fpsyg-11-00646 April 6, 2020 Time: 18:17 # 1ORIGINAL RESEARCHpublished: 08 April 2020doi: 10.3389/fpsyg.2020.00646Edited by:Guillaume Thierry,Bangor University, United KingdomReviewed by:Alba Tuninetti,Bilkent University, TurkeyFeier Gao,Indiana University Bloomington,United States*Correspondence:Yu-Fu Chienchien_yc@fudan.edu.cnSpecialty section:This article was submitted toLanguage Sciences,a section of the journalFrontiers in PsychologyReceived: 09 October 2019Accepted: 18 March 2020Published: 08 April 2020Citation:Chien Y-F, Yang X, Fiorentino Rand Sereno JA (2020) The Roleof Surface and Underlying FormsWhen Processing Tonal Alternationsin Mandarin Chinese: A MismatchNegativity Study.Front. Psychol. 11:646.doi: 10.3389/fpsyg.2020.00646The Role of Surface and UnderlyingForms When Processing TonalAlternations in Mandarin Chinese:A Mismatch Negativity StudyYu-Fu Chien1,2* , Xiao Yang3, Robert Fiorentino3 and Joan A. Sereno31 Department of Chinese Language and Literature, Fudan University, Shanghai, China, 2 Department of Modern Languages,DePaul University, Chicago, IL, United States, 3 Department of Linguistics, University of Kansas, Lawrence, KS, United StatesPhonological alternation (sound change depending on the phonological environment)poses challenges to spoken word recognition models. Mandarin Chinese T3 sandhi issuch a phenomenon in which a tone 3 (T3) changes into a tone 2 (T2) when followedby another T3. In a mismatch negativity (MMN) study examining Mandarin Chinese T3sandhi, participants passively listened to either a T2 word [tùu2 je4] /tùu2 je4/, a T3 word[tùu3 je4] /tùu3 je4/, a sandhi word [tùu2 jen3] /tùu3 jen3/, or a mix of T3 and sandhiword standards. The deviant in each condition was a T2 word [tùu2]. Results showedan MMN only in the T2 and T3 conditions but not in the Sandhi or Mix conditions. Allconditions also yielded omission MMNs. This pattern cannot be explained based on thesurface forms of standards and deviants; rather these data suggest an underspecifiedor underlying T3 stored linguistic representation used in spoken word processing.Keywords: event-related potentials, mismatch negativity, tone, tone 3 sandhi, phonological alternation, MandarinChinese, spoken word recognition, lexical representationINTRODUCTIONSpeech is variable. The acoustic characteristics of phonemes, words, and sentences change due toa variety of factors including speaking rate, speaking style, talker, and phonological environment(Cheng, 1966; Goldinger et al., 1991). Recognition of spoken words thus requires mapping thevariable speech input onto stored lexical representations. Understanding how the brain representsthese lexical categories is relevant for theories of speech perception, as theories differ in terms of theabstract nature of the mental representations as well as the role that surface acoustic and underlyinglinguistic representations play in speech processing.The phonological environment affects the acoustic realization of segments in different contexts(e.g., Reetz and Jongman, 2009). In American English, for example, the phoneme /t/ is voicelessaspirated [th] in initial position of a stressed syllable (as in “top”), voiceless unaspirated [t] whenpreceded by a sibilant (as in “stop”), and a voiced alveolar flap [R] between a stressed and anunstressed vowel (as in “matter”). The phonological environment also influences the phoneticrealization of tones in tone languages (Wang and Li, 1967). For instance, a Mandarin Chinese tone3 syllable (a low-dipping tone) changes into a tone 2 syllable (a high-rising tone) when followedFrontiers in Psychology | www.frontiersin.org 1 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 2Chien et al. Investigating Mandarin Tonal Alternations With MMNby another tone 3 syllable, a phenomenon known as MandarinChinese tone 3 sandhi. Given such phonological alternations, akey question arises as to how listeners map variable acoustic-phonetic information onto their linguistic representations storedin the mental lexicon. The current study reports a mismatchnegativity experiment, investigating the neural processing ofMandarin Chinese tone 3 sandhi words, to inform ourunderstanding of the role of the acoustic input and the storedlinguistic representations as listeners map the signal duringspoken word recognition.Psycholinguistic Studies on MandarinChinese Tone 3 SandhiMandarin Chinese (henceforth, Mandarin) is a tone language inwhich the fundamental frequency (F0) of syllables distinguishesword meanings. Mandarin has four tones. Tone 1 is a high-level tone (55); Tone 2 is a high-rising tone (35); Tone 3 is alow-dipping tone (214); and Tone 4 is a high-falling tone (51).In Mandarin, around 70 percent of words are disyllabic (Shei,2014). For disyllabic compounds in Mandarin, words undergoa tonal alternation phenomenon called tone 3 sandhi in whichan initial tone 3 syllable changes into a tone 2 when followedby another tone 3 syllable (Cheng, 1966, 1973; Chao, 1968; Dow,1972; Shih, 1997; Peng, 2000; Duanmu, 2007; Lin, 2007). Acousticanalyses on the realization of this sandhi tone 3, compared tothe canonical tone 2, has shown an overall lower F0 height(Peng, 2000), a lower F0 peak, as well as a later F0 risingonset (Yuan and Chen, 2014) for the sandhi tone 3, suggestingthat the realization of sandhi tone 3 may be influenced by itsunderlying tone 3, and that sandhi tone 3 and canonical tone 2 arenot completely neutralized in production. However, perceptionresults have demonstrated that native speakers often cannotdistinguish sandhi tone 3 from canonical tone 2 (Wang and Li,1967; Peng, 2000), posing a mismatch between the productionand perception of sandhi tone 3.Zhou and Marslen-Wilson (1997) contrasted three alternativeviews concerning how tone sandhi words are represented inthe mental lexicon. The canonical representation view statesthat tone sandhi words are represented as the concatenation ofthe citation forms of their constituent morphemes. Therefore,according to the canonical representation view, the Mandarinsandhi word [pAw2 CjEn3] 保 “insurance”, would be representedas /pAw3 CjEn3/ in Mandarin speakers’ lexicon. The surfacerepresentation view posits that tone sandhi words are storedbased on their surface form (represented as /pAw2 CjEn3/). Theabstract representation view states that since the morphemesundergoing tone sandhi could have either of two surfacetones, the lexical representations of them are underspecified,so that they are compatible and do not conflict with eitheralternation (for segmental underspecification, see Lahiri andMarslen-Wilson, 1991; Lahiri and Reetz, 2002, 2010). Zhou andMarslen-Wilson’s (1997) first auditory priming study on theprocessing of tone sandhi in Mandarin showed slower reactiontimes for tone 3 sandhi targets (e.g., [tshaj2 tChy3] /tshaj3tChy3/ “take an action”) when preceded by disyllabic tone 2primes (e.g., [tshaj2 xua2] /tshaj2 xua2/ “talent”) compared towhen preceded by unrelated control primes (e.g., [thjan1 72]/thjan1 72/ “swan”), and faster reaction times when precededby disyllabic tone 3 primes (e.g., [tshaj3 xoη2] /tshaj3 xoη2/“rainbow”) compared to when preceded by the unrelatedcontrols. Their second experiment showed slower reaction timesfor disyllabic targets with an initial tone 2 (e.g., [tshaj2 phan4]/tshaj2 phan4/ “referee”) when preceded by disyllabic tone 2primes (e.g., [tshaj2 tùhan3] /tshaj2 tùhan3/ “asset”), tone 3primes (e.g., [tshaj3 na4] /tshaj3 na4/ “adopt”), and sandhi primes(e.g., [tshaj2 faη3] /tshaj3 faη3/ “interview”) compared to whenpreceded by disyllabic unrelated control primes (e.g., [y4 ljao4]/y4 ljao4/ “predict”). In addition, the inhibitory effect was evenstronger in the tone 3 prime condition. Zhou and Marslen-Wilson claimed that the results from these two experimentswere not consistent with the abstract representation view, yetthey could not distinguish between the canonical and surfacerepresentation views.To further investigate how Mandarin tone 3 sandhi words areprocessed, Chien et al. (2016) conducted an auditory-auditorypriming lexical decision experiment examining the role of surfaceand underlying forms. Each disyllabic Mandarin tone 3 sandhitarget (e.g., /fu3 tao3/ , “to counsel”) was preceded by oneof three corresponding monosyllabic primes, namely, a tone 2prime (e.g., /fu2/ 服, “to assist”), a tone 3 prime (e.g., /fu3/ 輔 , “toguide”), or a control prime (e.g., /fu1/ 敷, “to put on”). Resultsshowed that tone 3 primes (underlying form overlap with thefirst syllable of the tone 3 sandhi targets) facilitated participants’lexical decision responses, while tone 2 primes (surface formoverlap with the first syllable of the tone 3 sandhi targets) didnot show facilitation relative to the unrelated control primes.Moreover, these priming effects occurred regardless of the targets’word frequency. These data show that tone 3 primes (e.g.,/fu3/ 輔 , “to guide”) facilitated reaction times to all tone sandhitargets (e.g., /fu3 tao3/ , “to counsel”) due to morpheme levelactivation between primes and the underlying form of the firstsyllable of the targets, while tone 2 primes (e.g., /fu2/ 服, “toassist”), although matching targets on the surface, did not showpriming effects, indicating that surface matching did not result infaster sandhi word recognition. Chien et al. (2016) concluded thatMandarin tone 3 sandhi words are represented as /tone 3 tone 3/in listeners’ mental lexicon.Given that previous auditory priming studies haven’t yetreached consensus regarding how Mandarin tone 3 sandhiwords are represented in the mental lexicon, the currentstudy further investigated tone sandhi processing by examiningcortical responses utilizing a mismatch paradigm, which hasbeen argued to reflect pre-attentive processing of auditorylinguistic representations independent of any behavioral task(Kujala et al., 2007; Näätänen et al., 2007). A number ofprevious studies have shown that the mismatch paradigm can beused to examine higher-level processes outside of the focus ofattention (Phillips et al., 2000; Pulvermüller and Shtyrov, 2006;Gu et al., 2012). Additionally, using the mismatch paradigmalso allowed us to isolate the effect elicited by the first tone oftone 3 sandhi words, so that we could examine how Mandarinlisteners process the initial rising tone, which could potentiallybe the first syllable of a tone 3 sandhi word, during spokenword recognition.Frontiers in Psychology | www.frontiersin.org 2 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 3Chien et al. Investigating Mandarin Tonal Alternations With MMNNeurophysiological Studies on LexicalTone ProcessingIn the past few decades, event-related brain potentials (ERPs)have been used to investigate the neural basis of perception withhigh temporal resolution. Although ERPs do not transparentlyreflect where neural processes occur, they do provide amillisecond by millisecond record of brain activity, yielding anumber of ERP components linked to perception. One ERPcomponent that the speech perception literature has widely usedin order to investigate the representation and processing ofspoken language is the mismatch negativity (MMN). Previousstudies have shown that the mismatch negativity (or mismatchfield, MMF, in MEG) reflects deviations from previouslypresented events or regularities (Näätänen et al., 1997; Phillipset al., 2000; Kazanina et al., 2006; Kujala et al., 2007; Wang et al.,2012). MMN is traditionally recorded using an oddball paradigmin which a sequence of similar sounds (“standards”) is presentedand rare sounds (“deviants”) are introduced occasionally. Thesimilar sounds form a memory trace in the auditory system towhich deviant sounds contrast, yielding MMN effects around100–250 ms after the onset of deviant sound presentation(Näätänen, 1992; Näätänen and Winkler, 1999).The mismatch response has also been used to examine theprocessing of tonal information in tone languages. Xi et al. (2010)used MMN to investigate categorical perception of Mandarinlexical tone 2 and tone 4 with the syllable structure /pa/. A 10-interval lexical tone continuum from tone 2 (stimulus 1) to tone4 (stimulus 11) was created, producing a total of 11 stimuli.Based on the results of a previously conducted AX discriminationtask, an across-category (stimulus 3 and stimulus 7) pair wasimplemented in an oddball experiment in which stimulus 3 wasthe deviant and stimulus 7 was the standard. In another oddballexperiment, a within-category (stimulus 11 and stimulus 7) pairwas selected, with stimulus 11 used as the deviant and stimulus 7as the standard. Their results showed that two physically differenttones across two distinct tonal categories elicited greater MMNamplitude than did two physically different tones within the sametonal category. Interestingly, physically different tones acrosstonal categories yielded a stronger MMN amplitude than didphysically different tones within the same tonal category in theleft hemisphere but not in the right hemisphere. These resultssuggest that the left-hemisphere may be more involved withthe long-term phonemic processing of lexical tones while theright hemisphere may play a more important role for acousticprocessing (Ren et al., 2009; Xi et al., 2010).Research has also demonstrated that speakers can extractregularities from complex acoustic signals at the tonal level.Wang et al. (2012) conducted a mismatch negativity study toinvestigate whether Mandarin speakers possess a mechanism thatcan help them extract abstract linguistic rules from complexacoustic information at a pre-attentive stage. Their experimentwas conducted using an oddball paradigm in which 90 percentof the monosyllabic stimuli were level tones (standards), carriedby different Mandarin vowels ([i, e, a, u]), with various F0height (10 levels from 78 Hz to 150 Hz) as well as intensity(3 intensity levels). The remaining 10 percent of the monosyllabicstimuli were either rising or falling tones (deviants), carriedby Mandarin vowels ([i, e, a, u]), with three F0 levels (i.e., anidentical F0 contour beginning at three different F0 heights)and intensity (same as the three intensities used for thestandards). Results showed a mismatch negativity, indicating thatthe effect emerged when the rule formed by level tones wasviolated. Wang et al. (2012) concluded that listeners can extractlinguistically meaningful rules from complex auditory signalsduring speech processing.Recently, MMN studies have begun to examinerepresentations involving phonological alternations. Li andChen (2015) conducted a mismatch negativity study to examineMandarin lexical tone representations involving phonologicaltonal alternations. An oddball experiment was conducted inwhich monosyllables [ma] with tone 1, tone 2 and tone 3 wereset as stimuli. Four conditions were generated: T1 standard/T3deviant, T3 standard/T1 deviant, T2 standard/T3 deviant, andT3 standard/T2 deviant. Results showed that all four conditionselicited MMN effects. However, the patterns of MMN effects weredifferent. T1 standard/T3 deviant and T3 standard/T1 deviantyielded comparable MMN effects. However, T2 standard/T3deviant and T3 standard/T2 deviant showed asymmetrical MMNeffects. When T3 was the standard and T2 was the deviant, theMMN effect was significantly reduced compared to when T2was the standard and T3 was the deviant. Moreover, the MMNelicited in the T3 standard/T2 deviant condition demonstrated aright hemispheric distribution, with all the other three conditionsshowing a left hemispheric MMN distribution. Li and Chen(2015) claimed that the reduced, right lateralized MMN effect inthe T3 standard/T2 deviant condition may be due to the fact thatwhen participants heard a tone 3 syllable, both tone 3 and tone3V (the sandhi tone, which is acoustically similar to tone 2; Yuanand Chen, 2014) representations were activated, leading to thereduced MMN effect. Based on these results, Li and Chen (2015)proposed that tone 3 is represented as both /tone 3/ and /tone3V/ in the mental lexicon.Similar to Li and Chen (2015), Politzer-Ahles et al. (2016)recently investigated the nature of Mandarin tone 3 by looking atthe MMN response to tone 3 in contrast with other tones. In threeexperiments, participants were presented with conditions wheremonosyllabic tone 3 words and monosyllabic words with othertones (tone 1, tone 2, tone 4) alternatively served as standardsand deviants. Across the experiments, they found the predictedasymmetrical MMN response elicited by conditions with tone3, while no asymmetry was observed in conditions with othertones. Mandarin tone 3 has several properties that underspecifiedsegments also have, such as surface alternation, low pitch register,and use as attentional cues (Kuo et al., 2007; Hsu et al., 2015;Politzer-Ahles et al., 2016). Politzer-Ahles et al. (2016) arguedthat the putatively underspecified nature of tone 3 weakens thecontrast between standards and deviants. Thus, when tone 3contrasted with other tones, asymmetrical MMNs were observed.Politzer-Ahles and colleagues suggested that the findings areconsistent with the underspecification accounts of Mandarin tone3, in that listeners do not seem to assign a clear tonal category totone 3 words. Although Li and Chen (2015) as well as Politzer-Ahles et al. (2016) provided evidence regarding Mandarin tone3 representations, these patterns of data were observed withFrontiers in Psychology | www.frontiersin.org 3 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 4Chien et al. Investigating Mandarin Tonal Alternations With MMNisolated syllables. A further question is how Mandarin nativespeakers process disyllabic tone sandhi words.Zhang et al. (2015) began to address this question in aspeech production task, using Mandarin disyllables to investigatethe phonological encoding of Mandarin tone 3 sandhi words.Participants were asked to concatenate two auditorily presentedsyllables together and covertly pronounce them as a whole chunk(e.g., T2 + T3 and T3 + T3). Participants were then askedto overtly pronounce these disyllables, as a unit, after a silentinterval (varying between 1000 and 1600 ms). Zhang et al.’s (2015)goal was to test whether Mandarin tone 3 sandhi words involvecomputational conversion of the first tone 3 syllable into a tone2 syllable, or whether tone 3 sandhi words undergo a lexicalmechanism in which a tone 3 sandhi word’s surface form isdirectly accessed.ERP recordings acquired during covert production showed asignificantly greater P200 (P2) amplitude at midline electrodeselicited by tone 3 syllables which are preceded by a tone 3syllable than those preceded by a tone 2 syllable. The strongerP2 amplitude yielded by the second tone 3 syllable of Mandarintone 3 sandhi words indicated that Mandarin speakers did expendadditional effort, as P2 has been shown to be related to higher-order perceptual processing, including attention, repetitioneffects and task difficulties (Dunn et al., 1998; Kim et al., 2008;Landi et al., 2012). Moreover, no significant P2 difference wasobserved between real tone 3 sandhi words and tone 3 sandhinon-words, suggesting that Mandarin speakers process tone 3sandhi words computationally, applying the tone 3 sandhi ruleeven to pseudowords.However, the greater P2 effect elicited by the tone 3 syllablespreceded by other tone 3 syllables could also be because itwas easier for the participants to produce words (T2 + T3)that matched the two syllables they heard (T2 + T3) than toproduce a word that did not match the two syllables they heard(T3 + T3). Since tone 3 syllables were always presented as acanonical falling-rising tone [214] rather than as their sandhiform [35], participants had to transform the canonical falling-rising tone 3 of the first syllable of a disyllabic sequence into a tone2 in order to produce the correct output. Additionally, the factthat participants were provided the canonical tone 3 form mightencourage them to focus on the underlying form (falling-risingF0 contour) of tone 3 in these sandhi syllables, which might thenrequire participants to use extra effort to convert what they heard(a falling-rising F0 contour) to what they needed to produce(a rising F0 contour).Taken together, previous neurophysiological studies ontone sandhi provide some differing evidence regarding therepresentation of Mandarin tones and how Mandarin tone 3sandhi words are processed. While Li and Chen (2015) proposedthat Mandarin tone 3 monosyllables should be representedas both tone 3 and tone 3V in the mental lexicon, Politzer-Ahles et al.’s (2016) data indicated that tone 3 may bephonologically underspecified, although neither study directlyexamined disyllabic tone 3 sandhi words.The current study directly examines the processing ofdisyllabic tone sandhi words. In word recognition, Mandarinnative speakers must process the tone of the first syllable ofMandarin tone 3 sandhi words, which is acoustically similar toa tone 2. Whether Mandarin speakers process the initial sandhitone 3 of sandhi words as a tone 2 or a tone 3, or whether speakersassign a special status to the initial sandhi tone 3 and wait forthe context (i.e., either a tone 3 syllable or not) is unclear. Toaddress this issue, we conducted a mismatch negativity study totest the role of surface and underlying forms of Mandarin tone 3sandhi words during spoken word processing. Investigating theprocessing of Mandarin tone 3 sandhi words in this way allowsone to examine processing independent of any overt behavioralresponse, such as lexical decision.Current StudyAn oddball paradigm was used to compare mismatch responsesacross four conditions: Tone 2, Tone 3, Sandhi, and Mix. In allconditions, the standards were disyllables and the deviant wasa tone 2 monosyllable (see Figure 1). We utilized a varying-standard design in which the standards within each conditionwere not acoustically identical to each other, as our main interestin the current study is to examine the contexts in which thebrain is able to extract a shared tonal representation across thephysically different standards within each condition, rather thanto test whether the brain can distinguish single tokens that eithermatch or mismatch in tone.The Tone 2 condition included disyllabic words with an initialtone 2 syllable followed by a tone 4 syllable ([tone 2 tone 4]/tone 2 tone 4/). The lack of a tone 3 sandhi environment in thiscondition ensured that the initial tone 2 of the standards had asurface and an underlying tone 2 form. In this condition, there isno mismatch between the standards and deviants in underlyingtone category or surface tone category (see Figure 1), as both theinitial syllable of standards and deviants are tone 2. However, asthis study utilizes a design in which multiple physically differenttokens of the standards are used, this condition is intendedto assess whether the physical difference between the set ofstandards and deviants is sufficient to yield an MMN (Winkleret al., 1999; Kasai et al., 2001; Xi et al., 2010), or whether MMNwill be absent due to the lack of many-to-one ratio at the levelof tone category.The Tone 3 condition included disyllabic standards with aninitial tone 3 syllable followed by a tone 4 syllable ([tone 3 tone 4]/tone 3 tone 4/). Since none of the standards in this conditionwere sandhi words, the first syllable should be interpreted astone 3 in both the surface and underlying form. For the Tone3 condition, we expected to observe an MMN effect due to themismatch between the first syllable of standards and the deviantin both underlying and surface forms (see Figure 1). Althoughprevious studies using monosyllables had found a reduced MMNeffect when a tone 3 was the standard and a tone 2 was the deviant(Li and Chen, 2015; Politzer-Ahles et al., 2016), we predicted thatin the current study, the disyllabic tone 3 standards might reducethe possibility for Mandarin-speaking participants to activatea tone 3V or a phonologically underspecified representationbecause the initial tone 3 syllable is never followed by anothertone 3 syllable (i.e., it is never in a sandhi context). The presentdisyllabic context would encourage participants to assign a clearFrontiers in Psychology | www.frontiersin.org 4 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 5Chien et al. Investigating Mandarin Tonal Alternations With MMNFIGURE 1 | Stimuli in the Tone 2 condition, Tone 3 condition, Sandhi condition, and Mix condition.tone 3 interpretation, potentially leading to a strong MMN effectin this condition.The Sandhi condition included disyllabic tone 3 sandhi words([tone 2 tone 3] /tone 3 tone 3/), with a tone 2 on the surfaceand a tone 3 underlyingly. For the Sandhi condition, whether ornot an MMN effect would emerge depends on how participantsprocess Mandarin tone 3 sandhi words. If participants processMandarin tone 3 sandhi words based on the underlying tone3 and the deviant is interpreted as a canonical tone 2 syllablein isolation, an MMN effect should be elicited since the tone2 deviant would violate the regularity formed by repetitivelyoccurring underlying tone 3 of the first syllable of standards. Ifparticipants focus on the surface tone 2, then no MMN effectwould be expected due to the lack of tonal mismatch withthe tone 2 deviant at the level of surface tone category (seeFigure 1). If both the underlying tone 3 and the surface tone2 are activated, then an MMN effect may still be observed,but with a weaker amplitude compared to that yielded in theTone 3 condition.In a fourth Mix condition, standards consisted of disyllabictone 3 sandhi words ([tone 2 tone 3] /tone 3 tone 3/) mixedwith disyllabic words starting with a tone 3 syllable ([tone 3tone 4] /tone 3 tone 4/). In the Mix condition, whether or notan MMN effect would emerge depends on whether Mandarin-speaking participants could extract the underlying tone of the firstsyllable of standards despite the differences in surface tone. If theycould, then an MMN effect is predicted because the tone 2 deviantwould violate the regularity formed by frequently occurringunderlying tone 3 of the first syllable of the standard, similarto Wang et al. (2012) where Mandarin-speaking participantsextracted a level tone for standards and two contour tones fordeviants despite their varying acoustic information, F0 heights,and intensity. If they could not extract the underlying tone of thefirst syllable of standards, then no MMN effect would be elicited(see Figure 1).For all four conditions, the initial syllable of the standardswas matched for segmental information, and the same tone2 monosyllabic deviant, which also matched the segmentalinformation of the first syllable of the standards, was used.These four conditions allowed us to directly examine howMandarin tone 3 sandhi words are processed by comparingthe ERPs elicited by the deviant and those yielded by thefirst syllable of standards in the same condition. In addition,utilizing disyllabic standards in this MMN study allowed forexplicitly testing the perception of tone 3 sandhi at a pre-attentivestage, and teasing apart the role of surface and underlyingforms. Specifically, we were able to determine whether Mandarinspeakers processed the first syllable of sandhi words as a tone2 or a tone 3 by examining whether MMN is elicited bythe deviant that contrasts with the first syllable of standardsin each condition.Regarding the three contrasting views on the representation oftone sandhi words (Zhou and Marslen-Wilson, 1997), the surfacerepresentation view would predict that the Sandhi condition andTone 2 condition show similar MMN effects since the surfacetonal information is matched across these two conditions, whilethe Sandhi condition compared to the Tone 3 condition and/or the Mix condition would yield different MMN patterns,given the different surface tonal information between the Tone 2condition and these two conditions. The canonical or underlyingrepresentation view would predict similar MMN patterns forthe Tone 3 condition, the Sandhi condition, and the Mixcondition, since all these conditions possess an underlying tone3 as the initial syllable of standards while the MMN patternbetween the Sandhi condition and Tone 2 condition would bepredicted to be different since these two conditions have differentunderlying tones as the initial syllable of standards. The abstractrepresentation view would predict dissimilar MMN patternsamong the Sandhi condition and all the other conditions giventhat it is the initial tone of the sandhi words that repeatedlyFrontiers in Psychology | www.frontiersin.org 5 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 6Chien et al. Investigating Mandarin Tonal Alternations With MMNalternates with tone 2, which would license an underspecifiedenvironment while the initial tone of standards in the otherconditions does not consistently undergo tone sandhi.Finally, the current design also allowed us to examine within-condition comparisons between the second syllable of standardsand the omission position (following the monosyllabic deviant).Given that our standards were disyllabic and our deviants weremonosyllables, we were able to utilize the omission MMNto directly investigate whether the deviants were processed asmonosyllables or as the first syllable of an expected disyllabicword. In our study, we would thus expect to observe anomission MMN effect in all conditions given that the deviantis a monosyllable, which would violate the prediction formedby the disyllabic standards (Raij et al., 1997; Yabe et al., 1997,1998; Janata, 2001). Additionally, we may also predict differentomission MMN amplitudes across conditions given that thesecond syllable of standards used in the four conditions varies,which might lead to different predictability (Bendixen et al.,2009, 2012, 2014). Finding these omission effects would provideevidence that Mandarin-speaking participants were processingdisyllabic standards as disyllabic stimuli and monosyllabicdeviants as the first syllable of disyllabic stimuli.MATERIALS AND METHODSParticipantsThirty-two right-handed native Mandarin speakers wererecruited (14 males and 18 females), with ages from 18 to35 years old. All participants were students from the Universityof Kansas with no reported language disability or hearingimpairment. This research was reviewed and approved by theHuman Subjects Committee at the University of Kansas. Allparticipants were asked to provide informed consent beforethe EEG experiment and were paid 10 dollars per hour fortheir participation.StimuliStimuli were produced by a male native Beijing Mandarinspeaker and recorded in an anechoic chamber with a cardioidmicrophone (Electrovoice, model N/D767a) and a digital solid-state recorder (Marantz, model PMD671), using a samplingrate of 22,050 Hz at the University of Kansas. Stimuli used inthis experiment were ten disyllabic Mandarin words and onemonosyllabic Mandarin word. For the ten disyllabic words, onewas a tone 2 + tone 4 word; five were tone 3 sandhi words, andfour were tone 3 + tone 4 words. The monosyllable was a tone2 word. All the disyllabic stimuli were normalized to 630 ms induration (first syllable 280 ms+ second syllable 350 ms), and themonosyllabic word was adjusted to 280 ms in duration. Intensityof all stimuli was normalized to 65 dB. All the manipulations wereconducted using Praat (Boersma and Weenink, 2015).DesignStimuli for each of the four test conditions were presentedusing the variable-standards structure in an oddball paradigm(see Figure 1). In each condition, 87.5 percent of the stimuliwere standards, while 12.5 percent of the stimuli were deviants.In the Tone 2 condition, 7 physically distinct instances of[tùu2 je4] /tùu2 je4/ “bamboo leaf” were used as standards.In the Tone 3 condition, 7 physically different tokens of[tùu3 je4] /tùu3 je4/ “cover page” were selected as standards.In the Sandhi condition, 7 physically different instances of[tùu2 jen3] /tùu3 jen3/ “to star” were set as standards. In theMix condition, four tone 3 sandhi words (T3 + T3 tonalsequence) ([tùu2 ta3] /tùu3 ta3/ “to feature”; [tùu2 pi3] /tùu3pi3/ “editorial writer”; [tùu2 ùCn3] /tùu3 ùCn3/ “chief umpire”;[tùu2 kwan3] /tùu3 kwan3/ “director”) and three disyllabicwords with a T3 + T4 tonal sequence ([tùu3 üCn4] /tùu3üCn4/ “manager”; [tùu3 pan4] /tùu3 pan4/ “to host”; [tùu3li4] /tùu3 li4/ “main force”) were used as standards. Forall four conditions, monosyllable [tùu2] /tùu2/ “bamboo” wasused as the deviant.Critically, in the Tone 2 condition, the first syllable of thestandard [tùu2 je4] /tùu2 je4/ and the deviant [tùu2] /tùu2/ werematched in both the surface and underlying forms, as shownin Figure 1. In the Tone 3 condition, the first syllable of thestandard [tùu3 je4] /tùu3 je4/ did not match the tone in thedeviant [tùu2] /tùu2/ in either surface or underlying form. In theSandhi condition, the first syllable of the standard [tùu2 jen3]/tùu3 jen3/ shared only the surface form with the deviant [tùu2]/tùu2/ but mismatched in the underlying form (tone 3). In theMix condition, there was not a many-to-one ratio with respect tothe first syllable of the standards and the deviant on the surface(either tone 2 or tone 3). However, there was a many-to-oneratio between the first syllable of the standards and the deviantunderlyingly (all tone 3).To equate the Tone 2 and Sandhi conditions, tonal contoursfrom four canonical tone 2 (four of the seven first syllables ofthe [tùu2 je4] /tùu2 je4/ standards) in the Tone 2 condition weresuperimposed upon the first syllable of 4 standards ([tùu2 jen3]/tùu3 jen3/) in the Sandhi condition. Likewise, tonal contoursfrom three sandhi tones (three of the seven first syllables ofthe [tùu2 jen3] /tùu3 jen3/ standards) in the Sandhi conditionwere superimposed on the first syllable of the other 3 standards([tùu2 je4] /tùu2 je4/) in the Tone 2 condition. Therefore, thefirst syllable of standards in the Tone 2 and Sandhi conditionshad identical tonal contours, with some from underlying tone 2stimuli and some from tone 3 sandhi stimuli.For the Tone 3 condition, tonal contours of the standardswere not manipulated. In the Mix condition, two of the fourstandards were the same as the first syllable of the two [tùu2 je4]/tùu2 je4/ Tone 2 condition standards; two of the four were thesame as the first syllable of the two [tùu2 jen3] /tùu3 jen3/ Sandhicondition standards. By doing so, the first syllable of standardsin the Mix condition had similar tone 2 contours as the firstsyllable of standards in the Tone 2 and Sandhi conditions. Forthe other three T3 + T4 words, no manipulation was conducted.The second syllables were not manipulated either, with four tone3 syllables and three tone 4 syllables varying in terms of thesegmental information.The same monosyllabic [tùu2] was used as the deviant acrossthe four conditions. The tonal contour of the deviant wasnot manipulated.Frontiers in Psychology | www.frontiersin.org 6 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 7Chien et al. Investigating Mandarin Tonal Alternations With MMNFundamental frequency (F0) was measured for the sevenstandard [tùu2] tokens and the deviant [tùu2] in the Tone2 and Sandhi conditions. The fundamental frequency of theseven standard [tùu2] tokens in the Tone 2 and Sandhiconditions was, on average, 10.28 Hz higher than that ofthe deviant [tùu2] in these two conditions. The 1F0 value(i.e., the change in fundamental frequency from onset of thetone to the lowest pitch value) of the average F0 contourof the standard [tùu2] tokens was 9.93 Hz, while the 1F0value of the deviant [tùu2] was 15.23 Hz. Similar measureswere also collected for the seven standard [tùu3] tokens inthe Tone 3 condition. The F0 of the [tùu3] tokens wasrealized as a low-falling tone, which was distinct from thatof the deviant [tùu2]. Acoustic details of the stimuli areshown in Appendix A.The four conditions were separately presented in 8 blocks,with each condition being split into two blocks. The order ofthe four conditions was counterbalanced across participants. Atthe beginning of each block, 14 standards were presented first tofamiliarize the participants with the standards. Then another 400trials were presented in a pseudo-randomized order, with everytwo deviants separated by 2 to 9 standards. There was a 600 mspause between stimuli. In total, participants heard 828 stimuli ineach condition, with 728 standards and 100 deviants. A self-pacedbreak was offered after each block.ProcedureParticipants were invited to the Neurolinguistics & LanguageProcessing Laboratory at The University of Kansas. They filledout a consent form, the Edinburgh Handedness Inventory(Oldfield, 1971), as well as a language background questionnaire.The participants were then fitted with an electrode cap beforethey were seated in front of a computer screen. Auditory stimuliwere presented using insert earphones (Etymotic Research, Inc.)while participants watched a silent movie with Chinese subtitleson the computer screen in front of them (Chandrasekaran et al.,2007, 2009; Xi et al., 2010; Gu et al., 2012; Wang et al., 2012;Politzer-Ahles et al., 2016). Participants were instructed to ignorethe auditory stimuli while watching the movie. The total durationof the experiment was around 2 h.EEG RecordingsElectroencephalogram (EEG) was recorded at a sampling rateof 1 kHz using an electrode cap (Electro-cap International,Inc.) from 32 sintered, Ag/AgCl scalp electrodes, arranged in amodified 10–20 layout. Electrode sites were FPz, Fz, FCz, Cz,CPz, Pz, Oz, FP1/2, F7/8, F3/4, FT7/8, FC3/4, T3/4, C3/4, TP7/8,CP3/4, T5/6, P3/4, and O1/2, with electrode AFz serving asthe ground electrode and the left mastoid as online reference.The recording was amplified by using a Neuroscan Synamps2amplifier system (Compumedics Neuroscan, Inc.). Additionalelectrodes were placed above and below each eye, and on theleft and right outer canthi of each eye to detect blinks and eyemovements. Impedances were kept below 5 kOhms. Data werecontinuously recorded in AC mode using an online high-passfilter of 0.1 Hz and a low-pass filter of 200 Hz.Data AnalysisEvent-related brain potentials analyses were time-locked at theonset of the first syllable of standards (0 ms post onset of standardpresentation), the onset of deviants (0 ms post onset of deviantpresentation), the onset of the second syllable of standards(280 ms post onset of standard presentation), and the onset ofomission positions (280 ms post onset of deviant presentation).Continuous EEG data were re-referenced offline to the mean ofthe left and right mastoids using Neuroscan Edit (CompumedicsNeuroscan, Inc.). The subsequent analyses were carried outusing EEGLAB (Delorme and Makeig, 2004) and MATLAB (TheMathWorks, Inc.). Bad channels were interpolated; no morethan 2 bad channels were found per participant. The continuousdata were then segmented into −500 ms to 500 ms epochsrelative to all the triggers and demeaned using the mean ofthe whole epoch (as recommended by Groppe et al., 2009).The data were then decomposed into independent componentsusing ICA (Makeig et al., 1996). For each participant one to fiveindependent components that are typical for eye movements,blinks, or muscle artifacts were identified by visual inspection andpruned from the data.Trials were then divided into smaller epochs based on Status(standard, deviant), Condition (Tone 2, Tone 3, Sandhi, Mix) andPosition (first syllable/onset, second syllable/omission) resultingin 16 different combinations. The epoch for the first syllable ofstandards and deviants at the onset position in all four conditionsspanned a −100 ms to 400 ms time window with respect tothe onset of the first syllable (0 ms). The epoch for the secondsyllable of standards and deviants at the omission position inall four conditions spanned a −380 ms to 400 ms time windowwith respect to the onset of the second syllable (0 ms). Thefirst-syllable-of-standard epochs and deviant epochs were thenbaseline-corrected using the −100 ms to 0 ms interval precedingthe first syllable, while the second-syllable-of-standard epochsand omission epochs were baseline-corrected using the −380 msto −280 ms interval (this part of the epoch was used for baselinecorrection as it is the 100-ms interval preceding the first syllable).Epochs with amplitudes exceeding ± 100 µV at any channelwere automatically rejected, resulting in the elimination of twoparticipants’ data due to excessive artifacts (i.e., more than 20%of deviant trials at both onset and omission positions wereexcluded). For the remaining 30 participants, less than 2% of allthe trials were excluded. Epochs were then low-pass filtered at30 Hz. The remaining data were carried forward for statisticalanalysis. Averaged ERP waveforms were generated for the first-syllable-of-standard epochs, deviant epochs, second-syllable-of-standard epochs, and omission epochs within each condition.The traditional MMN method was employed by comparingthe ERPs elicited by the deviant with those by the standards in thesame condition, since the main focus of this study is to compareresponses to the same tone 2 deviant following different standardcontexts, rather than having the same tone 2 stimuli as standardsand deviants in different conditions. Moreover, note that themonosyllabic deviant [tùu2] would be ambiguous if it was treatedas the standard in a reverse block used for the same-stimulusMMN method (i.e., comparing the ERPs to the deviant withFrontiers in Psychology | www.frontiersin.org 7 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 8Chien et al. Investigating Mandarin Tonal Alternations With MMNthose to the same stimulus when it is presented as the standardin another block). Such ambiguity was unlikely to happen whenthe surface form [tùu2] served as the deviant in the currentstudy, since the participants would interpret the deviant [tùu2]based on their predictions of the upcoming stimuli, and thesepredictions were made based on the standard of each condition.The same-stimulus method would not allow us to compare thesame status of [tùu2] within each condition. Moreover, previousresearch showed that the traditional and same-stimulus methodsproduced similar MMN patterns yielded by frequency (Peteret al., 2010), and we also compared responses to the physicallyidentical deviant [tùu2] across conditions to ensure that theresults were compatible with those generated by comparing theMMN effects across conditions using the traditional method.The midline electrode Fz was selected for mean ERP amplitudeanalyses since previous studies have shown that mismatch effectshave a fronto-central distribution (Näätänen et al., 2007; Renet al., 2009; Gu et al., 2012). Additionally, we analyzed themean ERP amplitudes on 9 representative electrodes (F3, Fz,F4, C3, Cz, C4, P3, Pz, and P4) in order to characterize thetopographic distribution of the MMN effects we may observe.For each participant, mean ERP amplitude was calculated byaveraging the ERP voltages in the 100–280 ms time window[comparable to that used in Li and Chen (2015) and Politzer-Ahles et al. (2016)] across epochs by Status (standard vs. deviant),Condition (Tone 2, Tone 3, Sandhi, and Mix) and Position(onset vs. omission).RESULTSStatistical analyses were conducted on the mean ERP amplitudesobtained from the MMN experiment. The MMN resultsgenerated by comparing the mean ERP amplitudes at the deviantposition and those at the first syllable of standards will bepresented first. Then the MMN results produced by comparingthe mean ERP amplitudes at the omission position and those atthe second syllable of standards will be introduced.Deviant Onset PositionFz AnalysisMean ERP amplitudes were analyzed via a series of linear mixed-effects models using the lme4 package (Bates et al., 2015) in R,with p-values calculated by the lmerTest package (Kuznetsovaet al., 2017). Status (First Syllable of Standard, Deviant) andCondition (Tone 2, Tone 3, Sandhi, Mix) were entered ascategorical fixed factors. For Status, Deviant was set as thebaseline to which First Syllable of Standard was compared inorder to derive MMN, while for Condition, Tone 2 was coded asthe baseline. Participant was included as the random intercept,while Status and Condition were entered for by-Participantrandom slopes. A series of likelihood ratio tests were performedto compare three by-Participant random slopes models in orderto evaluate effects of Status, Condition, and Status × Condition.Results showed that the model containing Status, Condition,as well as Status × Condition was significantly better than theone consisting of Status and Condition (χ2 = 24.311, df = 3,p < 0.001), thus, determined to be the best model. GivenFIGURE 2 | The four panels depict mean ERP waveforms for electrode Fzaveraged across 30 participants in the 0–680 ms time window (using the–100–0 ms time window as the baseline) for the four experimental MMNconditions (Tone 2 condition is the top panel, Tone 3 condition is the secondpanel, Sandhi condition is the third panel, Mix condition is the bottom panel).The solid vertical line delineates the onset of the first syllable of the standards(blue line) and the onset of the deviants (red line). The dotted line delineatesthe end of the first syllable of standards which is also the onset of the secondsyllable of standards (blue line) and the offset of the deviants which is also theonset of omissions (red line).that Status (First Syllable of Standard, Deviant) is the crucialvariable to index MMN effects, only the results relevant to Statusare reported below.As for the best model, linear-mixed effects analysisdemonstrated that ERPs elicited by the deviant were significantlymore negative than those produced by the first syllable ofstandards, indicating an MMN effect (β = 0.744, SE = 0.187,t = 3.987, p < 0.001) (see Figure 2). Moreover, a significantinteraction between Status and Condition was also obtainedFrontiers in Psychology | www.frontiersin.org 8 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 9Chien et al. Investigating Mandarin Tonal Alternations With MMN(df = 3, F = 8.696, p < 0.001). This interaction effect resulted fromthe significantly stronger MMN effect in the Tone 2 conditionthan that in the Sandhi condition (β = −0.903, SE = 0.255,t = −3.543, p < 0.001), and than that in the Mix condition(β = −0.812, SE = 0.255, t = −3.1888, p = 0.002). However, theMMN effect in the Tone 2 condition was not different from theMMN effect in the Tone 3 condition (β = 0.114, SE = 0.255,t = 0.447, p = 0.656) (see Figure 3).Given that we observed a main effect of Status withall conditions included, we conducted four follow-up linear-mixed effects analyses between First Syllable of Standard andDeviant to evaluate whether MMN effects emerged in eachindividual condition (see Figure 2). For each analysis, meanERP amplitude was regarded as a dependent variable, whileStatus (First Syllable of Standard, Deviant) was treated as acategorical fixed factor, with Deviant coded as the baselinein order to derive MMN. In addition, Participant was setas a random intercept, while Status was coded for by-participant random slopes.Results showed that for Tone 2 condition, ERPs elicited by thedeviants were significantly more negative than those yielded bythe first syllable of standards (β = 0.744, SE = 0.226, t = 3.290,p = 0.003), indicating an MMN effect. For Tone 3 condition,ERPs yielded by the deviants were significantly more negativethan those produced by the first syllable of standards (β = 0.858,SE = 0.153, t = 5.591, p < 0.001), also confirming an MMNeffect. However, with regard to Sandhi condition, ERPs elicitedby the deviants were not significantly different from those yieldedby the first syllable of standards (β = −0.159, SE = 0.192,t = −0.829, p = 0.414), nor were ERPs yielded by the deviantsin the Mix condition compared to those produced by the firstsyllable of standards in the Mix condition (β =−0.068, SE = 0.190,t =−0.360, p = 0.721) (see Figure 2).Since the deviant was physically identical across all the fourconditions, we could directly compare ERPs elicited by thedeviant across conditions. We ran an additional mixed-effectsanalysis on the mean ERP amplitude of the deviant for Fz,with Condition as a fixed factor (i.e., Tone 2 condition was setas the baseline condition) and Participant as a random effect.The results patterned the same as those reported above (seeFigure 3), with the mean ERP amplitude of the deviant in theTone 2 condition being significantly more negative than that inthe Sandhi (β = 1.014, SE = 0.250, t = 4.065, p < 0.001) and Mixconditions (β = 0.954, SE = 0.250, t = 3.823, p < 0.001).Multi-Electrode AnalysisFor the multi-electrode analyses, 9 representative electrodes wereselected: F3, Fz, F4, C3, Cz, C4, P3, Pz, and P4. The electrodeswere coded for Laterality (Left, Midline, Right) and Anteriority(Frontal, Central, Posterior): F3 as Left Frontal, Fz as MidlineFrontal, F4 as Right Frontal, C3 as Left Central, Cz as MidlineCentral, C4 as Right Central, P3 as Left Posterior, Pz as MidlinePosterior, and P4 as Right Posterior. A baseline level in eachof these two variables was chosen to which the other levelswere compared: Laterality: Left; Anteriority: Posterior. Mean ERPamplitudes were analyzed via linear mixed-effects models usingthe lme4 package (Bates et al., 2015) in R, with p-values calculatedby the lmerTest package (Kuznetsova et al., 2017).A series of linear-mixed effects analyses were conducted.The dependent variable in the model was the mean ERPamplitude in the 100–280 ms time window at a given electrode.Likelihood ratio tests were conducted to evaluate these fixedeffects: Status (First Syllable of Standard vs. Deviant in orderto derive MMN), Condition (Tone 2, Tone 3, Sandhi, Mix),Laterality (Left, Midline, Right), Anteriority (Frontal, Central,Posterior), Status × Condition, Status × Anteriority, Status ×Laterality, Condition × Anteriority, Condition × Laterality,Anteriority × Laterality, Status × Condition × Anteriority,Status × Condition × Laterality, Status × Anteriority ×Laterality, Condition × Anteriority × Laterality, and Status ×Condition × Anteriority × Laterality. Fifteen models werecreated by adding one of these fixed variables at a time and werecompared via likelihood ratio tests; the model which included themost variables and fit significantly better than the one with oneless variable was chosen as the optimal model. Participant wasincluded as the random intercept, while Status and Conditionwere entered for by-Participant random slopes in all the models.Results showed that the model containing Status, Condition,Anteriority, Laterality, Status× Condition, Status× Anteriority,Status × Laterality, and Condition × Anteriority was the bestfit (χ2 = 16.582, df = 6, p = 0.011). The other more complexmodels failed to explain more variance of the ERP data. Giventhat Status is the crucial variable to index MMN effects, only theresults relevant to Status are reported below.Within the best model, a main effect of Status indicated thatERPs elicited by the deviant were significantly more negativethan those yielded by the first syllable of standards, confirmingthe presence of an MMN effect overall (β = 0.561, SE = 0.113,t = 4.980, p < 0.001). A Status× Condition interaction suggestedthat the MMN effect was significantly stronger in the Tone2 condition than in the Sandhi (β = −0.745, SE = 0.080,t = −9.277, p < 0.001) and Mix conditions (β = −0.743,SE = 0.080, t = −9.246, p < 0.001). However, the MMN effectin the Tone 2 condition did not differ from that in the Tone 3condition (β = 0.037, SE = 0.080, t = 0.465, p = 0.642). We alsofound that the MMN effect yielded by the frontal electrode wassignificantly more negative than that by the posterior electrode(β = 0.145, SE = 0.070, t = 2.087, p = 0.037), indicating a frontallydistributed MMN effect.To summarize, both Fz and multi-electrode analyses showedthat MMN effects were observed in the 100–280 ms timewindow for the Tone 2 and Tone 3 conditions, indicatingthat Mandarin speaking participants treated the deviants inthe Tone 2 and Tone 3 conditions differently from the firstsyllable of standards in these two conditions. The MMN effectobserved was frontally distributed. However, no MMN effectwas obtained in the 100–280 ms time window for the Sandhior Mix condition. In these two conditions, Mandarin-speakingparticipants seem to process the deviant and the first syllable ofstandards similarly.Deviant Omission PositionFz AnalysisSimilar to the onset position analyses, a series of linear-mixedeffects analyses were conducted on the mean ERP amplitudeselicited by the omission position and the second syllable ofFrontiers in Psychology | www.frontiersin.org 9 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 10Chien et al. Investigating Mandarin Tonal Alternations With MMNFIGURE 3 | Mean MMN difference waves (Deviant – Standard) for electrode Fz for the Tone 2 condition, Tone 3 Condition, Sandhi condition, and Mix condition atboth onset and omission positions.standards within a 100–280 ms time window using the lme4package in R (Bates et al., 2015), with p-values calculated bythe lmerTest package (Kuznetsova et al., 2017). Status (SecondSyllable of Standard, Omission) and Condition (Tone 2, Tone3, Sandhi, Mix) were set as categorical fixed factors. For Status,Omission was treated as the baseline to which Second Syllableof Standard was compared in order to derive omission MMN.For Condition, Tone 2 was entered as the baseline. Participantwas coded as random intercepts, while Status and Conditionwere entered as by-Participant random slopes. Likelihood ratiotests were conducted to compare among a series of linear-mixedeffects models in order to evaluate effects of Status, Condition,as well as the Status× Condition interaction. Three models werecreated by adding one of these fixed variables at a time and werecompared between each other; the model which included themost variables and fit significantly better than the one with oneless variable was chosen as the optimal model. Results showedthat the best model contained Status and Condition (χ2 = 14.854,df = 3, p = 0.002), suggesting that the model consisting of Status,Condition, and Status × Condition interaction failed to explainmore variance of the ERP data than the best model (χ2 = 1.766,df = 3, p = 0.622). Such a result demonstrated that mean MMNamplitudes in the 100–280 ms time window did not statisticallydiffer across conditions (see Figure 3). Below, only the effectsrelated to Status will be presented.With respect to the best model, there was an effect of Status,indicating that ERPs elicited by the omission position weresignificantly more negative than those yielded by the secondsyllable of standards (β = 1.706, SE = 0.237, t = 7.186, p < 0.001).Four follow-up linear mixed-effects analyses were performed tofurther examine whether MMN effects emerged in each conditionby comparing mean ERP amplitudes related to the omissionposition with those related to the second syllable of standardswithin each condition. Moreover, Participant was set as a randomintercept, while Status (Second Syllable of Standard, Omission)was entered for by-participant random slopes. Results showedthat the ERPs yielded by the omission position were significantlymore negative than those elicited by the second syllable ofstandards in all the four conditions, confirming MMN effects(Tone 2 condition: β = 1.459, SE = 0.348, t = 4.187, p < 0.001;Tone 3 condition: β = 1.688, SE = 0.358, t = 4.722, p < 0.001;Sandhi condition: β = 1.753, SE = 0.277, t = 6.325, p < 0.001;β = 1.925, SE = 0.352, t = 5.463, p < 0.001) (see Figure 2).Multi-Electrode AnalysisSimilar to the onset analyses, the omission multi-electrodeanalyses used 9 representative electrodes: F3, Fz, F4, C3, Cz, C4,P3, Pz, and P4. A series of linear-mixed effects analyses wereconducted on mean ERP amplitudes at the 100–280 ms timewindow of a given electrode using the lme4 package (Bates et al.,2015) in R, with p-values calculated by the lmerTest package(Kuznetsova et al., 2017). Status (Second Syllable of Standard vsOmission in order to derive omission MMN), Condition (Tone2, Tone 3, Sandhi, Mix), Anteriority (Frontal, Central, Posterior)and Laterality (Left, Middle, Right) were entered as categoricalfixed factors in which Omission, Tone 2, Posterior and Left wereFrontiers in Psychology | www.frontiersin.org 10 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 11Chien et al. Investigating Mandarin Tonal Alternations With MMNselected as the baseline for each factor. Participant was includedas the random intercept, while Status and Condition were usedas by-Participant random slopes. Likelihood ratio tests wereperformed to evaluate effects of Status, Condition, Anteriority,Laterality, Status × Condition, Status × Anteriority, Status ×Laterality, Condition × Anteriority, Condition × Laterality,Anteriority × Laterality, Status × Condition × Anteriority,Status × Condition × Laterality, Status × Anteriority ×Laterality, Condition × Anteriority × Laterality, and Status ×Condition × Anteriority × Laterality. Fifteen models werecreated by adding one of these fixed variables at a time andwere compared via likelihood ratio tests; the model whichincluded the most variables and fit significantly better than theone with one less variable was chosen as the optimal model.Results showed that the model comprising Status, Condition,Anteriority, Laterality, Status× Condition, Status× Anteriority,Status × Laterality, Condition × Anteriority was the best(χ2 = 23.838, df = 6, p < 0.001). The other more complicatedmodels could not explain more variance of the ERP data. Giventhat Status, which indexes MMN effects, is most relevant tothe current research interests, only those effects involving Statuswill be presented.Within the best model, an effect of Status was obtained, asevidenced by the fact that ERPs elicited by the omission positionwere significantly more negative than those yielded by the secondsyllable of standards (β = 0.534, SE = 0.214, t = 2.497, p = 0.016).Furthermore, a Status×Condition interaction was also obtained,as revealed by significantly stronger omission MMN effects forthe Tone 2 condition than for the Tone 3 condition (β = 0.265,SE = 0.098, t = 2.695, p = 0.007), the Sandhi condition (β = 0.229,SE = 0.098, t = 2.328, p = 0.020), and the Mix condition (β = 0.382,SE = 0.098, t = 3.888, p < 0.001). There was also an interactionbetween Status and Anteriority, showing significantly strongeromission MMN effects for frontal and central electrodes than forposterior electrodes (Frontal vs Posterior: β = 0.866, SE = 0.085,t = 10.166, p < 0.001; Central vs. Posterior: β = 0.493, SE = 0.085,t = 5.787, p < 0.001), suggesting a more frontal distribution.In summary, the results revealed that omission MMN effectswere elicited by the omission position in all four conditions, andwere frontally distributed. Moreover, the omission effect in theTone 2 condition was significantly greater than that in the Tone3 condition, Sandhi condition, and Mix condition.DISCUSSIONOnset MMNThe current study used the mismatch negativity to examinethe processing of phonological alternation. Specifically, weinvestigated how Mandarin speakers process and representMandarin tone 3 sandhi words using MMN. At the deviant onsetposition, we observed frontally distributed MMN effects only inthe Tone 2 and Tone 3 conditions, but not in the Sandhi or Mixcondition. Moreover, the MMN effects in the Tone 2 and Tone 3conditions did not differ. These results were consistent with theanalysis that compared the ERPs elicited by the deviant acrossthe four conditions. In particular, we found that the ERP yieldedin the Tone 2 condition did not differ from that in the Tone 3condition, but was significantly more negative in amplitude thanthat in the Sandhi and Mix conditions.Most importantly for understanding the processing of sandhiwords, we found a significant difference between the Tone 2condition (standard [tùu2 je4] /tùu2 je4/ and deviant [tùu2]/tùu2/) and the Sandhi condition (standard [tùu2 jen3] /tùu3 jen3/and deviant [tùu2] /tùu2/) in that an MMN effect emerged in theTone 2 condition but not in the Sandhi condition. Recall that theacoustic information of the first syllable of standards in these twoconditions was identical. Moreover, the deviants used in thesetwo conditions were also identical. Despite this manipulation, weonly observed an MMN effect in the Tone 2 condition but not inthe Sandhi condition. This pattern of data suggests distinct neuralmechanisms when processing tone 2 standards ([tùu2 je4] /tùu2je4/) and sandhi standards ([tùu2 jen3] /tùu3 jen3/).The lack of an MMN effect in the Sandhi condition may bedue to the representation of Mandarin tone 3 sandhi. Recallthat Mandarin tone 3 sandhi is triggered by the phonologicalenvironment, that is, a tone 3 syllable undergoes tone sandhi,changing into a tone 2 syllable when followed by another tone3 syllable. For our Sandhi condition, participants heard manytone 3 sandhi words ([tùu2 jen3] /tùu3 jen3/). Since no MMNeffect was observed when hearing a tone 2 deviant in the Sandhicondition, participants did not seem to treat the standards aseither a tone 2 (tone 3V) or a tone 3; otherwise, an MMN effectshould have been elicited as in the Tone 2 and Tone 3 conditions.This result pattern contrasts to Li and Chen (2015) in whichboth sandhi tone 3V and tone 3 were reported to be activated.For the Sandhi condition, participants seem to be assigning anunderspecified representation for the first syllable of the tone3 sandhi words, which is neither a tone 2 nor a tone 3. Thedeviant ([tùu2] /tùu2/) did not elicit an MMN effect in the Sandhicondition even though the same acoustic differences between thefirst syllable of sandhi standards and the deviant yielded suchan effect in the Tone 2 condition. This pattern is consistentwith a proposal that Mandarin tone 3 may be underspecified.ERP studies (Politzer-Ahles et al., 2016) also offered evidence forthe underspecified nature of tone 3, showing that monosyllabictone 3 words, as contrasted with words with other tones, elicitedasymmetrical MMN effects in an oddball paradigm. However,as Politzer-Ahles et al. (2016) also note, unlike coronal place ofarticulation which has several underspecified properties (Lahiriand Reetz, 2002, 2010), the falling-rising F0 contour of Mandarintone 3 is one of the most extreme tones to produce, the mostdifficult tone to acquire for children (Wong et al., 2005; Wong,2012), and the most infrequent tone in token frequency (Zhangand Lai, 2010), suggesting that more research is necessary toexplore what acoustic properties of Mandarin tone 3 mightcontribute to its underspecified representation.Another possible explanation for the lack of MMN effectsin the Sandhi condition may be that participants productivelyapplied the tone 3 sandhi rule by always converting the initialsurface tone 2 into an underlying tone without exception (evenfor the deviant [tùu2] /tùu2/). Given that participants heard thesame sandhi word many times in this condition, they mighthave always predicted the upcoming stimulus to be a sandhiword, thus interpreting even the deviant [tùu2] /tùu2/ as thefirst syllable of the sandhi standard, focusing mainly on theFrontiers in Psychology | www.frontiersin.org 11 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 12Chien et al. Investigating Mandarin Tonal Alternations With MMNunderlying tone. Therefore, when hearing the deviant [tùu2],participants may have ignored the surface form and extracted itsunderlying tone, so that it would not be in conflict with the firstsyllable of sandhi standards, thus resulting in no MMN effect.Future research about the nature of the abstract representationof tone sandhi (as either underspecified or as a tone 3) is neededto differentiate these two accounts. If the lack of an MMNeffect in the Sandhi condition was really due to the above-mentioned reason, it would be consistent with Chien et al.’s(2016) priming results in that surface tonal overlap betweenprimes and targets did not significantly facilitate Mandarin-speaking participants’ lexical decision responses while underlyingtonal overlap between primes and targets did show significantfacilitation. It would further suggest that models of spokenword recognition that rely solely on surface forms need tobe modified.For the Tone 2 condition, notice that no sandhi environmentexisted in this condition. It is unlikely for the Mandarin listenersto posit that the tone 2 deviant was derived via the sandhi process,so that the first syllable of the standard [tùu2 je4] /tùu2 je4/and the deviant [tùu2] /tùu2/ should match in both surface andunderlying forms. However, our results showed an MMN effectin the Tone 2 condition. The significant MMN was most likelydue to the acoustic differences between the deviants and thefirst syllable of standards. Recall that the average fundamentalfrequency of the seven [tùu2] tokens of the standards in the Tone2 condition was 10.28 Hz higher than that of the deviant [tùu2].In addition, the 1F0 value of the average F0 contour of thestandard [tùu2] was 9.93 Hz, while the 1F0 value of the deviant[tùu2] was 15.23 Hz. The significant MMN effect in the Tone2 condition is consistent with previous studies (with tones andwith vowels) showing MMN effects when deviants and standardsdiffered acoustically (within the same phoneme category) but notphonemically (in different phoneme categories) (Winkler et al.,1999; Kasai et al., 2001; Xi et al., 2010).With respect to the Tone 3 condition, there was no sandhicontext either. It is unlikely for the Mandarin listeners to expectthat the tone 2 deviant came from the sandhi process. Therefore,the mismatch between the first syllable of tone 3 standards ([tone3 tone 4] /tone 3 tone 4/) and the deviant ([tùu2] /tùu2/) inboth surface and underlying forms yielded a significant MMNeffect. Being regularly exposed to the standards [tùu3 je4] /tùu3je4/ in the Tone 3 condition, Mandarin-speaking participantsinterpreted the first syllable they heard in a trial to be a tone 3.Encountering the deviant [tùu2] /tùu2/, mismatching both on thesurface and underlyingly, resulted in a significantly more negativeERP waveform relative to that yielded by the first syllable of thetone 3 standards ([tùu3 je4] /tùu3 je4/).However, the present data show no difference in MMNamplitudes between the Tone 2 and Tone 3 conditions. Previousneurophysiological studies have demonstrated that across-category differences elicited stronger mismatch responses thanwithin-category differences (Dehaene-Lambertz, 1997; Winkleret al., 1999; Sharma and Dorman, 2000; Kasai et al., 2001;Nenonen et al., 2003; Xi et al., 2010). The pattern of resultscould be due to the acoustic information of the tone 2 deviant,the first syllable of the T2T4 standards, and the first syllable ofthe T3T4 standards. Notice that there were acoustic differencesbetween the first syllable of T2T4 standards and tone 2 deviantin average F0 and 1F0. Moreover, a lower average F0 and alarger 1F0 value are the acoustic characteristics of tone 3, and thetone 2 deviant happened to have these acoustic characteristics.The fact that the tone 2 deviant was acoustically different fromthe first syllable of the T2T4 standards and had some acoustictraits similar to tone 3 may be able to explain why there wasno MMN difference between the Tone 2 and Tone 3 conditions.An alternative explanation for the lack of an MMN differencebetween the Tone 3 and Tone 2 conditions in the current studycould be that Mandarin tone 3 is underspecified (Politzer-Ahleset al., 2016) or represented as both tone 3 and sandhi tone (3V) (Liand Chen, 2015). For Politzer-Ahles et al. (2016) and Li and Chen(2015), weaker MMN effects were found when monosyllabic tone3 served as the standard and monosyllabic tone 2 served as thedeviant compared to a reverse condition. However, if Mandarintone 3 is represented as both tone 3 and sandhi tone (3V) as Liand Chen proposed, it would be difficult to explain why there wasnot an MMN effect in the Sandhi condition. The activated sandhitone (acoustically similar to tone 2) does match the surface toneof the first syllable of sandhi standards. If participants focus onthe surface tone, this would lead to an MMN effect. The fact thatwe did not observe an MMN effect in the Sandhi condition andfound comparable MMN amplitudes between Tone 2 and Tone3 conditions may be more consistent with the underlying or evenabstract (underspecified) representation view.For the Mix condition in which standards consisted ofdisyllabic tone 3 sandhi words ([tone 2 tone 3] /tone 3 tone 3/)mixed with disyllabic words starting with a tone 3 syllable ([tone3 tone 4] /tone 3 tone 4/), we did not observe an MMN effectwhen comparing the first syllable of standards with the deviant.Recall that this condition demonstrated a many-to-one ratio onlyin the underlying form between the first syllable of standards andthe deviant (/tùu2/). If Mandarin-speaking participants had paidattention to the underlying form and abstracted away from thesurface variability, an MMN effect should have been observed (allstandards are underlying tone 3 stimuli). The lack of an MMNeffect in the Mix condition indicated that participants were notsensitive to the many-to-one ratio in the underlying form, whichis necessary to elicit MMNs, or at least they could not extractthe underlying many-to-one ratio to generate a significant MMNeffect. The combination of tone 2 and tone 3 for the first syllableof standards on the surface, together with segmental variability inthe second syllable across standards (not present in the standardsin the other conditions) in the Mix condition may have made ittoo difficult for the participants to form any regularity for the toneof the first syllable.Omission MMNAn additional finding of the current study was the observationof omission MMN effects in all four conditions. These omissioneffects indicate a violation of participants’ prediction of hearingdisyllabic stimuli, suggesting that participants did process themonosyllabic deviant as the first syllable of a disyllabic chunk.When participants did not hear the second syllable in the(monosyllabic) deviants, an omission MMN effect was generated.Frontiers in Psychology | www.frontiersin.org 12 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 13Chien et al. Investigating Mandarin Tonal Alternations With MMNAn omission MMN effect has been reported in the previousliterature and used to suggest a predictive or rule formingmechanism in the auditory cortex (Raij et al., 1997; Yabe et al.,1997, 1998; Janata, 2001; Bendixen et al., 2009, 2014). Forexample, Bendixen et al. (2014) sought evidence for predictiveprocessing from mismatch responses to omitted speech segments.In their first experiment, German words Lachs ([laks], “salmon”)or Latz ([lats], “bib”) were presented to German listenersauditorily, with La ([la], no semantic meaning) occasionallymixed in. Results showed significantly stronger omission MMNeffects when the context licensed a prediction for the specificfinal segments (only [laks] as the standard, or only [lats] as thestandard, versus [la] as the deviant) compared to when suchprediction was not possible (both [laks] and [lats] as standardsversus [la] as the deviant). The Bendixen et al. (2014) results showpredictive coding mechanisms in the central auditory system forspoken word recognition.The omission MMN effects observed in the current studycould be interpreted in two ways. If the present omission MMNsmerely reflect the absence of a second syllable, amplitudes ofthe omission MMNs may be predicted to be similar acrossconditions. However, if the current omission MMNs reflect notonly the absence of a second syllable (i.e., same situation acrossconditions), but also predictability differences of the preciseproperties of the second syllable, differences in size of theomission MMNs across conditions may be expected.The current omission MMN results showed a significantlystronger MMN effect for the Tone 2 condition than for theTone 3 condition, Sandhi condition, and Mix condition for themulti-electrode analysis. The stronger omission MMN effect inthe Tone 2 condition compared to that in the Tone 3 and Mixconditions seems to result from differences in predictability forthe second syllable, suggesting either a violation of predictedsegmental or tonal information of the second syllable (Näätänenet al., 1997, 2007; Bendixen et al., 2012, 2014). In the Tone2 condition, since the deviant matched the first syllable ofstandards in both segmental and tonal information, participantsmight have deemed the deviant [tùu2] as the beginning ofthe exact word used as standards (i.e., [tùu2 je4] /tùu2 je4/)for the Tone 2 condition. Therefore, participants in the Tone2 condition could predict a precise second syllable for thedeviant. When they did not hear the expected second syllable,a strong omission MMN effect was elicited. In contrast, forthe Tone 3 condition, [tùu3 je4] /tùu3 je4/ was used as thestandard, which mismatched the deviant in tone. Upon hearingthe deviant, participants might have regarded it as the initialsyllable of a different word rather than that of the standard.Consequently, neither segmental nor tonal information for theomission could be precisely predicted, leading to a weakeromission MMN effect. Regarding the Mix condition, givensegmental and tonal variability on the surface, it is even moredifficult for participants to predict what could have been in theomission position. This may have led to the weaker omissionMMN effect observed; this condition also yielded a relativelylate ERP peak (based on visual inspection). The omission MMNdifference between the Tone 2 condition and Sandhi conditionwas also significant, which is more difficult to explain dueto similar predictability of the content of the second syllableacross the two conditions. Further research is required in orderto investigate whether and how the Sandhi context in thestandards may impact how the predictive coding of upcomingsegments occurs.Typicality of the MMN EffectsAt the onset position, MMNs emerged when the deviant andthe standard differed in the F0 contour (Tone 2 condition),and when the deviant and the standard mismatched in both theunderlying and the surface forms (Tone 3 condition). The MMNselicited by the deviant in the Tone 2 and Tone 3 conditionsappear to be relatively broad without a clear peak and theyhave a fronto-central scalp distribution, in contrast to previousstudies that have observed MMNs elicited by tonal contrast whichshowed obvious MMN peaks (e.g., Näätänen et al., 2007; Li andChen, 2015; Politzer-Ahles et al., 2016). However, the currentMMN effects were observed in the typical 100–280 ms timewindow. Recall that we used disyllables as standards in all fourconditions, in contrast to other studies where monosyllables wereused as standards. Another unique feature of our study is thatall of the deviants are expected to yield MMN not only at theonset of the first syllable, but also at the onset of the omittedsecond syllable. It is possible that these factors contributed tothe relatively broad MMNs at the onset position in our findings.Future studies could compare the deviant response at the onsetof the first syllable that are followed by omission (as in thecurrent study) and deviant responses at the onset of the firstsyllable that are followed by a second syllable. This would shedlight on whether the presence of a second syllable impactsthe MMN waveforms.On the other hand, the MMNs that we observed atthe omission position in all four conditions show typicalcharacteristics of the component found in previous studies(Näätänen et al., 2007; Bendixen et al., 2012, 2014). Theomission MMNs emerged in the 100–280 ms time windowafter the onset of the omission position with an overall fronto-central scalp distribution. Unlike the broad waveforms atthe onset position, the omission MMNs showed clear peaksaround 100–200 ms after the onset of the omission positionin the waveforms especially for the Tone 2, Sandhi, andMix condition, although less so for the Tone 3 condition.The Tone 3 condition is the only case where the deviantand the first syllable of standards have no overlap in eithersurface or underlying forms. Future studies could explorethis issue by testing whether the omission MMN waveformsvary depending on the similarity between the deviants andthe corresponding standards across more than one level oflinguistic representation.Finally, previous studies have shown that the MMN effectselicited by durational differences between standards and deviantscould be severely reduced or completely eliminated whenthe traditional method was employed compared to when thesame-stimulus method was conducted (Shelley et al., 1991; Cattset al., 1995; Peter et al., 2010). Peter et al. (2010) explained that thereduction/absence of MMN could be due to the confound in timebetween the shorter deviant offset P2 and the MMN. However,Frontiers in Psychology | www.frontiersin.org 13 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 14Chien et al. Investigating Mandarin Tonal Alternations With MMNthe current study showed significant and typical omission MMNeffects, suggesting that the omission MMN was strong enough toresist being canceled out by the P2, or the shorter deviant offsetP2 and the omission MMN did not overlap in time.In sum, the current study used disyllabic Mandarin tone3 sandhi words to investigate how Mandarin speakers processthe first syllable of tone 3 sandhi words, which is realized asa tone 2 on the surface. The present results suggest differentneural mechanisms for processing canonical tone 2 syllablesand sandhi tone 2 syllables in the initial position of disyllabicwords. Due to the similarity between the canonical tones andthe surface tones of the first syllable of standards in the Tone2 and Tone 3 conditions, Mandarin speakers did not need toabstract away from the surface tones and process the standards.When processing tone 3 sandhi words, due to the mismatchbetween the canonical and surface tones, Mandarin speakers maybe required to abstract away from the surface form and focus onretrieving the underlying form (Chien et al., 2016), which is eitheran underspecified representation or a tone 3 representation, sothat they could recognize the sandhi words.CONCLUSIONThe present study investigated how Mandarin tone sandhi wordsare processed using MMN, a cortical response allowing one todirectly assess the pre-attentive processing of auditory linguisticrepresentations. Our data suggest that different neural processingmechanisms are involved in the processing of canonical tone 2syllables and sandhi tone 2 syllables in disyllabic sequences. Weproposed two possibilities for the processing of Mandarin tone 3sandhi words: An underspecified tonal representation assignedto the first syllable of Mandarin tone 3 sandhi words, or anunderlying form /tone 3 tone 3/, both of which require Mandarinspeakers to computationally convert the surface tone 2 into anunderlying tone using a tone 3 sandhi rule.DATA AVAILABILITY STATEMENTThe datasets generated for this study are available on request tothe corresponding author.ETHICS STATEMENTThe studies involving human participants were reviewed andapproved by the Human Subjects Committee at the University ofKansas. The participants provided their written informed consentto participate in this study.AUTHOR CONTRIBUTIONSY-FC came up with the research ideas, wrote themanuscript, conducted the experiment, ran statistics,interpreted the results, and revised the manuscript. XYconducted the experiment, analyzed the data, interpretedthe results, and revised the manuscript. RF discussedthe research idea and methods with Y-FC, analyzed thedata, interpreted the results, and revised the manuscript.JS discussed the research ideas and methods with Y-FC,analyzed the data, interpreted the results, and revisedthe manuscript.FUNDINGThis research was supported by the Department of Linguistics,University of Kansas, Lawrence, KS, and Humanity and SocialScience Youth Foundation of Ministry of Education of Chinaunder grant 18YJC740075. This study is partially based on thefirst author’s dissertation (Chien, 2016).REFERENCESBates, D., Maechler, M., Bolker, B. M., and Walker, S. (2015). lme4: linear mixed-effects models using Eigen and S4. J. Stat. Softw. 67, 1–48.Bendixen, A., SanMiguel, I., and Schröger, E. (2012). Early electrophysiologicalindicators for predictive processing in audition: a review. Int. J. Psychophysiol.83, 120–131. doi: 10.1016/j.ijpsycho.2011.08.003Bendixen, A., Scharinger, M., Strauβ, A., and Obleser, J. (2014). Predictionin the service of comprehension: modulated early brain responses toomitted speech segments. Cortex 53, 9–26. doi: 10.1016/j.cortex.2014.01.001Bendixen, A., Schröger, E., and Winkler, I. (2009). I heard that coming: event-related potential evidence for stimulus-driven prediction in the auditorysystem. J. Neurosci. 29, 8447–8451. doi: 10.1523/JNEUROSCI.1493-09.2009Boersma, P., and Weenink, D. (2015). Praat: Doing Phonetics by Computer[Computer program]. Version 5.3.56. Available online at: http://www.praat.org/(accessed September 15, 2013).Catts, S. V., Shelley, A. M., Ward, P. B., Liebert, B., McConaphy, N., Andrews, S.,et al. (1995). Brain potential evidence for an auditory sensory memory deficit inschizophrenia. Am. J. Psychiatry 152, 213–219. doi: 10.1176/ajp.152.2.213Chandrasekaran, B., Krishnan, A., and Gandour, J. T. (2007). Mismatch negativityto pitch contours is influenced by language experience. Brain Res. 1128, 148–156. doi: 10.1016/j.brainres.2006.10.064Chandrasekaran, B., Krishnan, A., and Gandour, J. T. (2009). Sensory processing oflinguistic pitch as reflected by the mismatch negativity. Ear Hear. 30, 552–558.doi: 10.1097/AUD.0b013e3181a7e1c2Chao, Y. (1968). A Grammar of Spoken Chinese. Berkeley, CA: University ofCalifornia Press.Cheng, C.-C. (1973). A Synchronic Phonology of Mandarin Chinese. The Hague:Mouton.Cheng, R. L. (1966). Mandarin phonological structure. J. Linguist. 2, 135–158.doi: 10.1017/s0022226700001444Chien, Y.-F. (2016). Observing the Contribution of both Underlying and SurfaceRepresentations: Evidence from Priming and Event-related Potentials. Ph.D.dissertation, University of Kansas, KansasChien, Y.-F., Sereno, J. A., and Zhang, J. (2016). Priming the representation ofMandarin tone 3 sandhi words. Lang. Cogn. Neurosci. 31, 179–189. doi: 10.1080/23273798.2015.1064976Dehaene-Lambertz, G. (1997). Electrophysiological correlates of categoricalphoneme perception in adults. Neuroreport 8, 919–924. doi: 10.1097/00001756-199703030-00021Frontiers in Psychology | www.frontiersin.org 14 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 15Chien et al. Investigating Mandarin Tonal Alternations With MMNDelorme, A., and Makeig, S. (2004). EEGLAB: an open source toolbox for analysisof single-trial EEG dynamics including independent component analysis.J. Neurosci. Methods 134, 9–21. doi: 10.1016/j.jneumeth.2003.10.009Dow, F. (1972). A discussion on tone sandhi problems in Chinese. J. Int. Phon.Assoc. 2, 13–19.Duanmu, S. (2007). The Phonology of Standard Chinese. Oxford: Oxford UniversityPress.Dunn, B. R., Dunn, D. A., Languis, M., and Andrew, D. (1998). The relation ofERP components to complex memory processing. Brain Cogn. 36, 355–376.doi: 10.1006/brcg.1998.0998Goldinger, S. D., Pisoni, D. B., and Logan, J. S. (1991). On the nature of talkervariability effects on recall of spoken word lists. J. Exp. Psychol. Learn. Mem.Cogn. 17, 152–162. doi: 10.1037/0278-7393.17.1.152Groppe, D. M., Makeig, S., and Kutas, M. (2009). Identifying reliable independentcomponents via split-half comparisons. Neuroimage 45, 1199–1211. doi: 10.1016/j.neuroimage.2008.12.038Gu, F., Li, J., Wang, X., Hou, Q., Huang, Y., and Chen, L. (2012). Memorytraces for tonal language words revealed by auditory event-related potentials.Psychophysiology 49, 1353–1360. doi: 10.1111/j.1469-8986.2012.01447.xHsu, C., Evans, J., and Lee, C. (2015). Brain responses to spoken F0 changes: is Hspecial? J. Phon. 51, 82–92. doi: 10.1016/j.wocn.2015.02.003Janata, P. (2001). Brain electrical activity evoked by mental formation of auditoryexpectations and images. Brain Top. 13, 169–193.Kasai, K., Yamada, H., Kamio, S., Nakagome, K., Iwanami, A., Fukuda, M., et al.(2001). Brain lateralization for mismatch response to across- and within-category change of vowels. Neuroreport 12, 2467–2471. doi: 10.1097/00001756-200108080-00036Kazanina, N., Phillips, C., and Idsardi, W. (2006). The influence of meaning on theperception of speech sounds. Proc. Natl. Acad. Sci. U.S.A. 103, 11381–11386.doi: 10.1073/pnas.0604821103Kim, K. H., Kim, J. H., Yoon, J., and Jung, K.-Y. (2008). Influence of task difficultyon the features of event-related potential during visual oddball task. Neurosci.Lett. 445, 179–183. doi: 10.1016/j.neulet.2008.09.004Kujala, T., Tervaniemi, M., and Schröger, E. (2007). The mismatch negativityin cognitive and clinical neuroscience: theoretical and methodologicalconsiderations. Biol. Psychol. 74, 1–19. doi: 10.1016/j.biopsycho.2006.06.001Kuo, Y., Xu, Y., and Yip, M. (2007). “The phonetics and phonology ofapparent cases of iterative tone change in standard Chinese,” in Tones andTunes: Experimental Studies in Word and Sentence Prosody, Vol. 2, eds C.Gussenhoven, and T. Riad, (Cambridge: Cambridge University Press), 211–237.Kuznetsova, A., Brockhoff, P. B., and Christensen, R. H. B. (2017). lmerTest –Tests in Linear Mixed Effects Models. R Package Version, 2-0. Available onlineat: https://github.com/runehaubo/lmerTestR(accessed December 13, 2019).Lahiri, A., and Marslen-Wilson, M. (1991). The mental representation of lexicalform: a phonological approach to the recognition lexicon. Cognition 38, 245–294. doi: 10.1016/0010-0277(91)90008-rLahiri, A., and Reetz, H. (2002). Underspecified recognition. Lab. Phonol. 7,637–675.Lahiri, A., and Reetz, H. (2010). Distinctive features: phonologicalunderspecification in representation and processing. J. Phon. 38, 44–59.doi: 10.1016/j.wocn.2010.01.002Landi, M., Crowley, M. J., Wu, J., Bailey, C. A., and Mayes, L. C. (2012). DeviantERP response to spoken non-words among adolescents exposed to cocaine inutero. Brain Lang. 120, 209–216. doi: 10.1016/j.bandl.2011.09.002Li, X., and Chen, Y. (2015). Representation and processing of lexical tone andtonal variants: evidence from the mismatch negativity. PLoS One 10:e0143097.doi: 10.1371/journal.pone.0143097Lin, Y.-H. (2007). The Sounds of Chinese. Cambridge: Cambridge University Press.Makeig, S., Bell, A. J., Jung, T.-P., and Sejnowski, T. J. (1996). “Independentcomponent analysis of electroencephalographic data,” in Advances in NeuralInformation Processing Systems, eds D. Touretzky, M. Mozer, and M. Hasselmo,(Cambridge, MA: MIT Press), 45–151.Näätänen, R. (1992). Attention and Brain Function. Hillsdale, MI: Erlbaum.Näätänen, R., Lehtokoski, A., Lennes, M., Cheour, M., Huotilainen, M., Iivonen, A.,et al. (1997). Language-specific phoneme representations revealed by electricand magnetic brain responses. Nature 385, 432–434. doi: 10.1038/385432a0Näätänen, R., Paavilainen, P., Rinne, T., and Alho, K. (2007). The mismatchnegativity (MMN) in basic research of central auditory processing: a review.Clin. Neurophysiol. 118, 2544–2590. doi: 10.1016/j.clinph.2007.04.026Näätänen, R., and Winkler, I. (1999). The concept of auditory stimulusrepresentation in neuroscience. Psychol. Bull. 125, 826–859. doi: 10.1037/0033-2909.125.6.826Nenonen, S., Shestakova, A., Huotilainen, M., and Näätänen, R. (2003). Linguisticrelevance of duration within the native language determines the accuracy ofspeech-sound duration processing. Cogn. Brain Res. 16, 492–495. doi: 10.1016/s0926-6410(03)00055-7Oldfield, R. C. (1971). The assessment and analysis of handedness: the Edinburghinventory. Neuropsychologia 9, 97–113. doi: 10.1016/0028-3932(71)90067-4Peng, S.-H. (2000). “Lexical versus ‘phonological’ representations of Mandarintones,” in Language Acquisition and the Lexicon: Papers in Laboratory PhonologyV, eds M. B. Broe, and J. B. Pierrehumbert, (Cambridge: Cambridge UniversityPress), 152–167.Peter, V., McArthur, G., and Thompson, W. F. (2010). Effect of deviance directionand calculation method on duration and frequency mismatch negativity(MMN). Neurosci. Lett. 482, 71–75. doi: 10.1016/j.neulet.2010.07.010Phillips, C., Pellathy, T., Marantz, A., Yellin, E., Wexler, K., Poeppel, D., et al.(2000). Auditory cortex accesses phonological categories: an MEG mismatchstudy. J. Cogn. Neurosci. 12, 1038–1055. doi: 10.1162/08989290051137567Politzer-Ahles, S., Schluter, K., Wu, K., and Almeida, D. (2016). Asymmetries inthe perception of Mandarin tones: evidence from mismatch negativity. J. Exp.Psychol. Hum. Percept. Perform. 42, 1547–1570. doi: 10.1037/xhp0000242Pulvermüller, F., and Shtyrov, Y. (2006). Language outside the focus of attention:the mismatch negativity as a tool for studying higher cognitive processes. Prog.Neurobiol. 79, 49–71. doi: 10.1016/j.pneurobio.2006.04.004Raij, T., McEvoy, L., Mäkelä, J. P., and Hari, R. (1997). Human auditory cortexis activated by omissions of auditory stimuli. Brain Res. 745, 134–143. doi:10.1016/s0006-8993(96)01140-7Reetz, H., and Jongman, A. (2009). Phonetics: Transcription, Production, Acoustics,and Perception. Hoboken, NJ: Wiley-Blackwell.Ren, G.-Q., Yang, Y., and Li, X. (2009). Early cortical processing of linguisticpitch patterns as revealed by the mismatch negativity. Neuroscience 162, 87–95.doi: 10.1016/j.neuroscience.2009.04.021Sharma, A., and Dorman, M. F. (2000). Neurophysiological correlates ofcrosslanguage phonetic perception. J. Acoust. Soc. Am. 107, 2697–2703. doi:10.1080/87565641.2013.805218Shei, C. (2014). Understanding the Chinese Language: A Comprehensive LinguisticIntroduction. Abingdon: Routledge.Shelley, A. M., Ward, P. B., Catts, S. V., Michie, P. T., Andrews, S., and McConaghy,N. (1991). Mismatch negativity: an index of a preattentive processing deficitin schizophrenia. Biol. Psychiatry 30, 1059–1062. doi: 10.1016/0006-3223(91)90126-7Shih, C. (1997). “Mandarin third tone sandhi and prosodic structure,” in Studiesin Chinese Phonology, eds J. Wang, and N. Smith, (Berlin: Walter de Gruyter),81–124.Wang, W., and Li, K.-P. (1967). Tone 3 in Pekingese. J. Speech Hear. Res. 10,629–636.Wang, X.-D., Gu, F., He, K., Chen, L.-H., and Chen, L. (2012). Pre-attentiveextraction of abstract auditory rules in speech sound stream: a mismatchnegativity study using lexical tones. PLoS One 7:e30027. doi: 10.1371/journal.pone.0030027Winkler, I., Lehtokoski, A., Alku, P., Vainio, M., Czigler, I., Csepe, V., et al. (1999).Preattentive detection of vowel contrasts utilizes both phonetic and auditorymemory representations. Cogn. Brain Res. 7, 357–369. doi: 10.1016/s0926-6410(98)00039-1Wong, P. (2012). Acoustic characteristics of three-year-olds’ correct and incorrectmonosyllabic Mandarin lexical tone productions. J. Phon. 40, 141–151. doi:10.1016/j.wocn.2011.10.005Wong, P., Schwartz, R. G., and Jenkins, J. J. (2005). Perception and production oflexical tones by 3-year-old, Mandarin-speaking children. J. Speech Lang. Hear.Res. 48, 1065–1079. doi: 10.1044/1092-4388(2005/074)Xi, J., Zhang, L., Shu, H., Zhang, Y., and Li, P. (2010). Categorical perception oflexical tones in Chinese revealed by mismatch negativity. Neuroscience 170,223–231. doi: 10.1016/j.neuroscience.2010.06.077Frontiers in Psychology | www.frontiersin.org 15 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 16Chien et al. Investigating Mandarin Tonal Alternations With MMNYabe, H., Tervaniemi, M., Reinikainen, K., and Näätänen, R. (1997). Temporalwindow of integration revealed by MMN to sound omission. Neuroreport 8,1971–1974. doi: 10.1097/00001756-199705260-00035Yabe, H., Tervaniemi, M., Sinkkonen, J., Huotilainen, M., Ilmoniemi, R. J.,and Näätänen, R. (1998). Temporal window of integration of auditoryinformation in the human brain. Psychophysiology 35, 615–619. doi: 10.1017/s0048577298000183Yuan, J. H., and Chen, Y. (2014). 3rd tone sandhi in Standard Chinese: a corpusapproach. J. Chin. Linguist. 42, 218–237.Zhang, C., Xia, Q., and Peng, G. (2015). Mandarin third tone sandhi requires moreeffortful phonological encoding in speech production: evidence from an ERPstudy. J. Neurolinguist. 33, 149–162. doi: 10.1016/j.jneuroling.2014.07.002Zhang, J., and Lai, Y.-W. (2010). Testing the role of phonetic knowledgein Mandarin tone sandhi. Phonology 27, 153–201. doi: 10.1017/s0952675710000060Zhou, X., and Marslen-Wilson, W. D. (1997). “The abstractness of phonologicalrepresentation in the Chinese mental lexicon,” in Cognitive Processing ofChinese and other Asian Languages, ed. H.-C. Chen, (Hong Kong: The ChineseUniversity Press), 3–26.Conflict of Interest: The authors declare that the research was conducted in theabsence of any commercial or financial relationships that could be construed as apotential conflict of interest.Copyright © 2020 Chien, Yang, Fiorentino and Sereno. This is an open-access articledistributed under the terms of the Creative Commons Attribution License (CC BY).The use, distribution or reproduction in other forums is permitted, provided theoriginal author(s) and the copyright owner(s) are credited and that the originalpublication in this journal is cited, in accordance with accepted academic practice. Nouse, distribution or reproduction is permitted which does not comply with these terms.Frontiers in Psychology | www.frontiersin.org 16 April 2020 | Volume 11 | Article 646fpsyg-11-00646 April 6, 2020 Time: 18:17 # 17Chien et al. Investigating Mandarin Tonal Alternations With MMNAPPENDIX AAcoustic details of the seven standard [tùu2] tokens in the Tone 2 and Sandhi conditions, the seven standard [tùu3] tokens in the Tone3 condition, the four standard [tùu2] tokens and the three standard [tùu3] tokens in the Mix condition, and the deviant [tùu2] for allfour conditions (measured in Hz, rounded to the nearest integer).Tone 2 condition and Sandhi condition: Standard [tùu2].Time (% of tone) 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1F0 Turning Point (ms)[tùu2] (1) 160 159 157 157 160 166 173 183 194 205 208 3.19 51[tùu2] (2) 167 164 161 160 161 166 173 183 194 203 205 7.34 61[tùu2] (3) 169 164 159 157 157 160 166 176 186 194 196 12.45 71[tùu2] (4) 167 165 165 169 178 190 203 216 227 235 236 2.63 41[tùu2] (5) 164 161 157 154 154 156 159 165 173 182 186 10 81[tùu2] (6) 176 164 155 152 150 150 152 157 165 174 176 26.45 91[tùu2] (7) 155 152 149 148 149 152 157 164 174 186 191 7.44 61Mean (SD) 165 (6.2) 161 (4.4) 158 (4.6) 157 (6.1) 158 (9.1) 163 (12.6) 169 (15.7) 178 (18) 187 (19.2) 197 (18.6) 200 (18.1) 9.93 (7.47) 65 (15.9)Tone 3 condition: Standard [tùu3].Time (% of tone) 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%[tùu3] (1) 158 154 147 142 138 134 131 128 126 125 125[tùu3] (2) 167 160 151 143 136 129 124 121 118 117 118[tùu3] (3) 157 154 148 142 135 131 127 126 124 124 124[tùu3] (4) 167 161 153 146 140 134 128 124 121 121 122[tùu3] (5) 176 164 153 145 138 132 127 124 122 122 122[tùu3] (6) 158 155 151 146 140 135 130 127 125 124 124[tùu3] (7) 166 160 153 147 142 138 135 132 128 126 125Mean (SD) 164 (6.3) 158 (3.6) 151 (2.2) 144 (2.0) 138 (2.3) 133 (2.9) 129 (3.2) 126 (3.3) 124 (3.2) 123 (2.8) 123 (2.4)Mix condition: Standard [tùu3].Time (% of tone) 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%[tùu2] (1) 160 159 157 157 160 166 173 183 194 205 208[tùu2] (2) 167 164 161 160 161 166 173 183 194 203 205[tùu2] (5) 164 161 157 154 154 156 159 165 173 182 186[tùu2] (7) 155 152 149 148 149 152 157 164 174 186 191[tùu3] (8) 158 153 147 141 135 130 126 124 122 121 121[tùu3] (9) 165 156 146 138 131 126 123 121 121 122 122[tùu3] (10) 161 157 151 146 141 135 129 126 124 123 123Deviant [tùu2]Time (% of tone) 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1F0 Turning Point (ms)[tùu2] (8) 159 151 146 144 145 149 156 166 178 191 196 15.23 61Frontiers in Psychology | www.frontiersin.org 17 April 2020 | Volume 11 | Article 646

ORIGINAL RESEARCH ARTICLE

Abstract

Extracted data

ORIGINAL RESEARCH ARTICLE

Abstract

Extracted data

Related items

Related items