Words matter more than morphemes: Evidence from masked priming with bound-stem stimuli Hélène Giraudo Madeleine Voga Laboratoire CLLE (Equipe ERSS) University Paul-Valéry Montpellier III CNRS & Université Toulouse Jean Jaurès madeleine.voga@univ-montp3.fr giraudo@univ-tlse2.fr se representations within the mental lexicon is a Abstract matter of ongoing research. Two hypotheses can be drawn: according to the first, morphemic units Five masked priming studies were carried correspond to concrete pieces of words (i.e., in order to shed light on the processing of stems and affixes). Complex words are therefore bound-stem words (e.g., terr- in terrible). processed through a decomposition mechanism Both orthographic (e.g., termite) and un- that strips off the affix in order to isolate the related (e.g., montagne ‘montain’) condi- stem. The morphemic nature of the remaining tions stand as baselines for controlling letters is then verified by the system and access morphological effects. The results of the to word representations (i.e., word forms coded experiments using unrelated word con- in the orthographic lexicon) operates via the pre- trols suggest that in the particular case of activation of their constituent morphemes, i.e., bound-stem words, only genuinely de- morphemic representations stand as access units. rived word primes (terrible) produce pos- This mechanism is exemplified by Taft’s model itive effects differing from formal over- (1994), the basic principles of which are fol- lap effects. Morphological effects are in- lowed by many psycholinguistic studies (e.g., terpreted as resulting from both “morce- Crepaldi, Rastle & Davis, 2010). Morphemic me” and “base-lexeme” activations. units are situated between the level of let- 1 Introduction ters/syllables and the word level; consequently, they can only be matched to concrete letter clus- As is broadly admitted, morphologically related ters (i.e., bound-stems, free-stems and affixes) words prime each other in various languages that constitute words. This decompositional (Arabic: Boudelaa & Marslen-Wilson, 2001; mechanism is also insensitive to any semantic English: Rastle, Davis, Marslen-Wilson & Tyler, characteristics of words (i.e., transparent vs. 2000; French: Giraudo & Grainger, 2000; Ger- opaque morphological formation) or to their lex- man and Dutch: Drews and Zwitserlood, 1995; ical environment (in terms of orthographic Hebrew: Frost, Deutsch & Forster, 1997) thus neighborhood or family size). One of the strong suggesting the existence of a morphological level predictions of the decompositional approach is of processing. This kind of study has used differ- that morphological priming effects should vary ent types of materials, words or pseudowords, as following the ease with which constituent mor- well as multiple settings: for the masked priming phemes can be identified/extracted. technique (Forster & Forster, 2003), widely used According to the second hypothesis, morphology to shed light on morphological processes as well is coded at the interface of word and semantic as in this study, the distinction can be made be- representations and corresponds rather to lex- tween designs using only unrelated controls and emes (Aronoff, 1994). Lexeme units are coded at those using both unrelated and orthograph- the interface of the word and the semantic level, ic/phonological controls, as suggested by Girau- organizing the lexicon in terms of morphological do & Grainger (2001) or Pastizzo & Feldman families. The recognition of any complex word (2002). triggers first the activation of all word forms that Even though the existence of a morphological can match with it; a competition is then engaged level of processing is unanimously acknowl- between the pre-activated forms (forms matching edged, the exact nature, locus and the role of the- the input, i.e., those who are morphologically Copyright © by the paper’s authors. Copying permitted for private and academic purposes. In Vito Pirrelli, Claudia Marzi, Marcello Ferro (eds.): Word Structure and Word Usage. Proceedings of the NetWordS Final Conference, Pisa, March 30-April 1, 2015, published at http://ceur-ws.org 19 related but also those who are only orthograph- stem. Psycholinguists tested this difference and ically related) until the right lexical representa- found that processing for free and bound-stems tion reaches its recognition threshold (deter- may differ but both produce significant priming mined by its surface frequency). During this effects (Forster & Azuma, 2000; Järvikivi & competition phase, morphologically related Niemi, 2002; Marslen-Wilson, Tyler, Waksler & words send positive activation to their respective Older, 1994; Pastizzo & Feldman, 2004). Of base lexeme, feeding back activation to them. great importance to our study, Pastizzo and Morphological priming effects result from this Feldman (2004) observed that the magnitude of mechanism of co-activation. Following this su- facilitation varied following the baseline used in pralexical approach (Giraudo & Grainger, 2000; the experiments: equivalent magnitudes of prim- 2001), complex words are not “decomposed”, ing for free and bound-stems were obtained rela- but are able to trigger the activation of their con- tive to an unrelated baseline; with an orthograph- stituent morphemes. In this kind of architecture, ic control however, free-stems produced system- lexeme units are supposed to be abstract enough atically greater priming than bound-stems. The to tolerate variation induced by derivation and interpretation of this line of research suggests inflection (i.e., allomorphy, suppletion, phono- that morphological priming effects are not direct- logical/ morphological truncation, haplology, ly constrained by semantic similarity between verb-noun conversion). In other words, a mor- prime and target. The second handicap, in terms phological unit does not necessarily need to sur- of surface analysis, consists in the difficulty in face in the real world in order to be coded in segmenting the word forms into morphemes. At long-term memory. This organisation, compati- this point, the two different models presented ble with recent neuroimaging data (Lévy, Ha- above give rise to different predictions: accord- goort, Démonet, 2014), also implies that all mor- ing to the morpheme-based approach all complex phemes of a given language are not necessarily forms (free-stem as well as bound-stem words) represented within the mental lexicon: units such are first analyzed in morpheme fragments and as neologisms, hapaxes and nonce words are not then access word representation, in other words, necessarily directly connected with existing the lexicality of the base doesn’t matter. This morphological units; bound-stem words could be approach predicts morphological priming be- such a case. tween derivations (e.g., virus-viral) as well as between the base and its derivation (e.g., vir- viral). 2 The study According to the supralexical approach, the members of a morphological family are linked The present paper focuses on the processing of together by virtue of their common base at the bound-stem words by opposition to free-stem lexeme level; however, the base of bound-stem words. For ex., on one hand, the word viral com- words is not represented at the word level. In this posed of the bound-stem vir-, also present in vi- case, priming effects between related derived rus, virulent, virulence, virology and virologist words (e.g., virus-viral) are expected but no ef- and, on the other hand the word singer composed fect should be observed using their bound-stems with the free-stem sing that forms singing, song, as primes, the access to the base lexeme being etc. Both are defined as being morphologically conditioned by the prior activation of a word complex but while it is evident for the standard form at the word level. speaker/reader that the complex word singer de- Taft and Kougious (2004) investigated this issue rives from the root sing, it is less evident to say in English through a masked priming experi- from which root the complex word viral derives. ment. They compared both semantically and or- The morpheme vir-, which does not have any thographically related words (e.g., virus-viral) to clear meaning in English, can be considered as a merely orthographically related words (e.g., fu- bound-stem whereas sing- in singer is a free- ture-futile) and, unsurprisingly, found facilitation stem. From a processing point of view, the vir- in the former case but not in the latter. Neverthe- viral example can be viewed as a case where the less, the design of this study is not very informa- lexical unit is not directly connected to the mor- tive with respect to the decomposition issue, giv- phological unit, by virtue of its twofold handi- en that the critical condition examining the effect cap: the first aspect is semantic interpretability, of the bound-stem on its derivations has not been i.e., derivations composed with a bound-stem considered. could be less interpretable than those with a free- 20 Our study aims to fill this gap through five visual word primes (e.g., terrible) produced significant masked priming experiments with native French priming effects (33ms), though these conditions speakers. In this kind of protocol, subjects are did not significantly differ from the bound-stem unaware of the presence of the prime, which al- condition (e.g., terr-) whose effect (18ms) did lows minimizing strategy use and examining au- not manage to reach significance. tomatic processing during the early stages of Exp. 4 was designed to see if the advantage for word identification: all five experiments use a the complex word sharing the same bound-stem within-priming (Latin square) design, in which found in exp. 3 holds up to the comparison with we directly compare the effects of different non-word primes constructed with the same primes on the same target. A 57ms prime dura- bound-stem and an existing suffix. The three tion) was used and the task was lexical decision. priming conditions were the following: the mor- Exp. 1 examined morphological effects induced phologically related word sharing its bound-stem by words sharing the same bound-stem, e.g., ter- with the target, e.g., terrible – terreur; a non- rible – terreur ‘terrible-terror’ relative to an or- word made of the same bound-stem and a suffix thographic control baseline, e.g., termite – different to that of the target, e.g., terrage – terreur’ ‘termite-terror’ (where ‘termite’ is a terreur (where -age corresponds to an existing monomorphemic word), as well as an unrelated morpheme); an unrelated control, e.g., montagne baseline (montagne – terreur ‘mountain-terror’). – terreur. The statistical analysis of the results Results show that only truly derived word primes revealed that only related word primes (e.g., ter- produce facilitation, relative to unrelated (36ms rible) produced significant morphological prim- of facilitation) as well as orthographic controls ing (40ms) relative to the unrelated controls. (35ms). However, this first result does not in- Even if the non-word prime condition (e.g., ter- form us about how derived words constructed rage) led to quicker reaction times compared to with a bound-stem are processed: are they ana- the unrelated baseline (688 vs 703ms), it didn’t lyzed in terms of stem + affix or are they global- differ significantly from it. More importantly, the ly processed? Exp. 2 examined the extent to 25ms difference between the word prime condi- which the facilitation we take as morphological tion and the non-word one is statistically signifi- could be due to formal overlap: this is done by cant. This suggests that it takes a real word to using non-existent orthographic controls, sharing induce morphological priming, independently all but one letter with the ‘true’ bound-stem, e.g., and above orthographic low-level perceptual in- for the target terreur, the first possible prime is fluences, to which the masked priming technique the true bound-stem terr- presented in isolation is known to be sensitive. Our results show that (e.g., terr - terreur); the second priming condi- the presence of an existing bound-stem in a non- tion is the non-existing bound-stem condition word does not suffice to produce morphological (orthographic control) tarr- (e.g., tarr – terreur); priming, a finding which contradicts those pub- the third condition is an unrelated baseline (e.g., lished by Longtin and Meunier (2005) as we montag – terreur). Although only true bound- shall see in the discussion. Experiment 5 exam- stems induced significant facilitation relative to ined the extent to which the morphological facili- the unrelated baseline (28ms), the non-existing tation found in exp. 4 could be due to formal fac- stem condition (e.g., tarr-) exhibited reaction tors: in order to test this, we replaced the mor- times (RTs) that did not differ significantly from phologically related word primes by non-words those of the true bound-stem condition. This re- constructed with a bound-stem and a final letter sult highlights the fact, already pointed out by sequence that does not correspond to any suffix Forster (1999), that there is an influence of for- in French. The following three prime conditions mal factors in this kind of protocol, as well as the defined the three levels of the prime type factor: need to include orthographic controls in the de- a complex non-word formed by a bound-stem sign. Experiment 3 directly compared the effects and a suffix, e.g., terrage – terreur (where terr- of complex word primes to those of bound-stem and –age correspond to existing morphemes); a primes: the targets were the same as in Exp. 1 simplex non-word formed by a bound-stem and a and the three levels of the prime type factor were non-existing ending, e.g., terryme – terreur, in the following: a morphologically related suffixed which –yme is not a suffix; finally, an unrelated word sharing the same bound-stem, e.g., terrible non-word, e.g., moitagne – terreur. The statisti- – terreur ‘terrible-terror’; its bound-stem, e.g., cal analysis of the results revealed that both terr - terreur; an unrelated control, e.g., monta- complex and simplex non-word primes produced gne – terreur. Results showed that only complex shorter RTs than unrelated primes (31 and 27ms 21 of effect respectively): both types of prime are Taken together, the results of the experiments able to facilitate target recognition and produce using unrelated word controls (exp. 1, 3 and 4) thus morphological-like facilitation. Neverthe- suggest that in the particular case of bound-stem less, the fact that the effects produced by com- words, only genuinely derived word primes (ter- plex primes (e.g., terrage) did not differ from rible) produce positive effects differing from those produced by simplex non-word primes formal overlap effects. This is true with the ex- (e.g., terryme) leads us to reject any interpreta- ception of exp. 3, where the effect of genuinely tion based on pre-lexical morphological decom- derived word primes did not differ from bound- position. We suggest interpreting this pattern of stem primes (terr-); note however that in this results on the basis of formal criteria: for real experiment, the bound-stem condition did not words it takes a real word to facilitate processing differ from the unrelated condition, while the (exp. 4), but for non-words, given the absence of derived word condition did. This is a demonstra- representation in the word-level, morphological- tion of the fact that “nonwords would be always like priming does nothing but reflect low-level better form-primes than words, even when perceptual similarities, such as between the two masked. The reason is simply because a related non-words (both complex and simplex, ter- word prime will compete more vigorously with rage/terryme) and the target terreur. Besides the the target than a related nonword prime” (For- role attributed to formal factors, the point that ster, 1999: 8). These results are not in accordance should be stressed in the interpretation of exp. 4 with those found by Longtin and Meunier (2005) and 5 is that while in exp. 4 the nonword made using roughly the same priming conditions. In up from an existing bound-stem and an existing their study, derived non-word primes (e.g., suffix (terrage) seems to interfere with pro- garagité) systematically produced significant cessing of the target (terreur) by virtue of its priming effects on target recognition relative to morphological structure, in exp. 5 this interfer- unrelated word controls, while non- ence disappears. The fact that, in the ‘terrage’ morphological non-word primes (e.g., rapiduit) condition (exp. 4) we observe RTs that are not yielded a 29 ms non-significant effect. Two fac- significantly quicker than the unrelated condi- tors can explain these contradictory results: a) tion, despite the existence of a formal overlap the type of unrelated controls: contrary to Long- combined with morphological-like structure tin and Meunier, we examined priming effects (terr-age/terr-eur), can only be due to some kind relative to unrelated non-word primes when the of interference, otherwise we should observe at prime conditions included non-words and word least a small formal effect. This interference nev- primes when the prime conditions included ertheless disappears in exp. 5, since both types of words; b) the type of word targets: given that our non-words (with existing suffix, e.g., terrage, as study focuses on bound-stem words, our targets well as well as non-existing suffix, i.e., simplex are mandatorily complex words, and not bare- non-words such as terryme) lead to significant bases, as in the Longtin & Meunier study. Bare- facilitation. We therefore obtain a different pat- bases are by definition more frequent, and, sub- tern of priming for words (exp. 4) and for non- sequently, easier to activate because of their low- words (exp. 5) which leads us towards an ap- er activation threshold (due to their residual acti- proach where lexicality of the prime does matter vation; for a discussion on this point based on in the overall pattern of results. Even if the pro- McClelland & Rumelhart 1981, see Voga & Gi- cessing system can take advantage of ortho- raudo, 2009; Giraudo & Voga, 2014). graphic similarities between prime and target (and will not prevent itself from doing so, as exp. 2 showed) this does not tell the whole story, and 3 Discussion: On the representation of it certainly does not tell a morphological story: it bound-stem words is just another demonstration of a fact that re- searchers working with masked priming are fa- On the basis of the above results, we can con- miliar with, namely that this technique is sensi- clude that recognition of complex words benefits tive to formal factors (Forster, Mohan & Hector, from two springs of facilitation: a bottom-up ex- 2003). The experiments presented here provide citation from a sublexical level and a top-down evidence that we can use this valuable technique facilitation from a supralexical level. The idea of in order to shed light on truly morphological ef- a double representation for morphology was re- fects, as opposed to morphological-like effects. cently expressed by Diependaele, Sandra & Grainger (2005), suggesting that the morphologi- 22 cal level should be situated both above and be- Crepaldi, D., Rastle, K., Davis, C. 2010. Morphemes in low the word-form level. Subsequently, morpho- their place: Evidence for position-specific identification of suffixes. Memory & Cognition, 38:312-321. logical representations would be either defined as Drews, E. & Zwitserlood, P. (1995). Morphological and morphologically constrained orthographic repre- orthographic similarity in visual word recognition. Jour- sentations (depending on frequencies) or as mor- nal of Experimental Psychology: Human Perception & phologically constrained semantic representa- Performance, 21:1098-1116. Forster, K. I. (1999). The microgenesis of priming effects in tions (coded in terms of regularities in the map- lexical access. Brain and Language, 68:5–15. ping of word forms onto semantics). In the same Forster, K.I. and Azuma, T. 2000 Masked priming for pre- line, Crepaldi et al. (2010) proposed an extension fixed words with bound-stems: Does submit prime per- of Taft’s (1994) sublexical model integrating a mit? Language and Cognitive Processes, 15(4-5):539- lemma level comprised between an orthographic 561. Forster, K.I., & Forster, J.C. (2003). DMDX: A Windows lexicon and the semantic system. However, these display program with millisecond accuracy. Behavioral two models consider the two morphological lev- Research Methods: Instruments & Computers, 35:116- els equivalent, given that they both contain units 124. corresponding to concrete morphemes. One may Frost, R., Deutch, A., & Forster, K.I. 2000. Decomposing morphologically complex words in a nonlinear morphol- nevertheless assume that different locations im- ogy. Journal of Experimental Psychology: Learning, ply different contents: the hybrid model we pro- Memory and Cognition, 26:751-765. pose (Giraudo & Voga, 2014) is based exactly on Giraudo, H. & Grainger, J. 2000. Prime word frequency in this assumption. Within this model, morphologi- masked morphological and orthographic priming. Lan- cal complex words are coded according to two guage and Cognitive Processes, 15:421-444. Giraudo, H., & Grainger, J. 2001. Priming complex words: dimensions, their surface form and their internal Evidence for supralexical representation of morphology. structure. The first level captures the statistical Psychonomic Bulletin & Review, 8(1):127-131. regularities of morphemes translated in terms of Giraudo, H. & Voga, M. (2014). Measuring Morphology: perceptual saliency in the language. At this level, the tip of the iceberg ? A retrospective on 10 years of morphological processing. Carnets de Grammaire, morphologically complex and pseudo-derived 22:136-167. words as well as non-words whose surface struc- Järvikivi, J., & Niemi, J. 2002. Form-based representation ture can be divided into distinct morphemes, are in the mental lexicon: Priming (with) bound stem allo- equally processed. This level is not a morpholog- morphs in Finnish. Brain & Language, 81:412-423. ical level but rather a sub-orthographic level con- Lévy, J., Hagoort, P., & Démonet, J.F. 2014. A neuronal gamma oscillatory signature during morphological unifi- taining “morcemes”. The second level, i.e., the cation in the left occipitotemporal junction. Human morphological level is paradigmatically oriented, Brain Mapping, 35(12):5847-60. it deals with the construction of words according Longtin, M.-C. & Meunier, F. 2005. Morphological decom- to morphological rules (Booij, 2005; Corbin, position in early visual word processing, Journal of Memory and Language, 53(1), 26-41. 1987/1991); it contains “base-lexemes”, units McClelland, J. L., D.E. Rumelhart. 1981. An interactive abstract enough to tolerate orthographic and activation model of context effects in letter perception: phonological variations produced by derivation Part 1. An account of basic findings. Psychological Re- and inflection processes and connected to their view, 88: 375-407. related word forms on the basis of semantic Marslen-Wilson, W. D., Tyler, L. K., Waksler, R., & Older, L. 1994. Morphology and meaning in the English mental transparency. lexicon. Psychological Review, 101:3-33. Pastizzo, M. J. and Feldman, L. B. 2004. Morphological References processing: A comparison between free and bound-stem facilitation. Brain & Language, 90:31-39. Aronoff, M. (1994). Morphology by itself. Cambridge: MIT Rastle, K., Davis, M. H., Marslen-Wilson, W. D., & Tyler, Press. L. K. 2000. Morphological and semantic effects in visual Boudelaa, S., & Marslen-Wilson, W. D. 2001. Morphologi- word recognition: A time-course study. Language and cal units in the Arabic mental lexicon. Cognition, 81:65- Cognitive Processes, 15:507-537. 92. Taft, M. 1994. Interactive activation as a framework for Booij, Geert. 2005. Compounding and derivation: evidence understanding morphological processing. Language & for Construction Morphology. In W.U. Dressler, D. Cognitive Processes, 9:271-294. Kastovsky and F. Rainer (eds) Demarcation in Morphol- Taft, M., & Kougious, P. 2004. The processing of mor- ogy. Amsterdam / Philadelphia: John Benjamins. 109- pheme-like units in monomorphemic words. Brain & 132. Language, 90:9-16. Corbin, D. (1987/1991). Morphologie derivationnelle et Voga, M. & H. Giraudo. 2009. Pseudo-family size influ- structuration du lexique, vol. 2. Tubingen/Villeneuve d ences processing of French inflections: evidence in favor ’Ascq: Max Niemeyer Verlag / Presses Universitaires of a supralexical account in F. Montermini, G. Boyé, J. de Lille. Tseng (eds), Selected Proceedings of the 6th Décem- brettes: Morphology in Bordeaux. Somerville, MA: Cas- cadilla Proceedings Project, 148-155. 23