Word order, naturalness & conventionalisation:
evidence from silent gesture
Myrte Vos
July 2017
A research report submitted in partial fulfillment of the MSc degree in Brain and Cognitive Sciences (University of Amsterdam). Supervised by Prof. dr. Simon Kirby and dr. Marieke
Abstract
Of the six possible ways to order Subject, Object and Verb, two - SOV and SVO - account for the constituent word order in nearly 80% of the world’s languages. Why? The pragmatic principle ’Agent first’ accounts for the dominance of S-initial word orders; and recent work in word order typology, creole languages, emerging sign languages, and improvised silent gesture suggests that SOV word order is a kind of natural, intuitive ’default’. But if that is so, why is SVO word order nearly as prominent as SOV? One improvised silent gesture study, from Schouwstra & de Swart (2014), suggests that the usage of SOV versus SVO is dynamically influenced by the semantic content of the verb. Another study, by Marno, Langus and Nespor (2015) posits that SVO is preferred by the syntax-governing ’computational system’ of cognition, and that while improvised communication favors SOV, access to a lexicon frees up cognitive resources to use syntax, and ”consequently SVO, the more efficient word order to express syntactic relations, emerges.” In their improvised silent gesture task, wherein half the participants had to improvise their gesturing of simple transitive events and the other half were first taught a gesture lexicon before being asked to communicate, participants trained on a lexicon did indeed favor SVO. I replicated Marno et al’s experiment with stimuli restricted to event-types found to elicit SOV, as well as running an adapted condition using a lexicon of randomly assigned, arbitrary gestures, to further test their hypothesis.
Contents
1 Introduction 3
2 Experiment: gesture production with improvised, taught iconic and non-iconic
gestures 6
2.1 Methods . . . 6
2.1.1 Materials . . . 6
2.1.2 Coding . . . 7
2.1.3 Condition 1: improvised gesture . . . 8
2.1.4 Condition 2: taught iconic lexicon . . . 9
2.1.5 Condition 3: taught non-iconic lexicon . . . 11
3 Results 13 3.1 Comparison: did I replicate Marno et al.’s results? . . . 13
3.2 Further analysis . . . 16
4 Discussion 17 Acknowledgements 20 5 Appendix 21 5.1 Materials from Marno, Langus & Nespor (2015) . . . 21
5.2 First languages of participants . . . 21
5.2.1 Condition 1 . . . 21
5.2.2 Condition 2 . . . 22
5.2.3 Condition 3 . . . 22
5.3 Breakdown of all used word orders . . . 23
5.4 Summary output of generalised linear mixed model . . . 24
1
Introduction
All languages have conventions for defining and disambiguating the basic units of a sentence that tell us who did what to whom: subject, verb and object (henceforth S, V and O)1. One such convention, the order in which S, V and O are arranged, is known in language typology as basic word order. Some languages, like English, have a strict word order that requires users to infer S and O from their position relative to the verb; many languages have several permissible word orders, depending on e.g. the finiteness of the verb or the focus of the sentence; and some, like Modern Greek and Hungarian, have no dominant word order. Dryer’s World Atlas of Language Structures[9] records the basic word order of 1377 languages:
SOV SVO VSO VOS OVS OSV No dom. order
565 488 95 25 11 4 189
Table 1: Global distribution of word orders
What jumps out is the prevalence of two word orders in particular: SOV and SVO. They combine for nearly 80% of the world’s languages, an asymmetry that has interesting implications for the origins and evolution of language, and for our understanding of the human mind, because it suggests that these orders have some cognitive and/or communicative advantage.
The clear preference for S-initial word order was formulated by the American linguist Joseph Greenberg (1963) as Language Universal #1: ”In declarative sentences with nominal subject and object, the dominant order is almost always one in which the subject precedes the object.”[18] Jackendoff (1999) has called this the ’Agent First’ principle[27]: a default assumption that the first ’character’ brought to mention is the agent, observed in the languages of people with a lim-ited grammar, such as deaf children of non-signing parents[15], agrammatic aphasics[43], adult second-language learners without explicit instruction[31], and speakers of pidgins[3]. The under-lying motivation may be that placing the agent at the start of the phrase is more natural and intuitive, for a variety of reasons. DeLancey (1981) argues that speakers encode the way they perceive events in their grammar: the word order of a phrase is thus a reflection of attention flow, which with a transitive event would naturally begin with the agent[8]. Tomlin (1986) suggests that semantic characteristics of the agent, such as animacy, lend it prominence in the sentence[51]; prominence, in turn, informs the speed and ease with which a phrase is analysed, with the brain areas responsible for processing this information generally preferring more prominent arguments to precede less prominent ones (cf. Bornkessel-Schlesewsky & Schlesewsky 2009[4], and Kemmerer 2012[30]). Meir et al. (2007) note that speakers of sign languages tend to position their own body as the subject, which suggests a tendency to identify with and assume the viewpoint of the agent[38]; in recent work (Meir et al. 2017[39]) they found that both hearing subjects who were asked to improvise gestural utterances and deaf speakers of young sign languages strongly favoured S-initial word order when S was human but O was not. The subject, when it realises the agent, seems therefore to be conceptually salient, and most at home at the beginning of the phrase.
The order of Verb and Object, however, presents a puzzle. V and O are thought to be adjacent in the majority of languages because ”a transitive verb and its object form a more cohesive, unified
1Technically, subject, object and verb are syntactic elements that do not necessarily always align with the thematic roles agent, patient and action. For instance, in the passive construction the buffalo is chased by the lion, the subject is the buffalo, but the agent is the lion. This paper uses the nomenclature S, O and V to denote agent, patient and action respectively.
syntactic and semantic whole than do a transitive verb and its subject” (Tomlin 1986:74), but is the subsequent ordering of V and O then a matter of an evolutionary coin-toss, or are there functional, cognitive and socio-cultural factors driving a grammar to one or the other?
Studies on how languages change over time (diachronic variation) suggest that historically, the proportion of SOV languages was even greater, and that the earliest human languages may have had SOV word order(Newmeyer 2000[40]; Maurits & Griffiths 2014[36]); they also demonstrate that languages evolve unidirectionally away from SOV and towards SVO (Gell-Mann & Ruhlen, 2011[13]; Pagel 2009[41]). Another fascinating source of insight comes from gesture: in homesigns, the primitive gestural systems created by deaf children of hearing parents, the orders used (SV and OV) are consistent with SOV order, an observation that is robust cross-culturally (Goldin-Meadow et al. 1998[16]). In emerging sign languages such Nicaraguan Sign Language (Coppola et al. 1997[6]), Al-Sayyid Bedouin Sign Language (Sandler et al. 2005[47]), and the very young ’Tzotzil’ sign language in Mexico(Haviland 2011[23]), the preferred word order is SOV, despite the word orders of the ambient spoken languages being SVO or (in Tzotzil) VOS. A preference for SOV order has also been observed in psychological experiments, where hearing participants asked to describe simple transitive events using improvised gesture predominantly use SOV - regardless of the word order of their first language (Goldin-Meadow et al. 2008[17]; Gibson et al. 2013[14]; Langus & Nespor 2010[32]).
Arguments for the advantage of SOV order also come from functional and cognitive linguistics. Because verbs are conceptually more complex and semantically harder to access than nouns[17], it makes sense to place the verb last in the sentence(Hoeks 2016, unpublished[25]); this also makes them more predictable, because the preceding constituents have already provided information about what comes last(Ferrer-i-Cancho 2014[10]). Several authors have proposed that structural iconicity might inform word order, i.e. that the temporal, spatial, and conceptual ordering of an event is reflected in the word order of the sentence that describes it (Haiman 1980[19]; Croft 2002[7]; Conradie 2001[5]), particularly before a conventional grammar is in place. In an improvised-gesture task where participants were asked to describe extensional events (involving some kind of action through space) and intensional events (where the object is non-existent or non-specific, e.g. pirate dreams of sock ), SOV was predominantly used for extensional events, but for intensional events, an SVO preference emerged (Schouwstra & de Swart 2014[49]). In subsequent work, Schouwstra et al. (2016[48]) have demonstrated that, as the budding communication systems formed in improvisedgesture situations become more regular and conventional, they converge on a single word order -which order depends on the circumstances.
What circumstances, then, trigger a shift in word order preference to SVO? Several studies on improvised gesture (also called elicited pantomime by some authors) have investigated this question. Gibson et al. (2013[14]) found that participants used SVO instead of SOV to gesture events where the agent and patient of an event were plausibly reversible (e.g. fireman kicks girl ), and events where the object was an embedded clause (e.g. a man says that a woman pushes a boy). Their explanation, which they call the noisy-channel hypothesis, is that language users are sensitive to the possibility of noise corrupting the signal, and will try to present information as clearly as possible. In the absence of a case-marking system, SVO makes it easier to assume Agent-First and infer that the patient follows the verb. Hall et al. (2013[21], 2014[20] replicate Gibson et al.’s results, but argue that the reversibility of agent and patient is expressed as a constraint on production: because signers often impersonate the roles of agent and patient with their own body, confusion may arise when they are adjacent.
Finally, there is evidence that SVO may be more efficient in terms of online memory cost and processing effort (Hawkins 2004[24], Ferrer-i-Cancho 2014[10]). A study by Futrell et al. (2015[12]) on dependency length minimisation seems to support this idea.
The research cited above combines to paint a picture of SOV as the primordial word order, a ’natural’ grammatical default determined by functional and cognitive processing preferences, which diversifies and evolves into other word orders under pressure of any number of other biases and socio-cultural influences. However, there exist other perspectives. Langus and Nespor[32] posit that SVO, not SOV, is the natural default, departing from the assumptions that a) the human language faculty is modular and distributed across several specialised cognitive systems[11]; b) one of these systems, the so-called ’computational system’, underlies the generation of grammatical structures[22]; and c) ”...there is a universal underlying structure [the specifier-head-complement configuration (cf. Kayne (1994)[29]] from which the surface syntactic forms of all languages are derived”[32], for which the corresponding word order is SVO. Combining this perspective with the observation, from Hudson & Eigsti (2003[26]), that speakers of pidgins only begin to use consistent word order after achieving a certain amount of lexical competence, Marno et al. (2015[35]) hypo-thesised that using a consistent lexicon frees up cognitive resources, which effects a switch from the ’conceptual-motor’ system (which they say underlies basic, pre-syntactic communication) to the ’computational’ system, with a resultant switch in word order preference from SOV to SVO. In a gestural description task, wherein half the participants had to improvise their gesturing of simple transitive events and the other half were first taught a gesture lexicon, participants that had to improvise used SOV, and participants trained on a lexicon did indeed strongly favour SVO - regardless of whether their first spoken language had SVO or SOV order.
Because this was such a significant and surprising result, and because I was skeptical about the explanation given for why acquiring a lexicon should influence word order preference so strongly, I replicated Langus & Nespor’s and Marno et al.’s improvised-gesture and taught-lexicon exper-iments. I also did a third experiment in which participants were taught a gesture lexicon that was non-iconic: that is, the gestures were designed not to resemble their referents in any obvious, intuitive way. My hypothesis was that if learning a lexicon did lead to an increased preference for SVO, a lexicon with non-iconic gestures should be even further removed from the semantically-conditioned gesture strings produced during improvisation, and should therefore lead to an even stronger SVO preference.
2
Experiment: gesture production with improvised, taught
iconic and non-iconic gestures
I sought to replicate a) Goldin-Meadow’s finding that using improvised gesture to describe simple transitive events results in predominant use of SOV word order, regardless of the word order of the gesturer’s first language; and b) Marno et al.’s finding that learning and using a gesture lexicon to describe these events results in a predominant use of SVO word order. As such, I endeavored to remain as faithful to Marno et al.’s methodology as possible, with these exceptions:
• Materials. The materials used by Marno et al. (see Figure 10 in the Appendix) were, in my view, conceptually too confusing in a number of ways. E.g. in half the stimuli, the agent and the patient are not only both animate, but it is uncertain whether the agent is human or animal; the verb ’feed’ involves an instrument (a fluid is poured from a jug) and an object (a dish) competing for patient-hood; and the patients are nearly indistinguishable. The choice was made to create a new set of stimuli that precluded as many of these potential confounds as possible.
• Instruction. Both in Marno et al.(2015) and in Langus & Nespor (2010), which presented the improvised gesture experiment with which Marno et al. compare their ’taught lexicon’ experiment, participants were instructed to use exactly three gestures to describe the stim-uli.2 I chose not to include this instruction in any of my experiments, in case it might bias participants to structure their mental representation of the stimuli in terms of a known grammar.
• First language. The experiments in Marno et al. (2015) and Langus & Nespor (2010) were run on two groups: one whose first language had SVO order (Italian), and one whose first language had SOV order (Turkish/Farsi). In my experiments, participants were not selected or excluded for any particular first language, the rationale being that based on earlier results[17][32], first language should have no significant influence on the word order used in improvised gesture. The benefit of that approach is that any strong tendency to use a certain word order will be more robust from a cross-linguistic perspective; the downside is that if participants’ first language is not controlled for, it is impossible to discern how much variance in the data derives from that.3 However, because Marno et al. found no significant interaction between Instruction (improvised gesture vs. taught lexicon) and Language Group in their experiment, I hypothesised that I should be able to replicate their results with participants of any language background.
2.1
Methods
2.1.1 Materials
Stimuli were 16 pictures, drawn by the experimenter, depicting simple transitive events involving combinations of four agents (artist, pirate, viking, and nun), four patients (ball, guitar, mitten, and
2Interestingly, in Langus & Nespor’s improvised gesture experiment, about 40% of all productions consisted of 2-gesture strings, despite explicit instructions to use 3 gestures.
3By coincidence, most participants (across conditions) were speakers of an SVO language. A small number spoke a language with no dominant word order; 3 spoke a V2-SOV language (German and Luxembourgish) and only 2 spoke a true SOV language (Hindi). It must be noted, however, that many of the represented languages have topic-prominence (e.g. Mandarin, Hungarian, Vietnamese and Indonesian): this can affect word order, in ways that are difficult to anticipate or account for. One anecdotal example is a Mandarin-speaking participant in Condition 2, who consistently a) omitted S, and b) conditioned the order of V and O on the verb (VO for throw and drop, OV for hang and cut ) - they unfortunately did not say why they did so.
scarf), and four actions (throw, drop, cut, and hang [on a clothesline]). Elements were combined in such a way that no event could be predicted from seeing only some of its constituents: that is, when seeing the events gestured to them, an observer who was familiar with the stimuli should not be able to ’finish the sentence’.
Figure 1: Stimuli. The 16 pictures that participants were asked to describe in gesture
Per Goldin-Meadow et al.’s (2008[17]), and subsequently Schouwstra & de Swart’s (2014[49]), observation that extensional events drive a preference for SOV word order, all verbs were exten-sional, so that an observed preference for SVO word order could not be attributed to any semantic pressures. Verbs and Objects were chosen and designed such that no verb’s gestural form would need to be conditioned on the object.
Stimuli were presented on Macbook sitting on the desk in front of the participant; their upper body was recorded via a Logitech C930e HD webcam, with the video software VideoBox (Kirby, 2014). The experiment was built and executed using PsychoPy (Peirce, 2007[42]).
2.1.2 Coding
Participants’ gestures were transcribed by the experimenter, who identified each gesture’s referent as subject, verb, or object. Multiple consecutive gestures for the same referent were considered to belong to the same constituent. For example, the string “BRAIDS HELMET HORNS BEARD SCARF SCISSORS-CUT would be coded as SOV (viking, scarf, cut). However, if multiple gestures for the same referent were non-consecutive, a repetition would be noted. For example, “HANG-UP SCARF HORNS BRAIDS CLOTHESLINE HANG-UP” be coded as VOSV (hang, scarf, viking,
hang). In the absence of any principled way to determine phrase or utterance boundaries in impro-vised gesture, everything that the participant produced was counted as a single utterance, unless there was a lengthy pause (about ¿2 seconds). In these cases, the most complete string was coded. Broken off or unfinished gestures were ignored if the participant treated them as mistakes (e.g. by flapping their hands, shaking their head, laughing or frowning); in such cases, the gesture made in correction was coded. Gestures referring to things other than the twelve main constituents were ignored in determining constituent order; e.g. ”CLOTHESLINE”.
2.1.3 Condition 1: improvised gesture
Participants
n=22 (4 male, 18 female; mean age 23.8). Participants with no knowledge of any sign language were recruited via the University of Edinburgh careers service (MyCareerHub), and paid £5. For a list of participants’ first languages, see Table 4 in the Appendix. The experiment was approved by the Psychology Research Ethics Committee of the School of Philosophy, Psychology and Language Sciences of the University of Edinburgh. Participants read and signed an information & consent form.
Design and procedure
Participants were seated across from the experimenter with a laptop on the table in front of them; the experimenter could not see what was on the screen. After receiving written instructions, participants were first given brief exposure to the stimuli. In order to keep the procedures of Conditions 1 and 2 as similar as possible (see section 2.1.4), this was done by presenting the stimuli in 12 different pairs: each pair had one common constituent, so e.g. for ’hang’, VIKING-HANG-MITTEN was presented together with NUN-HANG-SCARF. In each pair, the stimuli were presented simultaneously and side-by-side for 3 seconds; to control for left-to-right processing effects, the order of the stimuli in each pair was randomized. For the same reason, for half the participants, all stimuli were presented in mirror image.
After exposure came the test phase: participants were shown each individual stimulus, in random order, and were instructed to describe the event in the picture to the experimenter, using only gesture and no speech. If a participant required additional instruction regarding the level of detail they should aim for in their description, which usually became obvious within the first three trials, they were told to imagine the experimenter had all sixteen pictures laid out in front of them, and that from the participants’ description they should be able to pick out the right one. Participants could use as much time and as many gestures as they liked; after finishing gesturing a picture, a keyboard press presented the next one.
Responses
Because of the freedom participants had in this experimental condition, a wide variety of responses was recorded: most of these variations were only used once or twice (for a full breakdown, see Figure 11 in the Appendix), so for the purposes of data inspection and analysis, responses were binned into four categories: 3-gesture strings with SOV and SVO order; 3-gesture strings with the order VSO, VOS, OSV or OVS (Other); and all other responses, which were either missing 1 or 2 constituents, or repeated them so as to make their basic word order ambiguous, and were therefore considered Invalid. It is worth noting that, of the 104 responses categorised as Invalid, 52 are ’OV’, and of the 17 responses categorised as Other, 14 are ’OVS’. A first inspection of the data therefore confirms a strong tendency to place Object before Verb during improvised gesture.
Figure 2: By-participant breakdown of used word orders.
2.1.4 Condition 2: taught iconic lexicon
Participants
n=22 (7 male, 15 female; mean age 21.6). For a list of participants’ first languages, see Table 5 (Ap-pendix). The procedure and selection criteria for recruiting the participants were as in Condition 1, but because Condition 2 took longer to complete participants were paid £7.
Materials
As in Condition 1, with the addition of 12 looping movies of about 2 seconds, showing a young woman making a gesture referring to one of the 12 constituents in the gesture vocabulary set. All gestures were derived from the most common improvised gestures from Condition 1, to ensure that they were intuitive and easy to remember. The woman was recorded in the same room where most participants did the experiment, and positioned in the same manner: seated behind a desk, against a white backdrop. Permission was obtained to re-use the movies in academic presentations and future experiments.
Design and procedure
In this condition, participants were given a lexicon of 12 gestures to memorise prior to having to describe the stimuli. To ensure that they could use the lexicon with confidence, they received two rounds of teaching and practice. In each teaching trial, participants saw two pictures side-by-side on the upper half of their computer screen, and a looping movie on the bottom half of the screen. The two pictures always had only one element in common, depicting either (a) the same agent per-forming different actions with different patients; (b) different agents perper-forming the same action, with different patients, or (c) different agents, performing different actions, with the same patient. The movie at the bottom of the screen showed a person demonstrating the gesture referring to the common element of the pictures. This approach, used first by Marno et al. (2015), ensured that at no point during the teaching phase was there any reference to the categories (Subject, Object, and Verb) of the gestures.
Figure 3: Procedure for teaching the gesture lexicon. The black arrows in the video still signify continuous motion of the hand from right to left, and multiple ’scissoring’ movements of the index and middle fingers.
The teaching phase was self-paced, each trial lasting until the participant pressed the space bar to move on to the next trial. Each trial was presented twice. Following the teaching round of 24 trials total, participants were informed they were given the chance to practice, and presented with the same picture-pairs from the teaching phase, but without the accompanying movies. Par-ticipants were instructed to make the gesture referring to the common element of the picture-pair, and to press the space bar to continue to the next trial. Each picture-pair was presented just once. The teaching and practice trials were presented in a random order and, as in Condition 1, the left-right positioning of the picture pairs was randomised and half of the participants saw a mirrored version of the stimuli.
After the two rounds of teaching and practice came the test phase: participants were shown each individual picture in random order, and were instructed to describe the event in the picture to the experimenter using only gesture and no speech. Participants were not restricted to using only the taught gestures, though in practice they all did.
Responses
For a full breakdown of used word orders in Condition 2, see Figure 12 in the Appendix. Looking at the by-participant proportions of word orders (Figure 4), it seems that in this experiment, Invalid responses have decreased (27 in Condition 2, versus 104 in Condition 1), but variability in word order has increased, not only across but also within participants.4 Though use of SVO has
4It would have been interesting to quantify this variability with a measure of entropy, but time did not allow for that to be included in this report.
Figure 4: By-participant breakdown of used word orders
increased dramatically, it is not immediately clear whether it has replaced SOV as the preferred word order.
2.1.5 Condition 3: taught non-iconic lexicon
Participants
n=12 (2 male, 10 female; mean age 20.9). For a list of first languages of participants, see Table 6 (Appendix). The procedure and selection criteria for recruiting the participants was as in Condition 1 and 2; they were paid £7, as in Condition 2.
Materials
As in Condition 2, but with a different set of filmed gestures. In this condition, participants were taught a lexicon of ’arbitrary’ gestures, wherein gestures were randomly assigned to constituents for each participant, and the gestures were designed to bear no obvious or deliberate iconic resemblance to any of their possible referents. (Of course, this could not prevent participants from seeing any such resemblance, or from interpreting the gesture form in some idiosyncratic way to make it easier to memorise.)
Figure 5: Example of the difference between an ’iconic’ and ’non-iconic’ gesture for the constituent ’Artist’. The iconic gesture (middle image) represents the artists’ bowtie.
Design and procedure
As in Condition 2. Three participants received an extra round of teaching and practice, because they had indicated that they were not yet confident that they had memorised all the gesture-referent pairs correctly.
6 participants did not succeed in memorising the gesture vocabulary set5. This was determined
by checking the log file of the experiment to see the (randomly assigned) gesture-referent pairs for that participant, as well as the (randomly generated) order in which the stimuli were presented in the test phase, and comparing that to the productions of the participant.
Responses
For a full breakdown of used word orders in Condition 3, see Figure 13 in the Appendix. In this experiment, finally, we see that a clear majority of the responses have SVO word order.
Figure 6: By-participant breakdown of word orders
3
Results
All data processing and analysis was done using R and the RStudio environment[44]. Plots were generated using the R package ggplot2[53].
Figure 7 presents the mean word order proportions in all three experimental conditions, with error bars representing the standard error of the mean. Descriptive statistics are listed in Table 2.
Figure 7: Word order proportions in all 3 conditions
3.1
Comparison: did I replicate Marno et al.’s results?
To recapitulate, below are the graphs for Marno et al.’s results:
(a) Results lexicon experiment (b) Meta-analysis improv & lexicon experiment
Figure 8: Figures reproduced from Marno et al.(2015)
Marno et al.(2015) compared their results from the taught lexicon experiment (Figure 8a) to the results of an earlier improvised gesture experiment (Langus & Nespor 2010) done with Italian and Turkish speakers. The stimuli for the 2015 experiment comprised half of the set of stimuli used in 2010; I assume that the improvised-gesture results in Figure 8b represent only
Method Word order Mean SD SEM improv SOV 0.63 0.37 0.07 SVO 0.02 0.1 0.02 Other 0.04 0.11 0.02 Invalid 0.29 0.34 0.07 taught SOV 0.46 0.42 0.09 SVO 0.35 0.4 0.08 Other 0.1 0.25 0.05 Invalid 0.07 0.22 0.04 arbit SOV 0.23 0.3 0.08 SVO 0.64 0.4 0.11 Other 0.08 0.19 0.05 Invalid 0.03 0.1 0.03
Table 2: Descriptive statistics of word order distributions
the responses to those stimuli, though this is not explicitly stated. The bars in Figure 8b do not appear to sum to 100, suggesting that per Condition and Language Group, a small and invisible number of data points (about 10%, at a guess) consists of what I have termed Other and Invalid responses. No exact proportions were given for the improvised-gesture data in their Results section, so approximations of the mean and the standard error were extracted using WebPlotDigitizer[45]. This enabled a comparison of Marno et al.’s results6 to my results from Conditions 1 and 2:
(a) Marno et al.(2015) (b) Vos et al.(2017)
Figure 9: Replication comparison
6In Figure 9a, the abbreviations appended to the method names are It (Italian), Tur (Turkish), and Per (Persian). Turkish and Persian are both SOV languages; Italian is an SVO language.
To test their data for statistical significance, Marno et al.(2015) discarded all responses that were not SOV or SVO, and reproportioned the remaining responses to sum to 1. (Note that the numbers represented in Figure 8 are therefore not the same numbers that were used to test for significance.) They then used two Two-Way ANOVAs, with between-subjects factors Instruction and Language Group and absolute proportions of SOV and SVO as the dependent variables, to find significant p-values (¡0.05) for the effect of Instruction on the use of both orders, and for the effect of Language Group on the use of both orders, but not for the interaction between Instruction and Language Group. This suggests that though participants’ first language does still influence word order choice, the Instruction does still skew word order preference one way or the other, and does so regardless of first language.
Via direct pairwise comparisons, Marno et al. were able to conclude that both SVO and SOV speakers were significantly more likely to use SOV order in the improvised-gesture experiment, and also significantly more likely to use SVO order in the taught-lexicon experiment.
In the spirit of a replication, I initially followed the same approach in analysing my data. Re-sponses were reformulated as binary response values, representing a success or failure to produce SVO order as TRUE or FALSE. Statistical analysis was carried out using a one-way ANOVA with a 3-level group factor to compare mean absolute proportions of SOV across the three conditions. Bartlett’s test revealed that the data for SVO did not have homogeneity of variance (which violates ANOVA’s assumption of homoscedasticity), so a nonparametric test (Kruskal-Wallis) and post hoc (Dunn’s test) were used to analyse it.
SOV∼Method SVO∼Method F -value/χ2 F (2,53)=4.291 χ2= 23.595
p-value p <0.0187 p <0.000... improv-taught p <0.307 p <0.0015
taught-arbit p <0.226 p <0.03-0.06 improv-arbit p <0.014 p <0.001
Table 3: Significance values
Of interest is the asymmetry in significance of the decreasing use of SOV on the hand, and increased use of SVO on the other. The increase in SVO between improv and taught is clearly significant, but the increase between taught and arbit only barely so. (The p-value is given as a number between 0.03-0.06 because I ran Dunn’s test twice, using different packages but - ostensibly - the same parameters.) The decrease of SOV, from improv to taught and from taught to arbit, is never significant.
My improvised-gesture experiment confirms the already quite robust finding that, when asked to invent gesture utterances from scratch to describe simple transitive events, participants prefer to use SOV order, regardless of the word order of their first language. However, my taught-lexicon experiment has not completely replicated Marno et al.’s results. Though use of SVO saw a significant increase, there was no similar decrease in use of SOV, and the difference between the two proportions is not large enough to be able to conclude that acquiring a lexicon prompts participants to switch their word order preference over to SVO.
3.2
Further analysis
Most authors investigating improvised gesture/elicited pantomime([17][32][21][49][20][14][35]) use ANOVAs to analyse their data. There are several problems with that from the perspective of the-oretical statistics (cf. Jaeger 2008[28] for an introductory discussion), chief among which is that the ANOVA is meant to test the means of continuous, not categorical, data. Shoehorning categorical data into a pseudocontinuous format, though not necessarily always wildly misguided, can lead to spurious results. I therefore chose to also run a binomial generalised linear mixed model, using the lme4 package[2], to explore the effect of method on SVO word order (a dummy variable wherein ’SVO’ responses are coded as 1, and all others as 0). Fixed effects were the method of instruction (improv, taught, and arbit ), and random effects were subject and stimulus.7 The advantage of this
approach is that it can model the effect of method of instruction, while simultaneously accounting for individual variation associated with each subject and each stimulus.[54]
The taught data was set as the model intercept (β = −1.142, SE = 1.301). The difference in SVO usage between the taught and improv conditions is clearly significant, and that is borne out by this model as well (β = −7.697, SE = 2.165, p¡0.01). The difference in SVO usage between the taught and arbit conditions does not reach standard significance level (β = 3.84, SE = 2.286, p = 0.092).
Keeping in mind that in Condition 3, n=128, we can conclude that:
• There is a statistically significant shift in word order distribution from predominantly SOV in situations where gesture has to be improvised, to SVO when a non-iconic gesture lexicon must be used.
• There is a statistically significant increase in SVO use when an iconic lexicon must be used, compared to improvised gesture; but there is no concurrent significant decrease in the use of other word orders.
• There is a notable but not statistically significant increase in SVO use when the gesture lexicon is non-iconic, compared to when it is iconic. Concurrently, there is a notable but not statistically significant decrease in SOV use.
7The formula used was glmer(SVO method + (1|subject) + (1|stimulus)| 8:(
4
Discussion
Based on the results of my first two experiments, I am not convinced that the learning and use of a lexicon effects a switch from SOV to SVO order. Though the increase of SVO use during Condition 2 was significant, there was only a slight decrease in the use of SOV, and more use of other word orders, besides. In addition, word order variability within subjects seemed to be greater in Condition 2, though this suspicion has not been tested. Crucially, however, Marno et al.’s hypothesised and observed increase in SVO word order with the use of a gesture lexicon was replicated. Does this mean Marno et al.’s explanation, which is that learning a lexicon frees up cognitive resources to engage the SVO-preferring computational system, is also valid?
To recapitulate Marno et al.’s theoretical point of departure: they assume that a) the human language faculty is modular and distributed across several specialised cognitive systems[11]; b) one of these systems, the so-called ’computational system’, underlies the generation of grammatical structures[22]; and c) ”...there is a universal underlying structure [the specifier-head-complement configuration (cf. Kayne (1994)[29]] from which the surface syntactic forms of all languages are derived”[32], for which the corresponding word order is SVO.
The scope of this report does not permit a full discussion of the complaints that can brought against the theory of Universal Grammar in general and Chomsky’s Minimalist Program in par-ticular, or the claim that language in the human brain is subserved by multiple distinct cognitive systems that have their own preferred word order and compete with each other for dominance. It must suffice to say that there is accumulating cross-disciplinary evidence (e.g. from behavioural neuroscience[1] and cognitive neuroscience[30][4][52]) that the ’conceptual-motor system’ and the ’computational system’ cited by Marno et al. do not exist or operate as distinct cognitive modules; even if they did, there appears to be no neurocognitively grounded evidence that the computational system imposes SVO word order.
What about linguistic evidence? Marno, Langus & Nespor’s motivation for their experiment hinges on two ideas: the idea that lexical competence drives grammaticalisation (per Hudson & Eigsti 2003[26], who studied the emergence of grammar in pidgins9, and Hall et al. 2014[20], who in
their improvised-gesture study found an increased use of SVO when participants were instructed to teach their gestures to the experimenter), and the idea that this, in turn should yield a preference for ”syntactically unmarked SVO word order”(per Kayne 1994[29]). Does the literature support this idea in any way?
Though McWhorter (2001[37]) makes the case that Creole grammars are less complex than others, and observes that Creole languages almost always have SVO word order, Newmeyer (2000:380[40]) counters the thought that this should mean that SVO languages are therefore more basic: ”Creoles, by definition, derive from pidgins, and pidgins by definition are created by people who already have a language (indeed, different languages). In such a contact situation, inflected morphology would be expected to be the first grammatical feature to disappear. And there are obvious functional reasons [...] why SVO is a better word order than SOV when thematic roles are not overtly marked.” Indeed, Sinnem¨aki (2010[50]) found that word order in zero-marking languages is almost always SVO (noting also that zero-marking languages are typologically quite rare), and argues that this is because a) thematic role assignment is more economical in SVO, b) cognitive processing is more efficient due to shorter dependency lengths ([24][12]), and c) iconicity
9It is interesting to note that Hudson and Eigsti themselves considered, and then dismissed, the possibility that the predominance of SVO order in their data and in creoles more generally might be due to an underlying universal preference for SVO; stating that this position ”seems unnecessarily extreme”, they instead attributed this pattern to ”a tendency toward SVO in new, morphologically poor languages”, and to the fact that in many known creoles the super- and substrate languages have SVO order.
in the thematic or temporal sequencing of events[19]. Note that these arguments largely overlap with the explanations for SVO use put forward by other improvised-gesture studies (Gibson, Hall, Schouwstra). There also seems to be no indication that SVO order is more ’learnable’ than SOV: a connectionist modeling study by Lupyan and Christiansen (2002[33] suggests that it makes little to no difference to the learnability of a language whether it has strict (as opposed to flexible) word order, or a case system. In an artificial sign language study (de Sam Lazaro, 2014, unpub-lished[46]), participants trained on an artificial sign language with SOV or SVO order learned the language equally well regardless of their own first language. On the other hand, a recent, inter-esting avenue of research concerns the influence of socio-cultural circumstances on the structure of language: Wray and Grace (2007[55]) and Lupyan and Dale (2010[34]) propose that factors that make a language more ’outward-facing’ (e.g. a large number of speakers, wide geographical range, or many second-language learners) drive it towards evolving simpler inflectional morphology. This would, as discussed earlier, weight certain functional principles (processing efficiency, avoidance of confusability) more heavily, and perhaps lend greater advantage to SVO. There is therefore ample explanation for why SVO might emerge in a language whose word order started with SOV, without needing to appeal to proposed but unsubstantiated theories about cognitive systems struggling to impose their preferred approach to hierarchically ordering information.
What, then, can explain the results of my research? Given that the stimuli were designed to be straightforward (no embedded clauses), extensional (provoking an iconic bias for SOV), and non-reversible; and given that the gesture vocabulary set for Condition 2 was derived from the most commonly used gestures from Condition 1, to ensure maximum naturalness, what can explain the difference in word order distributions between Conditions 1 and 2, and Conditions 2 and 3?
Unlike with Marno et al., the word order distribution for my taught-lexicon experiment did not show a clear predominance of either SOV or SVO, and the within-subject variability of word orders was greater than in the improvised-gesture experiment. Rather than frame the emergence and conventionalisation of word order in these kind of proto-language situations as a struggle between or a transformation from SOV to (only) SVO, it may be more useful to view it as a process that, initially, has SOV as a strong default, but opens up to more general word order variability as the language becomes more regular, and eventually converges on a dominant word order that, depending on the circumstances, may or may not be SVO. Hall et al. (2013) also found that when participants used non-SOV word orders to gesture reversible events, SVO accounted for only 18-38% of the responses: the rest were other word orders, most of which are rare in natural language, and would therefore probably quickly fall in disuse as the gesture system regularised and evolved. There may be something about the isolation and lexicalisation of elements in an event that affects the conceptual understanding and processing of that event, which allows the language user to analyse and arrange those elements as she sees fit, before functional and communication constraints drive her (and her interlocutor) to converge on a fixed order.
The small n for the non-iconic lexicon experiment makes it difficult to theorise or draw any firm conclusions, but the high proportion of SVO in the responses is an intriguing result. Be-cause participants were only trying to correctly reproduce the gestures they had learned, rather than communicate the events effectively to someone else, it seems unlikely that functional con-siderations (e.g. communication over a noisy channel, body-as-subject) could drive the effect. It is possible that, in using a non-iconic gesture lexicon, users become so far removed from the natural, spontaneously-invented communication of the improvised-gesture paradigm that all the iconicity-driven word order biases become obsolete, and the user must fall back on something else: be it a cognitive predisposition to use specifier-head-complement structure[29], reduce dependency length[12] - or minimise online memory load[10]. All participants in Condition 3 had SVO as the order of their first language. Adults learning a new language typically do not abandon their first grammar, at least initially - and feedback given by participants after they had completed the
experiment indicated that many had found it difficult to memorise and correctly reproduce the gestures. It cannot be discounted that the predominance of SVO order was due to participants falling back on the grammar of their first language.
To conclude, we are beginning to understand the many functional, cognitive, and socio-cultural factors influencing word order preference in the earliest infancy of a language (through the elicited pantomime paradigm and other experimental methods), and in its adolescence (through the study of pidgins, creoles, and emerging sign languages), but the results from Marno et al. and from this paper suggest that there is a phase in the emergence of language structure that has so far been overlooked: a transient and most likely brief period in which processes like lexicalisation can have dramatic effects on the direction a fledgling grammar takes on its way to becoming a full-grown, conventionalised syntax.
Acknowledgements
I would like to thank my supervisors in Edinburgh, Simon Kirby and Marieke Schouwstra, for their unwavering support and enthusiasm for seeing this project through despite the many hurdles and setbacks it encountered. Others in Scotland that have been exceptionally kind and helpful include Maggie Hunt (and the entire Edinburgh University Chamber Choir), Yasamin Motamedi, the other members of Thursday-morning PhD supervision, and Miranda Alcock, the landlady who became a sort of ersatz godmother and whose friendship I was and am very grateful for.
The past two years have not been easy. I owe this finished product and my overall well-being and happiness to my friends and family, and probably above all: Jelle van Aanholt.
I also want to thank Sufjan Stevens, Bryce Dessner, Nico Muhly and James McAlister for cre-ating Planetarium, and Alt-J for crecre-ating This Is All Yours: every project has a soundtrack, and for this one, these albums are it.
A final thank-you goes out to the Pittsburgh Penguins, whose back-to-back Stanley Cup Cham-pionships have served both to entertain and inspire me, and also to put my suffering in some perspective. However difficult I may have found this work at times, it has never caused me to lose any teeth.
5
Appendix
5.1
Materials from Marno, Langus & Nespor (2015)
Figure 10: Stimuli used by Marno et al.(2015)
5.2
First languages of participants
5.2.1 Condition 1
First language # Word order
English 9 SVO Chinese/Mandarin 3 SVO Cantonese 2 SVO Romanian 2 SVO Slovenian 1 SVO Italian 1 SVO
Bahasa Indonesian 1 SVO
Hungarian 1 No dom. order
Greek 1 No dom. order
5.2.2 Condition 2
First language # Word order
English 8 SVO Mandarin 1 SVO Cantonese 1 SVO Hindi 2 SOV Luxembourgish 1 V2-SOV German 2 V2-SOV Polish 1 SVO Russian 1 SVO Lithuanian 2 SVO Spanish 2 SVO French 1 SVO
Hungarian 1 No dom. order
Vietnamese 2 SVO
Table 5: First languages of participants (Condition 2). Some participants were counted twice due to bilingualism.
5.2.3 Condition 3
First language # Word order
English 8 SVO Mandarin 1 SVO Polish 2 SVO Russian 1 SVO Italian 1 SVO Vietnamese 1 SVO
Table 6: First languages of participants (Condition 3). Some participants were counted twice due to bilingualism.
5.3
Breakdown of all used word orders
Figure 11: All word orders used in Condition 1
Figure 12: All word orders used in Condition 2
References
[1] Georgios PD Argyropoulos. ‘The cerebellum, internal models and prediction in ‘non-motor’aspects of language: a critical review’. In: Brain and language 161 (2016), pp. 4–17.
[2] Douglas Bates et al. ‘Fitting Linear Mixed-Effects Models Using lme4’. In: Journal of Stat-istical Software 67.1 (2015), pp. 1–48. doi: 10.18637/jss.v067.i01.
[3] Derek Bickerton. Roots of language. Karoma, 1981, pp. 268–284.
[4] Ina Bornkessel-Schlesewsky and Matthias Schlesewsky. ‘The role of prominence information in the real-time comprehension of transitive constructions: a cross-linguistic approach’. In: Language and Linguistics Compass 3.1 (2009), pp. 19–58.
[5] C Jac Conradie. ‘Structural iconicity’. In: The Motivated Sign: Iconicity in Language and Literature 2 2 (2001), p. 229.
[6] Marie Coppola et al. ‘The emergence of grammar: Evidence from family-based sign systems in Nicaragua’. In: 22nd Annual Boston University Conference on Language Development, Boston, MA. 1997.
[7] William Croft. Typology and universals. Cambridge University Press, 2002.
[8] Scott DeLancey. ‘An interpretation of split ergativity and related patterns’. In: Language (1981), pp. 626–657.
[9] Matthew S. Dryer. ‘Order of Subject, Object and Verb’. In: The World Atlas of Language Structures Online. Ed. by Matthew S. Dryer and Martin Haspelmath. Leipzig: Max Planck Institute for Evolutionary Anthropology, 2013. url: http://wals.info/chapter/81. [10] Ramon Ferrer-i-Cancho. ‘Why might SOV be initially preferred and then lost or recovered? A
theoretical framework’. In: The Evolution of Language-Proceedings of the 10th International Conference (EVOLANG10). 2014, pp. 66–73.
[11] Jerry A Fodor. The modularity of mind: An essay on faculty psychology. MIT press, 1983. [12] Richard Futrell, Kyle Mahowald and Edward Gibson. ‘Large-scale evidence of dependency
length minimization in 37 languages’. In: Proceedings of the National Academy of Sciences 112.33 (2015), pp. 10336–10341.
[13] Murray Gell-Mann and Merritt Ruhlen. ‘The origin and evolution of word order’. In: Pro-ceedings of the National Academy of Sciences 108.42 (2011), pp. 17290–17295.
[14] Edward Gibson et al. ‘A noisy-channel account of crosslinguistic word-order variation’. In: Psychological science 24.7 (2013), pp. 1079–1088.
[15] Susan Goldin-Meadow and Carolyn Mylander. ‘Beyond the input given: The child’s role in the acquisition of language’. In: Language (1990), pp. 323–355.
[16] Susan Goldin-Meadow and Carolyn Mylander. ‘Spontaneous sign systems created by deaf children in two cultures’. In: Nature 391.6664 (1998), pp. 279–281.
[17] Susan Goldin-Meadow et al. ‘The natural order of events: How speakers of different languages represent events nonverbally’. In: Proceedings of the National Academy of Sciences 105.27 (2008), pp. 9163–9168.
[18] Joseph H Greenberg. Some universals of grammar with particular reference to the order of meaningful elements. Universals of language, ed. by Joseph H. Greenberg, 73-113. 1963. [19] John Haiman. ‘The iconicity of grammar: isomorphism and motivation’. In: Language (1980),
[20] Matthew L Hall, Victor S Ferreira and Rachel I Mayberry. ‘Investigating constituent order change with elicited pantomime: A functional account of SVO emergence’. In: Cognitive science 38.5 (2014), pp. 943–972.
[21] Matthew L Hall, Rachel I Mayberry and Victor S Ferreira. ‘Cognitive constraints on con-stituent order: Evidence from elicited pantomime’. In: Cognition 129.1 (2013), pp. 1–17. [22] Marc D Hauser, Noam Chomsky and W Tecumseh Fitch. ‘The faculty of language: what is
it, who has it, and how did it evolve?’ In: science 298.5598 (2002), pp. 1569–1579.
[23] John B Haviland. ‘Nouns, verbs, and constituents in an emerging ‘Tzotzil’sign language’. In: Representing language: Essays in honor of Judith Aissen (2011), pp. 157–171.
[24] John A Hawkins. Efficiency and complexity in grammars. Oxford University Press on De-mand, 2004.
[25] Morwenna Hoeks. ‘From SOV towards SVO: Explaining the word order distribution in terms of changing preferences’. Master’s thesis. Radboud University Nijmegen, 2016.
[26] Carla L Hudson and Inge-Marie Eigsti. ‘The Lexical Competence Hypothesis: A cognitive account of the relationship between vernacularization and grammatical expansion in creoliz-ation’. In: Journal of pidgin and creole languages 18.1 (2003), pp. 1–79.
[27] Ray Jackendoff. ‘Possible stages in the evolution of the language capacity’. In: Trends in cognitive sciences 3.7 (1999), pp. 272–279.
[28] T Florian Jaeger. ‘Categorical data analysis: Away from ANOVAs (transformation or not) and towards logit mixed models’. In: Journal of memory and language 59.4 (2008), pp. 434– 446.
[29] Richard S Kayne. The antisymmetry of syntax. 25. mit Press, 1994.
[30] David Kemmerer. ‘The Cross-Linguistic Prevalence of SOV and SVO Word Orders Reflects the Sequential and Hierarchical Representation of Action in Broca’s Area’. In: Language and Linguistics Compass 6.1 (2012), pp. 50–66.
[31] Wolfgang Klein and Clive Perdue. ‘The Basic Variety (or: Couldn’t natural languages be much simpler?)’ In: Second language research 13.4 (1997), pp. 301–347.
[32] Alan Langus and Marina Nespor. ‘Cognitive systems struggling for word order’. In: Cognitive psychology 60.4 (2010), pp. 291–318.
[33] Gary Lupyan and Morten H Christiansen. ‘Case, word order, and language learnability: In-sights from connectionist modeling’. In: Proceedings of the Cognitive Science Society. Vol. 24. 24. 2002.
[34] Gary Lupyan and Rick Dale. ‘Language structure is partly determined by social structure’. In: PloS one 5.1 (2010), e8559.
[35] Hanna Marno et al. ‘A new perspective on word order preferences: the availability of a lexicon triggers the use of SVO word order’. In: Frontiers in psychology 6 (2015).
[36] Luke Maurits and Thomas L Griffiths. ‘Tracing the roots of syntax with Bayesian phylo-genetics’. In: Proceedings of the National Academy of Sciences 111.37 (2014), pp. 13576– 13581.
[37] John McWhorter. ‘The world’s simplest grammars are creole grammars’. In: Linguistic ty-pology 5.2 (2001), pp. 125–66.
[38] Irit Meir et al. ‘Body as subject’. In: Journal of Linguistics 43.3 (2007), pp. 531–563. [39] Irit Meir et al. ‘The effect of being human and the basis of grammatical word order: Insights
from novel communication systems and young sign languages’. In: Cognition 158 (2017), pp. 189–207.
[40] Frederick J Newmeyer. ‘On the reconstruction of ‘proto-world’word order’. In: The evolu-tionary emergence of language: Social function and the origins of linguistic form (2000), pp. 372–388.
[41] Mark Pagel. ‘Human language as a culturally transmitted replicator’. In: Nature Reviews Genetics 10.6 (2009), pp. 405–415.
[42] Jonathan W Peirce. ‘PsychoPy—psychophysics software in Python’. In: Journal of neuros-cience methods 162.1 (2007), pp. 8–13.
[43] Maria Mercedes Pi˜nango. ‘Syntactic displacement in Broca’s aphasia comprehension’. In: Grammatical disorders in aphasia: A neurolinguistic perspective (1999), pp. 74–89.
[44] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. Vienna, Austria, 2015. url: https://www.R-project.org/. [45] Ankit Rohatgi and ZlatanStanojevic. WebPlotDigitizer: Version 3.9 of WebPlotDigitizer.
Oct. 2015. doi: 10.5281/zenodo.32375. url: https://doi.org/10.5281/zenodo.32375. [46] Maria de Sam Lazaro. ‘Investigating iterated learning of an artificial gestural language: When
word order is and is not learned, and differences from improvised pantomime’. Master’s thesis. University of Edinburgh, 2014.
[47] Wendy Sandler et al. ‘The emergence of grammar: Systematic structure in a new language’. In: Proceedings of the National Academy of Sciences of the United States of America 102.7 (2005), pp. 2661–2665.
[48] Marieke Schouwstra, Kenny Smith and Simon Kirby. ‘From natural order to convention in silent gesture’. In: The Evolution of Language: Proceedings of the 11th International Confer-ence (EVOLANG XI). EVOLANG XI, New Orleans, United States. Vol. 20. 2016, p. 24. [49] Marieke Schouwstra and Henri¨ette de Swart. ‘The semantic origins of word order’. In:
Cog-nition 131.3 (2014), pp. 431–436.
[50] Kaius Sinnem¨aki. ‘Word order in zero-marking languages’. In: Studies in Language. Interna-tional Journal sponsored by the Foundation “Foundations of Language” 34.4 (2010), pp. 869– 912.
[51] Russell S Tomlin. Basic Word Order: Functional Principles. Croom Helm, 1986.
[52] Michael T Ullman. ‘A neurocognitive perspective on language: The declarative/procedural model’. In: Nature reviews neuroscience 2.10 (2001), pp. 717–726.
[53] Hadley Wickham. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York, 2009. isbn: 978-0-387-98140-6. url: http://ggplot2.org.
[54] Bodo Winter. Linear models and linear mixed effects models in R with linguistic applications. 2013. url: http://arxiv.org/pdf/1308.5499.pdf.
[55] Alison Wray and George W Grace. ‘The consequences of talking to strangers: Evolutionary corollaries of socio-cultural influences on linguistic form’. In: Lingua 117.3 (2007), pp. 543– 578.
