Introduction

Languages are the product of long-term cumulative cultural evolution [1]. As the product of social learning, transmission and use, languages respond to selection pressures posed by local communication contexts deriving from physical, social and cognitive environments [2,3]. Although evolutionary changes have resulted in a vast array of over 7000 languages spoken in the world today [4], the questions of “why so many languages” and “why so unevenly distributed” are far from being answered. Nonetheless, a vast corpus of research has emerged on the evolutionary, ecological and social correlates of the global distribution of linguistic diversity [5–15].

Various studies have focused on demography as a key determinant of rates of cultural and linguistic evolution [16–22]. The most discussed demographic factor is population size, but studies disagree on its effects. Some analyses suggest that larger populations promote innovation, are less prone to cultural drift and random loss of linguistic elements, and may exhibit less stringent enforcement of norms, thus allowing languages to change faster [1,16,17,19,23–25]. In contrast, other studies have argued that linguistic differentiation should be fastest in small populations due more rapid diffusion of new features [10], greater tolerance of diversity [26], and stronger response to contact resulting from trade and marriage across groups [24]. Recent models have added population structure as an essential demographic factor underlying cultural and linguistic evolution. For example, population density, local interconnectedness and migrations were claimed to play a key role in cumulative cultural evolution by facilitating the emergence, diffusion and survival of linguistic innovations [18,21,27–30]. In contrast, social, cultural and political settings [31,32] might also promote social group cohesion and the sharing of linguistic features within groups, which would reduce rates of linguistic differentiation [33].

In addition to social and demographic factors, geographical isolation may also constrain communication between populations and reduce diffusion of linguistic or cultural traits, increasing the likelihood of random losses of cultural items [17,34], in a process analogous to genetic drift [35]. However, lost items may be replaced by novel innovations that would be more easily fixated [36,37]. In contrast, other studies proposed that contact is more effective than isolation in promoting language differentiation, as it facilitates the introduction of novel linguistic forms via second language acquisition [38].

Another factor often associated with linguistic diversification is warfare. From this perspective, conflict between groups may foster deliberate linguistic differentiation [39,40] as the result of cultural group marking [41,42]. In other words, whilst the use of language as an index of cultural group identity is widespread [39,40,43,44], experimental paradigms [42], computational simulations [41] and ethnographic accounts [45] suggest that this indexical function of language may be particularly salient when neighbouring groups are in conflict with one another, as the ability to distinguish in-groups from out-groups may become a matter of life or death. Thomason [29] describes several examples of this last phenomenon, such as the meeting that a European missionary attended during the 17th century at which the Delaware Indians planned to substitute different words for their native lexicon when they went to war against the Iroquois, so that their enemies wouldn’t understand them. A similar trend was observed among sixteenth-century Portuguese speakers trying to deliberately differentiate their language from that of their Spanish opponents. A related argument for a prominent role of between-group conflict is cultural group selection [46,47], which proposes that linguistic diversity reflects the relative success of competing cultural groups and thus ascribes a crucial role to warfare, extinction and acculturation in the distribution of languages [47]. However, empirical validation of those claims remains limited.

One of the reasons for the ongoing debate over language diversification is that previous studies have not yet attempted to include all the key geographic, social and demographic variables discussed above in a single analysis. This task is even more challenging since many of these factors act simultaneously and tend to be more similar between closely related languages [22,48]. In addition, most have used contemporary speaker population sizes in order to explain linguistic evolution over many years [22,48]. This approach is in general problematic for establishing a direction of causality, and in particular in Austronesia given that the demographic, socioeconomic, and cultural landscape has dramatically changed over the past century first due to European colonialism and imposed linguistic policies[49], and then due to the current globalising trends [49–51] as well as high rates of population growth [48]. Hence, if we wish to evaluate the factors that have affected the evolution of Austronesian languages, they should ideally reflect those societies prior to colonial times (see Bromham et al. [48]). Here we provide the first integrated test of the effect of various socio-demographic and geographic features on linguistic diversification among 50 Austronesian languages. The Austronesian family is the second most diverse linguistic group in the world with nearly 1200 languages and has been extensively investigated [40,52]. Austronesian languages tend to be restricted to clearly defined islands or archipelagos across the Pacific and Indian oceans [49], which offers an ideal context to investigate how geography affects genetic or cultural diversity [53].

We obtained information concerning the historical (prior to large-scale modernisation) state of five variables: population size, geographical isolation, within-group conflict, between-group conflict (same cultural group), and between-group conflict (distinct cultural groups; Table 1) from 50 Austronesian ethnolinguistic groups (Fig 1). They were used in Poisson generalized linear models as predictors of three measures of linguistic diversification: The rate at which new basic vocabulary items are added to a languages’ vocabulary (word gains), the rate at which existing basic vocabulary items are lost (word losses), and the overall effect of these two processes in the divergence of vocabularies between pairs of sister languages (lexical turnover). Either a higher rate of word gains or word losses result in sister languages having less words in common to define a pre-specified list of concepts (see Methods and Greenhill et al. [22] for a use of the same method). We applied Bayesian Poisson regressions to our 25 sister pairs using a time-calibrated phylogenetic sister-pairs approach [22,54] to control for shared evolutionary histories and overcome the problem of statistical non-independence that often characterises comparative studies of cultural and linguistic evolution [47].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Map indicating the approximate geographical location of the 54 Austronesian languages considered in our analyses.

Source: ww.naturalearthdata.com.

https://doi.org/10.1371/journal.pone.0243171.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Predictor variables used in our analyses.

All predictors were taken from the Pulotu dataset [82]. In addition to the specified transformations, all predictors were standardised (we subtracted their means and divided the result by their standard deviation).

https://doi.org/10.1371/journal.pone.0243171.t001

Different from other approaches to evolutionary rates (such as least-squares regressions based on data diagnosed by Welch and Waxman tests), the sister-pairs approach does not increase uncertainty for more closely related languages, excludes data points with relatively unreliable rate estimates, does not require divergence time estimates, and has higher statistical power to detect associations between rates of lexical change and fast-changing demographic or social parameters [54]. As Bromham [55] points out, selecting phylogenetically independent sister pairs is equivalent to running an experiment over and over again, taking one language, splitting it in two, and seeing which one evolves faster.

Results

Multiple social, demographic and geographic factors exert independent effects on the rates of linguistic differentiation. For word gains, word losses as well as lexical turnover, the three full models including the five variables always provided the best fit, with a WAIC weight of 1 (S2–S4 Tables). The full models explained the majority of the proportion of the variance in rates of word gains, losses and overall lexical turnover (Bayesian R² values of 0.59, 0.70 and 0.66 respectively), also indicating that the five predictors had a stronger effect on rates of word losses.

Among the five factors, geographical isolation exerted the strongest effect on overall lexical differentiation, significantly increasing its pace (estimate = 1.09; 90% HPDI: [0.92, 1.24]) (Fig 2). This is because although it moderately increased the ability at which languages acquired new lexical items (estimate = 0.29, 90% HPDI: [0.06, 0.51]) it also severely increased the rate of word loss (estimate = 0.86, 90% HPDI: [0.70, 1.04]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Entire posterior distribution of the full models predicting the number of word gains (left), the number of word losses (centre) and the overall lexical turnover (word gains + word losses).

The thick line represents the median parameter estimate and the shaded region the 90% Highest Posterior Density Interval (HPDI) that reveals the narrowest interval containing the specified probability mass from 24,000 samples from the posterior distribution.

https://doi.org/10.1371/journal.pone.0243171.g002

In the models, the languages of cultural groups with larger populations gained words at a faster rate (estimate = 0.52, 90% HPDI: [0.34, 0.68]). Although we also found evidence that these languages also lost words at a slightly faster pace (estimate = 0.14, 90% HPDI: [0.08, 0.20]), the overall positive effect of speaker population size on lexical turnover was not significant (0.06, 90% HPDI: [-0.10, 0.23])(Fig 1).

Even if within-community conflict did not have a significant effect on the rate of creation of new words (estimate = 0.16, 90% HPDI: [-0.03, 0.36]), it accelerated the rate of word loss (estimate = 0.27, 90% HPDI: [0.09, 0.47]). As a result, conflict within communities of the same culture had an overall significant positive effect on lexical differentiation (estimate = 0.16; 90% HPDI: [0.01, 0.16]).

Conflict between communities of the same culture reduced words gains (estimate = -0.30; 90% HPDI: [-0.50, -0.11]) but increased the pace at which languages lost words (estimate = 0.37; 90% HPDI: [0.16, 0.58]) (Fig 2). Consequently, there was no significant effect of conflict between communities of the same culture on the overall rate of lexical differentiation (estimate = 0.13; 90% HPDI: [-0.01, 0.27]). In contrast, inter-cultural conflict significantly reduced both the rate of word gains (estimate = -0.55; 90% HPDI: [-0.69, -0.40]) and word losses (estimate = -0.38; 90% HPDI: [-0.51, -0.25])(Fig 2), resulting in a deceleration in the overall rate of lexical differentiation (estimate = -0.40; 90% HPDI: [-0.51, -0.30]).

Discussion

We applied a phylogenetic sister-pairs approach to three measures of lexical divergence among Austronesian languages and showed that the five sociodemographic factors were able to explain most variation in lexical differentiation between Austronesian languages. Among the five variables, geographical isolation exhibited the largest effect on lexical turnover. We found that being geographically isolated (further away from a landmass inhabited by members of another culture) greatly speeded up the rate of word losses and also moderately that of word gains. This is consistent with the idea that geographical isolation results in a higher risk of random loss of cultural items due to incomplete inter-generational sampling of existing variation [17,19]. However, we also found evidence that new lexical innovations were more likely to be driven to fixation in more isolated cultures. A powerful effect of drift acting on cultural evolution has also been reported for other traits such as technological toolkits [23], phonemes [56] or even baby names across the United States [57].

We found no association between geographical isolation and within-community conflict in our sample (χ² = 52.66, df = 44, p = 0.17; S2 Fig). Therefore it is unlikely that geographical isolation promotes differentiation either by increasing social group cohesion [33] or intensifying within-group conflict over resources [58]. In summary, we argue that geographical isolation may have contributed to differentiation among Austronesian languages mostly by preventing the random loss of existing lexical items, despite also hindering the ability of novel lexical innovations to emerge and be utilised [38,59], rather than by affecting the internal social dynamics of isolated groups.

Although population size did not have an effect on the overall rate of linguistic differentiation, larger population sizes increased the rate at which languages acquired new vocabulary items. Multiple studies have proposed that populations containing more individuals provide more chance for innovations to arise [19,60,61]. However, the positive association between population size and rates of word gains may also be due to large populations having less stringent norm enforcement allowing them to change faster [24,25]. At the same time, we also report a slight positive effect of population size on the rate at which languages lost existing lexical items. Whilst this is at odds with claims that smaller populations should be more prone word losses due to random sampling effects [16,17,19,23], influential linguistic theories, notably those by Lesley and James Milroy [26,62] have already remarked that small populations may have greater tolerance for diversity and malleable linguistic representations [63], facilitating their retention of linguistic variants. In contrast, we found no evidence for to smaller populations being able to gain new words more rapidly due to a faster uptake of innovations [25,64] nor for larger populations being less vulnerable to the loss of lexical items due to a larger number of sources of social learning and thus increased fidelity of information transmission [65].

Our findings are also at odds with Greenhill et al. [22], who did not find an association between population size and either rates of word gains or word losses in the Austronesian language family. We believe the reason for this discrepancy is twofold: First, they did not include any additional socioecological variables, whose confounding effects may have masked the effect of population size. When running our models including only population size as predictor, its effect decreased on the rate of word gains decreased from 0.52, 90% HPDI: [0.34, 0.68] to 0.09, 90% HPDI: [0.01, 0.18] and became insignificant for word losses (estimate = 0.01, 90% HPDI: [-0.10, 0.12]). But most importantly, the disparate results we observe are due to our choice of population size variable: whereas Greenhill et al. used contemporary speaker population sizes obtained from the Ethnologue [4], we used historical population sizes (prior to large-scale modernization trends taking place throughout past century). We believe that considering historical instead of contemporary population size estimates was important not only to match the temporal resolution of the other socioeconomic variables but also to establish a direction of causality. Since much of the lexical data used for the construction of the phylogenetic tree providing sister pairs for both our and Greenhill et al.’s analyses (see Materials and Methods) were obtained during the early twentieth century, contemporary speaker population sizes may not be representative of the samples when linguistic data were obtained. For example, Hawaii is coded by the Ethnologue as having 2,000 speakers, despite it being one of the languages with the largest numbers of speakers historically [66]. In addition, if demographic variables were collected more recently than linguistic ones, population size would be more properly interpreted as an outcome (rather than a predictor or cause) of linguistic dynamics.

To verify that our discrepancy with Greenhill et al. [22] was due neither to our smaller sample size nor to our choice of statistical methods, we ran their generalized linear models as well as our Bayesian Poisson regressions on our own data, but using contemporary in-area speaker population sizes from the Ethnologue [4] as the only predictor. Both methods failed to identify a significant effect of contemporary speaker population size on either rates of word gains (estimate = 0.05, P = 0.12 with their method; estimate = 0.05, 90% HPDI: [-0.03, 0.13] with our method) or word losses (estimate = -0.02, P = 0.57; estimate = 0.01, 90% HPDI: [-0.07, 0.09]). Finally, although our sample is smaller, it was still three times as large as those utilised in similar previous studies [48], and importantly, more evenly representative of the different geographical areas where Austronesian languages are spoken as well as their different linguistic subgroups (Fig 2)

Our findings seemingly contradict those obtained in other cultural domains where well interconnected groups tend to lose cultural diversity faster, possibly due to the propensity of individuals to learn from successful cultural models and hence accelerated population convergence [67]. This was illustrated by the fact that high levels of intra-cultural conflict (both within communities and between communities of the same culture), which presumably reduces connectivity-within groups, accelerated word losses. A possible explanation for this phenomenon is that internal conflict may reduce the density and connectivity of local social networks [68] to a point where it hinders their ability to maintain social conventions including words [69,70]. In addition, within-group conflict may disrupt a well-documented positive feedback between social cohesiveness and the establishment of group-level traits such as language [71]. However, our finding was independent of population size, as the two variables were not significantly correlated (S2 Fig). The different conflict variables were not associated with each other either, hence we find no evidence for some groups being generally more prone to conflict than others.

Social isolation, which acts as a social barrier to cultural and linguistic exchange, had the opposite effect of geographical isolation of rates of linguistic differentiation. Conflict between cultures decelerated both the rate of word gains and losses. Inter-cultural conflict may discourage communication between speakers of different languages and therefore impose social (as opposed to physical) barriers to the emergence of new variants via the process of second language acquisition [72,73]. At the same time, since conflict poses a risk of acculturation or extinction, groups may turn to linguistic prescriptivism and the use of language as a marker of identity to prevent losses of structural integrity [45] and preserve group boundaries [41], thus resulting in a reduced rate of word loss.

Finally, although it is not within the scope of this study to provide a formal test on the role of cultural group selection on language evolution [46,47] our results did not support the idea that between-group conflict promotes linguistic differentiation at least in the case of Austronesian languages. In fact, Austronesia is famous for the importance of long-reaching networks of institutionalised ritualistic alliances such as the Kula Ring [74] and other cultural practices such as spousal exchanges, collective defence arrangements and exchange of social information and technology [75,76] aimed at creating ties between groups separated by thousands of kilometres. Those processes have been interpreted as adaptations for mitigating the potentially detrimental effects of isolation on the genetic and cultural diversity of insular populations [17]. Therefore, peaceful contact rather than warfare seems to contribute to extensive multilingualism and fast lexical turnover among Austronesian populations, whose mean distance to the nearest landmass is over 128 km in our sample.

In summary, we identified increased isolation and internal conflict, and reduced between-group conflict, as factors contributing to linguistic differentiation. They operate by altering rates of word gain, word loss, or both. In addition, although population size did not have an effect on the overall rate of lexical evolution, languages spoken in larger communities gained words at a faster pace. Our conclusions may be specific to the case of Austronesian languages, spoken by populations separated by long distances and living on islands often unable to support multiple communities or cultural groups. It is not clear whether linguistic differentiation would reflect a different set of factors in groups where isolation does not pose an imminent risk, or on islands such as Papua New Guinea where a large number or groups are present and where social boundaries resulting from between-group conflict may be as relevant as geographical boundaries. In addition, the median number of speakers in our sample in only 26,485 compared to over seven million in Indo-European languages [4], which is relevant since the effect of population size may change beyond a given critical mass [77]. Finally, social and demographic factors may be more relevant to the evolution of lexicon than grammar or phonology, as the former have been shown to evolve in punctuational bursts rather than in a gradual fashion among Austronesian languages [39,40].

Materials and methods

The Austronesian language family is one of the most diverse on the planet, comprising between 1,100 and 1,200 languages [78]. Austronesian cultures and languages are the product of a recent expansion [52] and thus Austronesian cultures and languages share many similarities with one another [79]. Their spread into Oceania was part of an expansion starting from Taiwan at around 5,000 ya, reaching the Bismarck Archipelago by 3,400 ya and Remote Oceania by 3,200 ya, in association with the appearance of Lapita pottery [80].

Supporting information

Acknowledgments

The authors wish to acknowledge Simon Greenhill for help with the implementation of RateCounter and Andrea Migliano for helpful discussions on previous versions of the paper. We also want to thank Terhi Honkola and two anonymous reviewers for their contribution to improving the manuscript.