tiplex PageRank attacks, indicating that words in the viablecluster are versatile40,41,54 and hence fundamental for transi-tioning between the semantic and phonological aspects of themental lexicon through a random walk. The empirical multi-plex network is fragile, to a lesser extent, also to attacks basedon multiplex closeness, association degree and multidegree.Aphasia-inspired attacks targeting words with a higher prob-ability of correct naming are able to fully dismantle the LVCby removing only 20% of its nodes. All these attacks displayalso a sudden discontinuity over the LVC size, reminiscent ofthe explosive phase transition detected during cognitive de-velopment by Stella and colleagues26. Explosive phase tran-sitions were observed also in multidegree attacks of multiplexnetworks made of random graphs25. However, these discon-tinuous transitions were not reproduced by the configurationmodels investigated here (see Figure 3, right), mainly becauseof the heavy-tailed degree distribution observed in free asso-ciations and phonological similarities.

No psycholinguistic-based attack is able to fully disman-tle the LVC, indicating that the information encapsulated inconceptual associations and captured by network metrics arequalitatively different from the information identified by lin-guistic norms. This finding represents additional evidencethat psycholinguistic measures of word relevance and networkcentralities provide different information and should be there-fore used in a synergistic combination in order to obtain aricher description of word relevance for language acquisitionand use.

The fragility of the empirical LVC to multiplex PageRank,multidegree and multiplex closeness cannot be explained bydegree itself as in the random models no discontinuous phasetransition is present when only 800 or less words are attacked.This indicates that the catastrophic, sudden phase transitionsobserved in the empirical multiplex network are a linguisticpattern that is not a direct sequence of heavy-tailed degreedistributions. In other words, the LVC being more fragilethan random expectation to these kinds of attacks indicatesthat concepts included in the viable component are relevant tothe mental lexicon not only in terms of their local degree butalso in terms of meso-scale and large-scale network patterns.

Removing words of lower correct naming probability pc(w)for people with aphasia resulted in minimal damage and agraceful decline of the LVC size equivalent to word length.This indicates that the words that people with aphasia findthe most difficult to name in picture naming tasks are mainlywords in the periphery of semantic and phonological net-works, outside the core of relevant words identified by the

Robustness of the mental lexicon 9

FIG. 3. Left: Robustness profiles of the largest viable cluster (LVC) for a multiplex lexical network representing semantic and phonologicalassociations in the mental lexicon of an average adult English speaker. Right: Robustness profiles for random configuration models with thesame degree distribution of above. All results are averaged over 20 configuration models. In all subpanels, in targeted attacks, words wereranked and removed at uniform steps (i.e., removing top 50 nodes, top 100 nodes, top 150 nodes, etc.). Progressively attacking/removingnodes decreases the size of the LVC in different ways, depending on attack strategies. Removing words of lower pc(w) resulted in minimaldamage, equivalent to word length attacks, and was not reported for brevity.

LVC.On the other hand, the LVC is fragile to removal of those

words that are found easier to be named by people withanomic aphasia, i.e. attacking words with higher pc(w) first.This represents a strong quantitative indication that beingwithin the viable cluster correlates positively with a facilita-tive effect over word retrieval.

Although this analysis cannot shed light over causal rela-tionships, it highlights the clinical relevance of the LVC asa sub-component of the mental lexicon where having multi-ple semantic and phonological paths between words correlateswith a shielding effect against cognitive decline and failinglexical processing. Words in the LVC are found to be eas-ier to be produced than words outside of the LVC, hence thefragility to the "big-to-small" aphasia-inspired attack.

1. Persistence of catastrophic transitions in the robustnessof the mental lexicon across age groups

The mental lexicon is not a fixed cognitive system but itrather evolves over time. Free associations indicate memoryrecall patterns that can weaken with ageing58, thus leading tomore rarefied or less clustered network structures over time, asreported also in previous studies57. In this way, it is importantto assess the robustness of the mental lexicon across differ-ent age groups, in order to monitor potential changes due toageing. Ageing was modelled only at the semantic level, forwhich different networks of free associations were used (cfr.Methods).

Figure 4 reports the robustness profiles of the LCC (LVC)under multidegree (multiplex PageRank) attacks for differentageing stages. These two attack strategies were selected andtested across age groups because they resulted being the mostdisruptive for the LCC and LVC, respectively. Despite fluc-tuations due to small differences in the link density of eachnetwork, all multiplex representations of the mental lexicon

FIG. 4. Top: Robustness profiles under multidegree attacks of thelargest connected component (LCC) for a multiplex lexical networkrepresenting semantic and phonological associations in the mentallexicon of an adult English speaker of a given age group. Bottom:Robustness profiles under multiplex PageRank for the largest viablecluster (LCC) for a multiplex lexical network representing seman-tic and phonological associations in the mental lexicon of an adultEnglish speaker of a given age group.

Robustness of the mental lexicon 10

across age groups featured a catastrophic phase transition inthe LCC size under multidegree attacks when roughly 5000words were attacked. The attack of the LVC by means ofmultiplex PageRank displayed a larger variability across agegroups, with the layer of young adults between ages 22 and28 displaying the most fragile LVC. Across all age groups, thelargest viable component vanished when a rather small frac-tion of high multiplex PageRank words were removed.

As a result, the same fragility patterns observed above foran age-aggregated mental lexicon also persist across differentage-specific multiplex lexical networks.

IV. DISCUSSION

This study investigated the robustness of the mental lexi-con towards progressive word failure through a network ap-proach. The associative structure of concepts in the men-tal lexicon was modelled as a minimal multiplex networkcombining two aspects of language: semantic relationships(through free associations10,11,36) and sound patterns (throughphonological similarities15,18).

The combination of these two aspects of the mental lex-icon led to results absent in the individual single-layer net-works. In fact, the whole networked mental lexicon exhibiteda catastrophic62 phase transition over node connectivity whennodes of high multiplex degree where attacked first. Attack-ing nodes based only on their semantic richness11,35,39 (i.e.free association degree) or their phonological neighbourhooddensity15,17 (i.e. phonological degree) did not exhibit anyphase transition on the size of the largest connected compo-nent of concepts in the mental lexicon. Attacking words ac-cording to the sum of these degrees, i.e. their multidegree40,dismantled completely connectedness when only half avail-able concepts where attacked. Such evidence indicates afragility in the layout of conceptual associations that can beobserved only when semantics and phonology are representedin conjunction. This motivates the need for pursuing lexi-cal representations4,5,27,29 accounting for semantic and phono-logical interactions in acquiring, storing and processing lan-guage.

Considering null models preserving node degree but ran-domising conceptual links, the empirical networked mentallexicon was found to be more fragile than random expecta-tion. This quantitative comparison indicates that hubs in thecognitive layout of conceptual associations are important notonly because of their degree but also because of their overallorganisation and connectivity with other concepts. Remov-ing hubs efficiently disrupted connectedness across concepts.Connectedness is considered to be fundamental for lexical re-trieval in the commonly adopted revised model of spreadingactivation2,5,35,38,43. Activation is a signal propagating oversemantic and phonological conceptual associations and ac-cumulating over target concepts that have to be identified orrecalled. The absence of connectedness disrupts lexical re-trieval, inhibiting language processing within the mental lexi-con. In this way, the identified catastrophic phase transition isa trademark of the importance of hub concepts in the mental

lexicon for guaranteeing semantic and phonological associa-tions, which, combined, connect most concepts together andoffer channels for the flow of activation spreading, word pro-cessing, identification and recall3.

Although fragile to multiplex degree-attack, the mentallexicon was considerably robust to progressive word fail-ure driven by psycholinguistic features, e.g. attacking firsthigh frequency or shorter words. The graceful decline23 ob-served in connectivity when attacking words based on theirfrequency, length, age of acquisition or number of differentmeanings indicates that the mental lexicon structure includesadditional information that is not fully captured by individualpsycholinguistic features. This result indicates the importanceof investigating the mental lexicon structure not only in termsof individual word features48 but also in terms of networkmetrics, including information on how concepts are connectedand organised through conceptual associations.

Indeed, network attacks were able to dismantle the lan-guage kernel59 represented by the largest viable cluster44

(LVC) more efficiently than random expectation and psy-cholinguistic metrics. In agreement with previous studies26,30,the LVC was found to be richer in more commonly usedwords. Furthermore, previous approaches26 identified theLVC as providing short connections between concepts andhighlighted different access strategies to the LVC itself dur-ing the mental navigation performed in fluency tasks by peo-ple with high and low creativity levels30, respectively. Theanalysis of aphasic data32 reported here identified the LVCas richer in words whose correct naming is easier for peoplewith a diagnosis of anomic aphasia52,53 in comparison to therest of the mental lexicon. This facilitative effect in producinga word across aphasic populations, in addition to the aboveresults, identifies the LVC as a region of cognitive relevancefor language processing within the mental lexicon structure.

Multiplex degree was effective in dismantling also theLVC but not as much as multiplex PageRank. Previousstudies40,41,54 reported multiplex PageRank as capable ofidentifying bottle-neck nodes fundamental for driving randomwalks across two distinct network layers. Given that words ina viable component have to be connected by paths made onlyof semantic links and, at the same time, by paths made only ofphonological associations, it can be expected for such viablenodes, as captured by multiplex PageRank, to act as bridgingconcepts for the whole mental lexicon. In this way, an inter-esting future research direction opened by the above resultsis assessing the role of the LVC for language processing ofconcepts far apart in the mental lexicon, through the design ofspecific psycholinguistic tasks.

Also the LVC was found to be resilient to attacks basedon psycholinguistic norms but it displayed a higher fragilityto aphasia-inspired attacks, which at the same time did notcritically dismantle connectedness in the whole multiplex net-work. These aphasia-inspired attacks modelled cognitive de-cline as informed from data about correct word production andpicture naming in people with anomic aphasia. The currentapproach represents a first-of-its-own extension of pioneeringnetwork investigations of cognitive decline19,20 to the richerphenomenology of multiplex networks and psycholinguistics-

Robustness of the mental lexicon 11

driven attacks, a possibility opened by the recent availabilityof Big Data also in clinical populations48,50.

Attacking first words that are less correctly named by peo-ple with anomic aphasia did not hamper either the LCCor the LVC in any drastic way, providing quantitative sup-port to the graceful degradation hypothesis in cognitiveneuroscience52,53 that cognitive decline in anomia and re-lated cognitive impairments starts preferentially by impact-ing linguistic skills in non-severe ways. In the current anal-ysis, the first words that failed in aphasia-inspired attackswere indeed peripheral for the connectivity and viability ofthe whole, multiplex, associative structure of the mental lexi-con structure, thus having an expected limited negative impactover linguistic skills. Understanding the correlations betweenstructural degradation and linguistic skills remains an openchallenge3,8,45,53 of relevance for future research.

Attacking first words more correctly named by people withanomic aphasia did not dismantle effectively connectednessbut it provoked an explosive, discontinuous25,44 phase transi-tion in the LVC, effectively destroying it through a relativelysmall number of attacks. This difference in fragility providesevidence that words whose lexical retrieval is damaged lessby anomic aphasia are mainly located within the LVC but arenot as fundamental as hubs for connectedness of the wholemental lexicon. Alternatively, words correctly identified andproduced by people with anomic aphasia are not mainly hubwords but rather versatile nodes, as identified by multiplexPageRank and being part of the largest viable cluster. Theseresults enable the hypothesis that the LVC facilitates the pro-duction of words important for transitioning between phonol-ogy and semantics and not for achieving global connected-ness. These versatile words could be shielded against cogni-tive decline. Unfortunately the data-informed approach usedhere enables the identification of a correlation between "cor-rect naming", "being in the LVC" and cognitive decline, with-out allowing for the identification of causal links. The aboveresults open an interesting way for studying cognitive declinein people with aphasia by assessing better the role that wordsin the LVC might have in hampering or promoting the re-training7,46,52 of linguistic skills.

Interestingly, the above patterns are preserved even throughageing, when the multiplex structure is changed in order toaccount for alterations57,58 in semantic memory patterns (i.e.free associations) provided by young adults, adults and olderadults. At all ageing stages, the multiplex mental lexicon isfound to be fragile to multiplex degree-attacks while the LVCremains fragile to multiplex PageRank attacks. This persis-tence might be an indication that although memory patternscan weaken with ageing58, the organisation of the mental lexi-con in terms of interacting phonological and semantic systemsalways presents both hubs guaranteeing global connectednessand viable nodes brokering spreading activation flow betweensemantic and phonological layers.

It is important to underline that the above results are basedon specific assumptions and suffer from a few limitations. Allthe attacks reported here simulated progressive word failurewhere, once attacked, a concept did never recover20,23. Al-though in line with progressive disruptive clinical pathologies

where language can be completely lost, like in Alzheimer’sDisease3, in other cognitive pathologies language skills canbe recovered even after a long time since the onset of the dis-ease, like in some instances of aphasia63. The gathering ofmore specific data, linking language recovery and linguisticskills to different snapshots of the mental lexicon, could beused for designing realistic recovery probabilities after wordattack/removal. The main issue in this case, common also tothe aphasia-inspired attacks used here, is sample size: Datawas available only for 173 words from the PNT32,50. Al-though logistic regression is expected to work correctly alsowithin small sample sizes31, the extension of the PhiladelphiaNaming Task to larger pools of words would be highly advan-tageous for obtaining more realistic and detailed estimates ofword failures in people with cognitive impairments.

Another limitation of the current study is the assumptionthat the mental lexicon structure is fixed over time and is thesame across different people. Although the first downsidewas partially addressed by considering time-dependent freeassociations, at the best of the author’s knowledge there isno large-scale dataset about how phonological relationshipsevolve over time yet. The so-called Transmission DeficitHypothesis58 indicates that older adults can incur into seman-tic activation of words that do not correspond to a phonolog-ical activation, thus inhibiting word production and leavingwith a vague awareness of knowing a word without beingable to produce it, a phenomenon known also as tip-of-the-tongue6. In this way, ageing might have a more predominanteffect over semantic rather than on phonological links, thusjustifying the approximation of considering only weakenedsemantic associations age groups. Nonetheless, within the ro-bustness framework outlined in this work, novel datasets in-vestigating the structural decline of phonological and seman-tic associations, combined, would prove extremely valuablefor better understanding failed lexical retrieval. In order toaddress also linguistic variability, individual networks of freeand phonological similarities might be built through appropri-ate data collection34 and investigated within the same robust-ness framework summarised here.

Notice also that the current analysis is limited to inves-tigating only structural changes, and not functional nor dy-namical ones, in the mental lexicon. This approximationdoes not keep into account cognitive search strategies3,52 oversuch structure, that might also degrade along with cognitiveimpairments7,8,51 or ageing58. The debate between structureand dynamics represents a relatively open field of investi-gation in psycholinguistics, where the most commonly ac-cepted hypothesis is that search strategies and mental lexiconstructure are strongly inter-dependent3. Recent investigationslike fluency performance in people with different creativitylevels30 have indicated that structural multiplex measures insynergy with psycholinguistic data are powerful enough forhighlighting different search strategies on the mental lexicon,detecting differences in the way people with high and low cre-ativity levels explore and search concepts in the animal cat-egory task. Similarly, other approaches based on multiplexstructure have been successful in predicting picture namingperformance of people with aphasia27. Through a different

Robustness of the mental lexicon 12

network construction, structural network metrics were power-ful in predicting also cognitive degratation due to Alzheimer’sdisease14,33. All these approaches indicate that investigatingthe network structure of the mental lexicon can be considereda first important step for understanding language processing,providing the foundations for analysing other important ele-ments of the lexicon, such as search strategies, through addi-tional research.

The framework of network robustness outlined here opensnew ways for transdisciplinary investigations of the interplaybetween language processing and the degrading multi-layeredstructure of the mental lexicon. Such ambitious aim can beachieved only through a synergy of psycholinguistic word fea-tures, cognitive science theory, clinical datasets and access tolarge-scale, multiplex representations of conceptual associa-tions. Quantitative models exploiting such synergy representan innovative way for developing data-informed, potentiallyindividually developed models of cognitive decline and lan-guage restoration46 in clinical patients.

V. ACKNOWLEDGEMENTS

The author acknowledges Dr. Nichol Castro, University ofWashington, for insightful discussion in the early stages ofthis work.

1A. Baronchelli, R. Ferrer-i Cancho, R. Pastor-Satorras, N. Chater, andM. H. Christiansen, “Networks in cognitive science,” Trends in cognitivesciences 17, 348–360 (2013).

2C. S. Siew, D. U. Wulff, N. M. Beckage, and Y. N. Kenett, “Cognitivenetwork science: A review of research on cognition through the lens ofnetwork representations, processes, and dynamics,” (2018).

3J. Aitchison, Words in the mind: An introduction to the mental lexicon (JohnWiley & Sons, 2012).

4M. S. Vitevitch, C. Siew, and N. Castro, “Spoken word recognition,” TheOxford Handbook of Psycholinguistics , 31 (2018).

5B. Dóczi, “An overview of conceptual models and theories of lexical rep-resentation in the mental lexicon,” The Routledge Handbook of VocabularyStudies , 46 (2019).

6R. Brown and D. Mcniell, “The "tip-of-the-tongue" phenomenon,” Journalof Verbal Learning and Verbal Behavior 6 (1966).

7G. S. Dell, M. F. Schwartz, N. Martin, E. M. Saffran, and D. A. Gagnon,“Lexical access in aphasic and nonaphasic speakers.” Psychological review104, 801 (1997).

8V. Erdeljac and M. Sekulic, “Syntactic-semantic relationships in the mentallexicon of aphasic patients,” Clinical linguistics & phonetics 22, 795–803(2008).

9M. Newman, Networks (Oxford university press, 2018).10S. De Deyne, Y. N. Kenett, D. Anaki, M. Faust, and D. J. Navarro, “Large-

scale network representations of semantics in the mental lexicon,” (2016).11S. De Deyne, D. J. Navarro, and G. Storms, “Better explanations of lexical

and semantic cognition using networks derived from continued rather thansingle-word associations,” Behavior research methods 45, 480–498 (2013).

12N. M. Beckage and E. Colunga, “Language networks as models of cogni-tion: Understanding cognition through language,” in Towards a TheoreticalFramework for Analyzing Complex Linguistic Networks (Springer, 2016)pp. 3–28.

13Y. N. Kenett, D. Y. Kenett, E. Ben-Jacob, and M. Faust, “Global and localfeatures of semantic networks: Evidence from the hebrew mental lexicon,”PloS one 6, e23912 (2011).

14J. C. Zemla and J. L. Austerweil, “Analyzing knowledge retrieval impair-ments associated with alzheimer’s disease using network analyses,” Com-plexity 2019 (2019).

15M. S. Vitevitch, “What can graph theory tell us about word learning and lex-ical retrieval?” Journal of Speech, Language, and Hearing Research (2008).

16M. Stella and M. Brede, “Patterns in the english language: phonologicalnetworks, percolation and assembly models,” Journal of Statistical Mechan-ics: Theory and Experiment 2015, P05006 (2015).

17C. S. Siew and M. S. Vitevitch, “Spoken word recognition and serial recallof words from components in the phonological network.” Journal of Exper-imental Psychology: Learning, Memory, and Cognition 42, 394 (2016).

18K. D. Neergaard and C.-R. Huang, “Constructing the mandarin phonologi-cal network: novel syllable inventory used to identify schematic segmenta-tion,” Complexity 2019 (2019).

19J. Borge-Holthoefer, Y. Moreno, and A. Arenas, “Modeling abnormalpriming in alzheimer’s patients with a free association network,” PloS one6, e22651 (2011).

20J. Borge-Holthoefer, Y. Moreno, and A. Arenas, “Topological versus dy-namical robustness in a lexical network,” International Journal of Bifurca-tion and Chaos 22, 1250157 (2012).

21D. S. Callaway, M. E. Newman, S. H. Strogatz, and D. J. Watts, “Networkrobustness and fragility: Percolation on random graphs,” Physical reviewletters 85, 5468 (2000).

22T. L. Frantz, M. Cataldo, and K. M. Carley, “Robustness of centrality mea-sures under uncertainty: Examining the role of network topology,” Compu-tational and Mathematical Organization Theory 15, 303 (2009).

23S. Iyer, T. Killingback, B. Sundaram, and Z. Wang, “Attack robustness andcentrality of complex networks,” PloS one 8, e59613 (2013).

24D.-w. Zhao, L.-h. Wang, Y.-f. Zhi, J. Zhang, and Z. Wang, “The robustnessof multiplex networks under layer node-based attack,” Scientific reports 6,24304 (2016).

25G. Baxter, G. Timár, and J. Mendes, “Targeted damage to interdependentnetworks,” Physical Review E 98, 032307 (2018).

26M. Stella, N. M. Beckage, M. Brede, and M. De Domenico, “Multiplexmodel of mental lexicon reveals explosive learning in humans,” Scientificreports 8, 2259 (2018).

27N. Castro and M. Stella, “The multiplex structure of the mental lexicon in-fluences picture naming in people with aphasia,” Journal of Complex Net-works cnz012 (2019).

28M. Stella, N. M. Beckage, and M. Brede, “Multiplex lexical networks re-veal patterns in early word acquisition in children,” Scientific Reports 7,46730 (2017).

29M. Stella, “Cohort and rhyme priming emerge from the multiplex networkstructure of the mental lexicon,” Complexity 2018 (2018).

30M. Stella and Y. N. Kenett, “Viability in multiplex lexical networks andmachine learning characterizes human creativity,” Big Data and CognitiveComputing 3, 45 (2019).

31M. Stella, “Modelling early word acquisition through multiplex lexical net-works and machine learning,” Big Data and Cognitive Computing 3, 10(2019).

32A. Roach, M. F. Schwartz, N. Martin, R. S. Grewal, and A. Brecher, “Thephiladelphia naming test: Scoring and rationale,” Clinical aphasiology 24,121–133 (1996).

33J. C. Zemla and J. L. Austerweil, “Estimating semantic networks of groupsand individuals from fluency data,” Computational Brain & Behavior 1, 36–58 (2018).

34Y. N. Kenett, D. Anaki, and M. Faust, “Investigating the structure of se-mantic networks in low and high creative persons,” Frontiers in human neu-roscience 8, 407 (2014).

35Y. N. Kenett, E. Levi, D. Anaki, and M. Faust, “The semantic distance task:Quantifying semantic distance with semantic network path length.” Journalof Experimental Psychology: Learning, Memory, and Cognition 43, 1470(2017).

36S. De Deyne, D. J. Navarro, A. Perfors, M. Brysbaert, and G. Storms, “The“small world of words” english word association norms for over 12,000 cuewords,” Behavior research methods , 1–20.

37D. R. Amancio, O. N. Oliveira Jr, and L. da F Costa, “Using complexnetworks to quantify consistency in the use of words,” Journal of StatisticalMechanics: Theory and Experiment 2012, P01004 (2012).

38R. Goldstein and M. S. Vitevitch, “The influence of closeness centrality onlexical processing,” Frontiers in psychology 8, 1683 (2017).

39Y. N. Kenett, “What can quantitative measures of semantic distance tellus about creativity?” Current Opinion in Behavioral Sciences 27, 11–16

Robustness of the mental lexicon 13

(2019).40F. Battiston, V. Nicosia, and V. Latora, “The new challenges of multiplex

networks: Measures and models,” The European Physical Journal SpecialTopics 226, 401–416 (2017).

41G. Bianconi, Multilayer Networks: Structure and Function (Oxford univer-sity press, 2018).

42S. Arbesman, S. H. Strogatz, and M. S. Vitevitch, “The structure of phono-logical networks across multiple languages,” International Journal of Bifur-cation and Chaos 20, 679–685 (2010).

43K. Bock and W. J. Levelt, Language production: Grammatical encoding(Academic Press, 1994).

44G. J. Baxter, D. Cellai, S. N. Dorogovtsev, A. V. Goltsev, and J. F. Mendes,“A unified approach to percolation processes on multiplex networks,” inInterconnected Networks (Springer, 2016) pp. 101–123.

45M. S. Vitevitch and N. Castro, “Using network science in the language sci-ences and clinic,” International journal of speech-language pathology 17,13–25 (2015).

46T. McAllister and K. J. Ballard, “Bringing advanced speech processingtechnology to the clinical management of speech disorders,” InternationalJournal of Speech-Language Pathology 20, 581–582 (2018).

47M. Brysbaert and B. New, “Moving beyond kucera and francis: A criticalevaluation of current word frequency norms and the introduction of a newand improved word frequency measure for american english,” Behavior re-search methods 41, 977–990 (2009).

48M. Brysbaert and A. Biemiller, “Test-based age-of-acquisition norms for 44thousand english word meanings,” Behavior research methods 49, 1520–1523 (2017).

49G. A. Miller, “Wordnet: a lexical database for english,” Communications ofthe ACM 38, 39–41 (1995).

50D. Mirman, T. J. Strauss, A. Brecher, G. M. Walker, P. Sobel, G. S. Dell,and M. F. Schwartz, “A large, searchable, web-based database of aphasicperformance on picture naming and other tests of cognitive function,” Cog-nitive neuropsychology 27, 495–504 (2010).

51N. Martin, A. Roach, A. Brecher, and J. Lowery, “Lexical retrieval mech-anisms underlying whole-word perseveration errors in anomic aphasia,”

Aphasiology 12, 319–333 (1998).52A. E. Hillis, “Aphasia: progress in the last quarter of a century,” Neurology

69, 200–213 (2007).53M. Laine and N. Martin, Anomia: Theoretical and clinical aspects (Psy-

chology Press, 2013).54M. De Domenico, A. Solé-Ribalta, E. Omodei, S. Gómez, and A. Arenas,

“Ranking in interconnected multilayer networks reveals versatile nodes,”Nature communications 6, 6868 (2015).

55M. Stella and M. De Domenico, “Distance entropy cartography charac-terises centrality in complex networks,” Entropy 20, 268 (2018).

56M. E. Newman, “Communities, modules and large-scale structure in net-works,” Nature physics 8, 25 (2012).

57H. Dubossarsky, S. De Deyne, and T. T. Hills, “Quantifying the structure offree association networks across the life span.” Developmental psychology53, 1560 (2017).

58D. U. Wulff, S. De Deyne, M. N. Jones, R. Mata, A. L. Consortium,et al., “New perspectives on the aging lexicon,” Trends in cognitive sci-ences (2019).

59R. F. I. Cancho and R. V. Solé, “The small world of human language,”Proceedings of the Royal Society of London. Series B: Biological Sciences268, 2261–2265 (2001).

60T. L. Griffiths, M. Steyvers, and A. Firl, “Google and the mind: Predictingfluency with pagerank,” Psychological Science 18, 1069–1076 (2007).

61C. M. Morrison, A. W. Ellis, and P. T. Quinlan, “Age of acquisition, notword frequency, affects object naming, not object recognition,” Memory &Cognition 20, 705–714 (1992).

62J. L. Casti, Connectivity, complexity and catastrophe in large-scale systems,Vol. 7 (John Wiley & Sons, 1979).

63J. D. Stefaniak, A. D. Halai, and M. A. L. Ralph, “The neural and neuro-computational bases of recovery from post-stroke aphasia,” Nature 41582,019–0282 (2019).