The Role of Space and Body in Referent Distinction Strategies: A Cross‐Cultural Comparison

Abstract

The emergence of linguistic structure is shaped by interaction and cultural transmission as well as cognitive biases and social conventions. An important aspect of linguistic structure is reference distinction or “who did what to whom.” This paper explores how cross‐cultural differences in the conceptualization of space may influence the development of strategies for referent distinction in this domain. We conducted a silent gesture experiment employing dyadic interaction and iterated learning with participants from two cultural groups that differ in their predominant conceptualization of space: Balinese and Dutch. Focusing on silent gestures produced by hearing participants naïve to sign languages, we examined the effect of culturally embedded biases on the use of space and body, the resources of manual communication. Our findings reveal the emergence of various strategies that are equally successful but differ in their use of space and body. While the Dutch participants developed multiple different strategy types, the Balinese largely followed one strategy type. Crucially, they did not employ an abstract or grammaticalized conception of space. We argue that, rather than reflecting a developmental shift, this pattern stems from the geocentric conceptualization predominant in Balinese culture that inhibits grammatical use of space and favors alternative strategies, such as capitalizing on the iconic affordances of the body to represent grammatical relationships. Our findings thus provide evidence for how cultural conventions such as the conceptualization of space can influence the evolution of linguistic structure.

1. Introduction

When language users communicate, they often make clear “who did what to whom” by distinguishing different referents. All languages have developed a certain strategy to achieve this, though these strategies differ across languages. Spoken languages usually employ a grammatical case system, make use of verb agreement, or they maintain a dominant conventional word order (e.g., Payne, ¹⁹⁹⁷). Sign languages additionally employ modality‐specific strategies making use of their spatial and bodily nature, for example, through spatial modulations that use space and direction for referential loci (see Fig. 1) and spatial agreement of verbs, or by a change in the signer's bodily orientation to indicate a role shift (e.g., Engberg‐Pedersen, ¹⁹⁹³; Liddell, ²⁰⁰³). These modality‐specific strategies are particularly effective in complex narratives, where agents and patients may change back and forth, and tools are required to indicate this. For example, in the consecutive sentences: “Hannah is walking on the beach with her sister. She brings Hannah a pretty shell.”, here the agent has switched and a strategy is needed to distinguish the primary referent of the first and the second sentence. Even though the nature of visual‐manual devices for reference distinction is relatively well studied across sign languages, little is known about how they emerge and develop over time, and why we see cross‐linguistic variation. Here, we fill this gap by conducting a language evolution experiment in the visual‐manual domain, with participants from two different cultures (Bali, Indonesia, and the Netherlands) who have no experience with sign languages.

Fig. 1.

Example of spatial referent distinction in the manual modality. While telling a story, the boy points to two locations in space to represent his brother and mother, respectively. These locations are assigned arbitrarily and do not correspond to the referents’ actual position or a physically present referent. Later, he may point back to the indexed locations to indicate a specific referent (his brother or mother). This strategy has been referred to as abstract deixis in co‐speech gesture (McNeill et al., 1993) and referential loci in sign languages (Engberg‐Pedersen, 1993; Liddell, ²⁰⁰³), where it is often conventionalized and grammaticalized.

To answer questions about the origins of linguistic structure, research has been conducted using various experiments across modalities in controlled lab settings, simulating mechanisms and pressures believed to shape emerging communication systems (for an overview, see Tamariz & Kirby, ²⁰¹⁶). Other studies have related results from these lab experiments to data from natural languages, to provide a comprehensive image of conventionalization (e.g., Flaherty & Schouwstra, ²⁰²⁴; Meir et al., ²⁰¹⁷; Schouwstra, ²⁰¹⁷). Here, we contribute to this line of research by focusing on the cultural context that experimental participants bring into the lab. Specifically, we focus on the emergence of reference distinction across cultures and explore how conceptualization of space may help explain the variation observed in these strategies. Investigating the emergence of visual‐manual devices for referent distinction in this way offers a unique chance to assess in detail how existing conceptual devices (spatial conceptualization and properties of the body) are put to use to form linguistic conventions around reference over time.

1.1. Conveying information in the absence of established linguistic conventions

An experimental paradigm that allows for investigation of communication in the absence of linguistic conventions is called silent gesture (Gershkoff‐Stowe & Goldin‐Meadow, 2002; Goldin‐Meadow, So, Özyürek, & Mylander, ²⁰⁰⁸; Naegeli & Schouwstra, ²⁰²⁵). In silent gesture experiments, hearing participants, who have no previous experience with conventionalized visual‐manual communication systems (i.e., sign languages), are asked to convey information solely with their hands and bodies. Silent gestures are different from co‐speech gestures, which occur alongside speech, in that they bear the entire communicative load (Goldin‐Meadow & Brentari, 2015). As hearing participants are largely unfamiliar with communicating exclusively in the visual‐manual modality, they must improvise on the spot and in the absence of a conventionalized linguistic system. Previous research using this paradigm has shown that participants might express themselves differently in silent gesture than what is conventional in their spoken languages. For instance, participants from different linguistic backgrounds chose to order simple gesture sequences according to Subject‐Object‐Verb order even when the dominant order of their spoken language was different, for example, Subject‐Verb‐Object (Goldin‐Meadow et al., 2008; Hall, Mayberry, & Ferreira, ^{2013; 2014}). Therefore, these experiments can provide us with an insight into cognitive preferences for organizing information independent from learned linguistic systems.

1.2. Simulating mechanisms of cultural evolution

Languages are shaped by repeated interactions between language users, and by being acquired by new generations of learners over time. These mechanisms of cultural evolution can be simulated in lab settings by implementing experimental paradigms. To model language use, experiments can be interactive, pairing participants in dyadic or even group experiments in which they communicate with each other to successfully complete a task (e.g., Galantucci, ²⁰⁰⁵; Raviv, Meyer, & Lev‐Ari, ²⁰¹⁹; Smith, Feher, & Ritt, ²⁰¹⁴). To simulate transmission of language across generations of learners, the iterated learning paradigm (Kirby, Cornish, & Smith, 2008) can be used. In this paradigm, the output of a single or multiple participants is used as training input for the next participant(s) whose output is then in turn used as training input for the next participant(s) and so forth, to build chains in which each generation learns from the previous one (e.g., Beckner, Pierrehumbert, & Hay, ²⁰¹⁷; Kirby et al., ²⁰⁰⁸; Kirby, Tamariz, Cornish, & Smith, ²⁰¹⁵). Tamariz and Kirby (2016) compiled an extensive overview of experimental research that used interaction and iterated learning to understand the pressures of cultural evolution of language, concluding that together they shape emerging linguistic structures. Furthermore, combining these paradigms with silent gesture allows researchers to tease apart various pressures of cultural evolution that shape conventionalized communication systems from improvisation in the absence of linguistic structure (Motamedi, Schouwstra, Smith, Culbertson, & Kirby, 2019).

1.3. The biases that remain: Space across cultures and modalities

While silent gesture provides an opportunity to observe the emergence of communicative strategies with the influence of learned linguistic conventions minimized, it is important to acknowledge that this does not mean there are no other biases participants bring into the lab. Silent gesture experiments are by default bodily, as information must be conveyed by the body, and spatial, as gestures are always made in space. The use of the body in gestures and the conceptualization of space are not universal but differ across communities and come with their own set of cognitive and cultural biases. Here, we define culture as socially transmitted information (Mesoudi, 2016), including spatial conceptualization and gestural conventions. Specifically, conventions on how space is conceptualized, while also often reflected in language, are not merely linguistic phenomena. In fact, the linguistic conceptualizations of space are flexible and do not always match how space is mentally represented (Bohnemeyer et al., 2022; Kısa, Stengelin, Maurits, & Haun, ²⁰²⁶; Wassmann & Dasen, ¹⁹⁹⁸). This suggests that spatial conventions can work independently and exert influence even when language is restricted.

Most silent gesture studies have been conducted in Western settings with participants from similar cultural backgrounds who share broadly similar spatial and gestural conventions. For the current study, we placed our focus on how these cultural biases may influence new communication systems emerging. We did this by conducting a silent gesture study into the emergence of referent distinction strategies across two different cultural groups: speakers of Balinese (Bali, Indonesia) and speakers of Dutch (The Netherlands). For the current study, we chose participants from these origins, as the respective cultural groups differ substantially in their conceptualization of space and their gesture conventions.

In Dutch culture, as in many other Western cultures, space is predominantly conceptualized in an egocentric way relative to the viewpoint of the speaker (Levinson, 1996). When using their language, Dutch speakers commonly use the Dutch equivalent of relative terms such as left and right in their spatial descriptions. This conceptualization of space also translates to co‐speech gestures. When retelling a story, a Dutch person might refer to someone who stood on their left by pointing to their left, regardless of their absolute position when retelling the story.

In Bali, on the other hand, space is traditionally conceptualized in a geocentric way. Instead of relative to an observer's viewpoint, spatial references are absolute, anchored to landmarks on the island such as the mountainous center and the seashore surrounding the island or the directions of the sun rising and setting (Aryawibawa, Putra Yadnya, Ngurah Parthama, & Pye, 2018; Covarrubias, ¹⁹⁵⁰; Dasen & Mishra, ²⁰¹⁰; Wassmann & Dasen, ²⁰⁰⁶). Indeed, spatial references made in Balinese use directional terms like kelod (mountainward) and kaja (seaward). These directions are embedded in Balinese culture, that is, everyday life, customs, and religious practices, reaching far beyond just being used to communicate spatial information. Houses and compounds are built in adherence to this spatial system, and it dictates which direction temples and altars should be facing. Even the position of one's head when in bed is prescribed by this spatial system. This conceptualization of space also influences co‐speech gesture, with Balinese pointing gestures adhering to their geocentric spatial system (Dasen & Mishra, 2010; Wassmann & Dasen, ¹⁹⁹⁸). Thus, while retelling a story, a Balinese person would conventionally and consistently point to a location mountainward if the person stood on that side, irrespective of the location where the retelling takes place (see Fig. 2). Similarly to space, culture‐specific conventions have been found to shape the way we use our hands when we gesture (e.g. Kita²⁰⁰⁹& Özyürek, ²⁰⁰³). Notably, in the case of Balinese culture, pointing with an outstretched index finger, while common across Western cultures, is considered impolite and pointing gestures are more likely to use the thumb or the whole hand (Imelda & Harisal, 2024).

Fig. 2.

Visualization of the same event in two different spatial conceptualizations: egocentric (prevalent in Western cultures, e.g., Dutch and British) and geocentric (predominant way of conceptualizing space in Balinese culture).

The lab is an exciting site for research into the emergence of linguistic patterns and how participants who culturally conceptualize space differently might develop different strategies, as it offers high levels of experimental control. At the same time, it is constrained and artificial as it often targets one specific linguistic structure in isolation. Therefore, it is helpful to also look at sign languages, especially young sign languages,¹ to refine research questions or to put results from silent gesture experiments into perspective (e.g., Flaherty & Schouwstra, ²⁰²⁴; Flaherty et al., ²⁰¹⁸; Meir et al., ²⁰¹⁷; Motamedi, Schouwstra, & Kirby, ²⁰¹⁷). It also serves to consider these sign languages within the broader cultural and semiotic context in which they emerge (de Vos, ²⁰¹⁴; Nyst, ²⁰¹⁹).

This is especially relevant in the Balinese context, as Kata Kolok, a sign language used by a remote community in the north of the island, exhibits patterns not common to other sign languages. In Kata Kolok, in addition to only a few other sign languages (Bauer, 2014; Schuit, ²⁰¹⁴), space is used in an absolute, geocentric way (de Vos, ²⁰¹⁴). Second, while grammatical use of space like spatial modulations (such as using referential loci indexed in the signing space) are prevalent across many sign languages, it has not developed in Kata Kolok (de Vos, ²⁰¹⁴). Instead, Kata Kolok signers prefer using points toward the fingers of the nondominant hand known as list buoys (Liddell, 2003; Marsaja, ²⁰⁰⁸). Therefore, space is not used in a grammaticalized, abstract way.

Absence of grammatical use of space in sign language has often been linked to time: it is (implicitly) assumed that languages that lack it might still develop it (e.g., Padden, Meir, Aronoff, & Sandler, ²⁰¹⁰). Here, we remain open to the possibility that grammatical use of space is not the only “end point” of language development; in some cultural contexts, this might just not be the preferred strategy, and other strategies might work better. Whether or not the geocentric representation of space in Bali and the exceptional lack of grammatical space in Kata Kolok are causally related is currently not known. Our study serves to advance the understanding of how spatial conceptualization and reference devices in visual‐manual communication are related.

1.4. Previous work

To study whether and how space is used to distinguish referents by participants who are naïve to sign languages, we adapted an experimental methodology presented in Motamedi, Smith, Schouwstra, Culbertson, and Kirby (2021) in which emerging referent distinction strategies were studied in a silent gesture experiment. This experiment implemented interaction between participants as well as iterated learning with UK‐based, English‐speaking participants. The goal was to investigate how systematic argument distinction evolves over generations. Additionally, they examined whether strategies involving grammatical use of space could emerge. As we have seen, spatial modulation is a feature present in most sign languages but absent from some sign languages (e.g., Kata Kolok, de Vos, ²⁰¹⁴; Kafr Qasem Sign Language and Al‐Sayyid Bedouin Sign Language, Meir, Israel, Sandler, Padden, & Aronoff, ²⁰¹²), which is often attributed to their relative youth (Sandler, Meir, Padden, & Aronoff, 2005). The findings of Motamedi et al. (2021) suggest that participants with no prior signing experience can actually develop strategies that resemble those found in established sign languages.

Motamedi et al. (2021) conducted five iterated learning chains, which consisted of five generations each. Every generation in every chain comprised of two participants who were paired at random. In each chain, the first generation of participants improvised gestures in a director‐matcher task (with roles alternating between participants). All subsequent generations were trained by participating in an additional experimental phase during which they were shown the video output of the previous generation in their respective chain. The experimental stimuli consisted of orthographically presented sentence pairs describing two events consisting of an agent, an action, and in most cases, a patient or recipient. The sentence pairs appeared in one of two types: either the agent remained the same in both sentences of the pair (same‐agent sentence pairs), or the agent changed between sentences in the pair (different‐agent pairs) (see Table 1 for examples). To communicate successfully with their partner, participants had to develop a (referent distinction) strategy to clarify who the agent was in each sentence of the sentence pair.

Table 1.

Examples for the two sentence pair types

Same‐agent sentence pair	Different‐agent sentence pair
Hannah is swimming to Sarah, Hannah is cycling to Sarah	Hannah is scolding Sarah, Sarah is helping Hannah

Relevant to the current study, Motamedi et al. (2021) predicted that participants would rely on using their own bodies to iconically represent agents. Their results showed that participants developed strategies on which they converged within chains. They found that each of the five chains developed one of three unique strategy types which made use of gesture, body, and space in different ways. Two chains developed lexical strategies, in which referents were distinguished by a different agent‐type gesture. Concretely, participants used a 1‐handshape gesture for the first agent and a 2‐handshape gesture for the second agent in different‐agent pairs, while a 1‐handshape gesture was used for both sentences in a same‐agent pair. In contrast, two other chains used body (shift) strategies, in which participants shifted the orientation of their bodies to indicate a switch in referents. These strategies were similar to role shift strategies observed in sign languages, where the body of the signer is used as a grammatical subject, and confirm the prediction that participants would make use of the iconicity of the body to represent animate agents (Meir, Padden, Aronoff, & Sandler, 2007). Lastly, one of the chains developed an indexing strategy: both agents were assigned to locations arbitrarily chosen to represent them on an abstract level (McNeill et al., 1993; Perniss, ²⁰¹²). This strategy of using referential loci is also commonly found across sign languages (Engberg‐Pedersen, 1993; Perniss, ²⁰¹²). The development of this strategy, over the course of generations of participants, was remarkably similar to diachronic changes also found in sign languages. In earlier generations, one rather than two points in space were used: one agent was mapped on the participant's body, while the other agent was indicated somewhere opposite the body.

In a similar vein, in both Al‐Sayyid Bedouin Sign Language and Nicaraguan Sign Language, it has been observed that older generations (or “cohorts”) preferred to use their own body representing one of the agents, whereas younger generations use an abstract and grammaticalized version, in which referential loci are employed to refer to agents removed from the signer's body (Montemurro, Flaherty, Coppola, Goldin‐Meadow, & Brentari, 2019; Padden et al., ²⁰¹⁰). This trajectory, from anchoring an argument on the signer's body to projecting referents onto abstract locations in space, has been described both diachronically as well as developmentally for sign languages. While the term body‐anchored originally referred to a class of verb signs whose location is lexically specified on the signer's body (Shepard‐Kegl, 1985), the notion has been extended to the representation of arguments (Meir et al., 2007), and can also be applied more broadly to describe how reference systems can be anchored to the body. The shift from body‐anchored reference to abstract spatial systems described across cohorts of young sign languages has also been observed in sign language acquisition with young deaf children learning ASL, who initially use their bodies as reference points for arguments and only later project them into space (Meier, 1982; Newport & Meier, ¹⁹⁸⁵).

In all, the results by Motamedi et al. (2021) show that hearing participants with no experience in using systematic communication in the visual‐manual modality can develop and converge on manual referent distinction strategies in the absence of linguistic conventions. Moreover, they also demonstrate that the grammatical use of body and space (common in sign languages) can be employed by silent gesturers as well. Furthermore, the shift from a body‐anchored reference system to one that uses referential loci over generations, as observed in one chain, mirrors the trajectory observed both diachronically in young sign languages and developmentally in sign language acquisition.

1.5. The current study

The findings from Motamedi et al. (2021) are intriguing as they indicate that participants without any experience in conventionalized manual communication can quickly develop functioning communication strategies in a controlled experimental setting. Furthermore, the study demonstrated that the grammatical use of body and space, as is common in sign languages, was also employed by silent gesturers. This finding is in line with other research suggesting that spatial modulation is not exclusive to sign languages, as it has also been observed in co‐speech gestures of users of spoken language for reference tracking (Perniss & Özyürek, 2015).

In the study at hand, we took these findings as a starting point to evaluate the influence of cultural factors, such as the conceptualization of space, on these strategies. Therefore, we conducted a methodological replication of the study in Motamedi et al. (2021) in the Netherlands and in Bali with participants of the respective cultural backgrounds, who were otherwise comparable in age and educational background. For our study, we had two main objectives.

First, we investigated whether we could replicate the quantitative results found by Motamedi et al. (2021) across cultural groups. Specifically, we examined whether participants developed successful communication strategies to indicate whether the agent remained the same across sentences in the sentence pair or whether there was a switch in agent (see Table 1). Since this distinction was the main focus of our experimental design, rather than distinguishing multiple arguments within a sentence, we refer to the emerged strategies as referent distinction strategies.

We assessed this first objective in two steps: We evaluated whether participants in a dyad were able to communicate successfully by looking at the responses given by the matcher, and we analyzed whether the use of different gestures across a sentence pair was predicted by the type of sentence pair (i.e., same‐agent vs. different‐agent). Given the setup of the experiment, we predicted that dyads would establish successful communication, and, in order to do so, would develop strategies, conventionalized at the chain‐level, to differentiate sentence pair types irrespective of cultural group, as such globally replicating Motamedi et al. (2021).

Second, extending Motamedi et al. (2021), we explored to what extent and in what way the referent distinction strategies differed across the two cultures. As silent gesture relies on the body and space for communication, differences in manual referent distinction strategies can be expected as the respective conventions of Balinese and Dutch culture differ in these respects. However, if all chains develop similar strategies irrespective of cultural background, this might suggest that cultural conventions do not play a role in the emergence of such strategies. Using a qualitative approach, the strategies used for referent distinction in different‐agent sentence pairs and the mechanics (e.g., the use of body and space) of each strategy were examined. We predicted to find a difference in the strategies used across cultural groups in line with the difference in cultural conventions with regard to the conceptualization of space. Specifically, we expected the Dutch participants to develop strategies similar to the ones developed by the UK participants in Motamedi et al. (2021), as both these cultures share an egocentric conceptualization of space. As the experimental stimuli do not provide any spatial information, Balinese participants cannot know the absolute location of the agents in their landmark‐anchored, geocentric spatial system. Our prediction for the Balinese strategies was, therefore, based on what we know from Kata Kolok, a sign language that emerged in the same cultural context. We predicted that the strategies developed by the Balinese chains would rely less on the use of space in a grammaticalized way. Instead, we considered it possible that nonspatial methods like list buoys might be used, for instance, through using points to fingers on the nondominant hand (Cormier, 2012; Liddell, ²⁰⁰³; Marsaja, ²⁰⁰⁸).

2. Method

2.1. Transparency and openness

This study was preregistered on OSF (https://osf.io/w4sgx). The preregistration can be found there, alongside coded data, stimuli lists, translations, and analysis script. Since writing the preregistration, we have reordered our hypotheses and added the prediction formulated as part of our first objective.

As our study is a methodological replication of Motamedi et al. (2021) conducted with Balinese and Dutch participants, we matched the original study's sample size for each cultural group, using five iterated learning chains with 10 participants per chain, paired up into five dyads representing five generations. The experimental procedure closely followed the original design and setup, but we used a web‐hosted experiment rather than two directly connected computers. The experimental stimulus sentences were translated into the respective languages. While the English sentences used present continuous (e.g., “is swimming”), the Balinese and Dutch translations used present simple (akin to “swims”), as this form is more natural in those languages.

2.2. Participants

Fifty Balinese participants (38 female, 12 male) who were recruited from Universitas Udayana in Denpasar, Bali, Indonesia, took part in the Balinese experiment. Their average age was 20.6 years (age range 17–22). All participants were required to be native speakers of Balinese, as the experiment was conducted in Balinese, and to have had no previous experience with sign languages. Participants received 70,000 IDR as compensation for their time and effort in accordance with local conventions.

Fifty Dutch participants (25 female, 25 male) were recruited from Tilburg University, The Netherlands. Their average age was 21.7 years (age range 18–29). As the experiment was conducted in Dutch, participants were required to be native speakers of Dutch. They were also required to have had no previous experience with sign languages. For compensation, participants were given the choice between a 10 EUR gift voucher or course credits (if they signed up via the university's participant pool), following university policy. This study was approved by the TSHD Research Ethics and Data Management Committee.

2.3. Materials

The stimuli consisted of orthographically presented sentence pairs shown on a computer screen. Each sentence described an event involving one of two possible agents: Hannah or Sarah.² The events were either transitive or intransitive and consisted of motion events with no specified end point (e.g., “Hannah is swimming”), motion events with a specified end point (e.g., “Sarah is running to Hannah”), physical transfer (e.g., “Hannah is giving a book to Sarah”), and nonphysical or metaphorical transfer (e.g., “Sarah is praising Hannah”).³ To form a sentence pair, two sentences from the same event type were paired up. Sentence pairs varied in their agent constellation. Either the same person stayed in the role of agent for both sentences in the sentence pair (same‐agent pair), or the agent switched between the first and the second sentence of the pair (different‐agent pair). Events were paired up in a balanced way, and each event and each agent constellation appeared equally frequently throughout the experiment.

Participants were paired up in a director‐matcher task (see below). The director was tasked with conveying information to the matcher, who in turn had to choose the presumed “correct” sentence pair from an array of four pairs, including three distractor sentence pairs. The three distractor sentence pairs were made up from (1) the same events as the target sentence pair but with a different agent constellation, (2) the same agent constellation but using two different events, and (3) a different agent constellation and different events (see Figs. 3 and 4). The inclusion of the distractors encouraged participants to evolve a strategy to convey whether the agent had switched or remained the same across sentences in a pair, while also conveying the right event information.

Fig. 3.

Experiment interface during the Interaction phase for director (left) and matcher (right). A Zoom window was open on the other half of the screen for both participants. Text here translated to English for comprehension.

Fig. 4.

Experiment interface for the Training phase. Text translated to English for comprehension.

2.4. Procedure

The experimental design employed both interaction and iterated learning. Each participant was randomly paired up with another participant, and each of these dyads were randomly assigned to one of five iterated learning chains. In total, we conducted five experimental sessions per chain, simulating five generations of language transmission (cf. Motamedi et al., ²⁰²¹). Each dyad took part in the Interaction phase (see Section 2.4.1) of the experiment. Generations 2−5 also participated in the (iterated learning) Training phase (see Section 2.4.2). Generation 1 did not complete a Training phase but immediately started with the Interaction phase in which they improvised. We will now first provide in detail the general setup and interactive part of the experiment and then describe the iterated learning portion.

2.4.1. Generations 1–5: Interaction phase

For each experimental session, participants in a dyad were placed in different rooms or experiment booths and seated in front of a table with a laptop. We used chairs without arm rests, with wheels, and with the ability to move in all directions in order to least restrict participants in their movements. The laptops were set up in split‐screen view showing the web‐hosted experiment in a Google Chrome browser window on one side and an open Zoom meeting between the participants and the experimenter on the other. Once participants had read and completed the information and consent pages on the browser, they were directed to switch on their camera on Zoom to indicate that they were ready to receive instructions from the experimenter. Once both participants had switched on their cameras, the experimenter followed suit and gave them instructions about the procedure, the number of trials, and reminded the participants that they could use only their bodies to communicate. These instructions were delivered in the respective experiment language (Balinese or Dutch).

Subsequently, the experimenter instructed the participants to switch off their camera and sound and only switch the camera back on if they were in the director role. Participants then proceeded on the browser window, where they received an additional summary of the instructions and then started the experiment. The roles (director vs. matcher) participants started in were assigned randomly. On each trial, the director was shown a stimulus sentence pair and asked to press a button to start the trial. This button press triggered the array of sentence pairs to appear in a vertical list and in random order on the matcher's side of the experiment (see Fig. 3 for the Interaction phase interfaces for director and matcher). The director then switched on their camera to gesture. The matcher's camera was turned off throughout; therefore, they had no way of requesting clarification or repetition. Once the matcher made their choice by clicking one of the four sentence pairs on their screen, both the director and matcher received full feedback displayed on the screen informing them if the matcher made the right choice or not. If the matcher did not choose the target sentence pair, they were shown both the pair chosen by the matcher as well as the target pair. After 5 s, the next target stimuli appeared on the director's screen. Between feedback and the director starting the next trial, the matcher's screen would simply say “waiting for gesturing” in the respective language.

Participants swapped roles (director vs. matcher) every four trials until each of them had been in either role for 16 trials (i.e., four blocks of four consecutive trials). Blocks were introduced to give participants the opportunity to repeatedly track the same referents in the gesturing space (see Motamedi et al., ²⁰²¹). Per participant, two of the four blocks were entirely same‐agent and the other two entirely different‐agent. The order of the sentence pairs within the block was randomized and so was the order in which the eight blocks appeared. In total, we collected data from five chains per cultural group, each consisting of five generations. Each generation comprised of two participants who each produced gestures for 16 trials, amounting to 400 trials per cultural group.

2.4.2. Generations 2–5: Training phase and Interaction phase

The experimental procedure for Generations 2−5 included a Training phase before participants entered the Interaction phase. For this phase, the video output of one of the participants in the previous generation in the chain assigned to them was chosen at random and used as training input. Both participants in one dyad saw the same training videos. The Training phase started immediately after the participants received the instructions (which included instructions about this phase for Generations 2−5) via Zoom. The procedure in the Training phase was identical to what participants subsequently experienced in the matcher role. In each of the 16 training trials, participants watched a video while being shown the sentence pair array (see Fig. 4). They then chose one of the options by clicking on it and received feedback. The videos would play instantly on loading the page, and there was no way for participants to replay the video, similarly to the Interaction phase. Both participants saw the videos in a randomized order. Once one participant had finished the Training phase, instructions were shown on screen telling them to wait for their partner. As soon as the other participant also finished the Training phase, the screen changed to the brief instruction summary of the Interaction phase, and after confirming by button press that they were ready, the Interaction phase (see Section 2.4.1) began as described above.

2.5. Data collection

2.5.1. Data collection in Bali

Balinese data were collected in rooms of the Fakultas Ilmu Budaya of Universitas Udayana in Denpasar by the first author and a local research assistant who took on the role of experimenter. Participants were filmed on the laptops’ internal camera, and the video data were screen‐recorded via the experimenter's laptop in the Zoom application. As both the Zoom app and the web‐hosted experiment needed a stable and consistent internet connection, we had to opt for rooms where an Ethernet connection was available. For 15 of the 25 experiments, we were able to use the two computer labs on campus, but we had to move to different rooms for the remaining 10 experiments. In these rooms, we had to improvise the Ethernet connection from a portable modem and nearby offices. Despite our best efforts, the unstable connection resulted in some trials where the video was frozen or glitchy. This also influenced the procedure: Initially, we had asked the participant in the director role to switch off their camera in between the trials of the four‐trial block. However, we noticed that this often resulted in the video being frozen at the beginning of a gesture string. To avoid this issue, we omitted this from the verbal instructions as soon as we noticed this. It was, however, kept as written text displayed to the director during the experiment as a reminder to switch off the camera once the roles were switched (see Fig. 3). We also gave participants the option to communicate with the experimenter in the Zoom chat (each participant privately with the experimenter) to alert us in case of technical issues.

2.5.2. Data collection in the Netherlands

The Dutch experiments were conducted in single‐participant booths in a lab at Tilburg University. The data from the first generation were collected by a Dutch research assistant, and later generations were conducted by the first author. For the latter, a prerecorded video of a Dutch native speaker reading the instructions was shown to participants on Zoom. The hardware and software setup mirrored the Balinese setup; however, the dimension of the booths dictated a slightly closer and narrower angle. As the instruction to switch off the camera between director‐trials had to be omitted for the Balinese experiment, this was also applied for the Dutch experiment. There were no issues with the connection for the Dutch experiment and none of the participants contacted the experimenter via the Zoom chat function, despite in principle being able to do so. The experimenter occasionally used the chat function to instruct the participant to adjust the camera angle. This was sometimes necessary as some participants, once they were in their booth, moved the laptop computer and, therefore, changed the camera angle of the built‐in camera.

2.6. Data processing and coding

The collected video data were segmented by participant and trial using Adobe Premiere Pro. This had to be done during the running of the study as we needed the videos for the Training phases. After concluding data collection, each trial video was coded by the first author and a second coder using the coding scheme formulated and used by Motamedi et al. (2021). The original coding scheme was supplemented with variables required for strategies not observed by Motamedi et al. (2021) (see adapted version used for this study https://osf.io/w4sgx). The subsequent analysis included four parameters marked by Motamedi et al. (2021) as the ones that indicated referent distinctions: agent type (by changing what kind of gesture was used to identify the agent), agent location (by changing the location of the agent gesture), verb location (by changing the onset location of the gesture representing the action), and verb path (by changing the direction of the gesture representing the action).

From the Balinese data set, 17 trials (4.25%, eight different‐agent trials and nine same‐agent trials) had to be excluded due to the video footage being frozen. None of the Dutch trials had to be excluded. The remaining trials were binomially coded on the parameters mentioned above, receiving a 1 if there was a difference in the participants’ gesturing between the first sentence and the second sentence in the pair and a 0 if not. Proportional agreement between the two coders for the individual difference scores (see below) of the four parameters across both groups ranged from 74.96%−98.34% (Balinese data range: 74.96%−98.34%, Dutch data range: 79.25%−97.63%). For the statistical analysis, the four parameters were combined into one collapsed difference score. If any of the four parameters had received a 1 during the previous binomial coding, the collapsed difference score was also 1, indicating a difference. If all four parameters scored a 0, the collapsed difference score was also 0. Proportional agreement⁴ of the collapsed difference score between the first and second coder was at 90.42% for the Balinese data and at 90.25% for the Dutch data. Given the high agreement, the coding of the first coder was chosen as input for the analysis. It should be noted that the collapsed difference score was used as an overall measure to assess whether participants differentiated between sentence types and did not capture the complexity of how the strategies achieved this. The latter is addressed through the qualitative analysis of strategy types (Section 3.2).

3. Results

3.1. Quantitative results

Our first objective was to evaluate whether our participants would (1) be able to successfully convey the event information, and in order to do so, (2) develop referent distinction strategies to identify the sentence pairs in the experiment.

3.1.1. Communicative success

To examine whether participants communicated successfully, we calculated the proportion of correct responses by the matcher for each generation of each chain (see Fig. 5). For the Balinese data set, the mean proportion of correct responses was at 0.83 (sd = 0.30, range = 0.29−1); for the Dutch data set, the mean proportion was slightly higher at 0.92 (sd = 0.26, range = 0.72−1).⁵ We ran logistic generalized mixed‐effects regression using the lme4 package in R (Bates, Mächler, Bolker, & Walker, 2015; R Core Team, ²⁰²², v4.2.2). Following Motamedi et al. (2021), the model included communicative success (binary, 1 = correct response, 0 = incorrect response) as the outcome variable and Generation as a fixed effect. We also added Sentence Pair Type (sum‐coded: 0.5 = same agent, −0.5 = different agent) and Group (sum‐coded: 0.5 = Balinese, −0.5 = Dutch) as fixed effects, as well as all possible interactions among the fixed effects (see Table 2). We used sum‐coding for the two binary predictors, as we were interested in whether the communicative success varied across groups and sentence pair types in general and had no theoretical reason to treat either level as the reference level. The initial random effects structure of the model included by‐participant,⁶ by‐chain, and by‐item intercepts, but this resulted in singular fit. As the variance of by‐item intercepts was near zero, it was excluded from the final model.

Fig. 5.

Plots showing the mean proportion of communicative success by cultural group and chain over generations. The plot on the left is illustrating the Balinese data, and the plot on the right the Dutch data.

Table 2.

Outcome of logistic generalized mixed‐effects regression model for communicative success

	Estimate Std	Error	z value	Pr (>|z|)
(Intercept)	1.955	0.341	5.728	1.01e‐08***
Generation	0.115	0.098	1.180	0.238
Sentence Pair Type	0.178	0.388	0.459	0.646
Group	−0.687	0.493	−1.394	0.163
Generation × Sentence Pair Type	−0.032	0.121	−0.263	0.793
Generation × Group	−0.072	0.130	−0.554	0.579
Sentence Pair Type × Group	−1.620	0.777	−2.086	0.037 *
Generation × Sentence Pair Type × Group	0.282	0.241	1.171	0.242

Note: Model: Success ∼ Generation * Sentence Pair Type * Group + (1|Chain) + (1|Participant). ***p < .001, **p < .01, *p < 0.05.

None of the fixed effects emerged as a significant predictor of communicative success. This suggests that the difference in communicative success between the Balinese and Dutch chains was not pronounced enough to be statistically significant. Of the interaction terms, we only found a significant interaction effect between Sentence Pair Type and Group (β = −1.620, se = 0.777, z = −2.086, p =  .037). This finding indicated that the effect of Sentence Pair Type on communicative success differed between Balinese and Dutch participants. The proportion of correct responses per sentence pair type and group showed that the Balinese participants had a slightly higher success rate for different‐agent pairs (0.855) compared to the same‐agent pairs (0.816). The pattern was reversed for the Dutch participants, with a slightly higher success rate for same‐agent pairs (0.935) than for different‐agent pairs (0.9). However, additional separate models for each culture group showed that Sentence Pair Type was not a significant predictor of communicative success (see Supplementary Materials 1 and 2). Therefore, the interaction effect was likely caused by the two cultural groups showing opposing patterns.

3.1.2. Difference score

To analyze whether participants differentiated between the types of sentence pairs by using different gestures for the two sentences in the different‐agent sentence pairs, we calculated binomial difference scores on the basis of the four parameters described above: agent type (the type of gesture used to signify the agent), agent location (the location of the agent gesture), verb location (the start location of the gesture signifying the action), and verb path (the path of gesture signifying the action). Below, we provide plots of the difference scores for each parameter coded by sentence pair type and chain over generations (see Fig. 6a for the Balinese results and Fig. 7a for the Dutch results). We also provide plots for the collapsed difference score (see Fig. 6b for the Balinese results and Fig. 7b for the Dutch results). Chains indicated the agent switch for different‐agent sentence pairs by consistent differences in one or more of the four different dimensions:

• ‐agent type (chains 3 and 5 in the Dutch data set)
• ‐agent location (chain 1 in the Balinese and Dutch data set)
• ‐agent type and agent location (chains 2–5 in the Balinese data set)
• ‐agent location, verb location, and verb path (chains 2 and 4 of the Dutch data set)

Fig. 6.

Plots tracking the mean proportion of the different gestures by sentence pair type, chain, and generation for the Balinese data set. (a) Mean proportion of the difference score by parameter. (b) Mean proportion of collapsed difference score.

Fig. 7.

Plots tracking the mean proportion of the different gestures by sentence pair type, chain, and generation for the Dutch data set. (a) Mean proportion of the difference score by parameter. (b) Mean proportion of collapsed difference score.

To test whether Sentence Pair Type, Generation, and Group were predictors for the collapsed difference score, we ran a logistic generalized mixed‐effects regression model. The model included Generation, Sentence Pair Type (dummy‐coded with same‐agent as the intercept), and Group (sum‐coded: 0.5 = Balinese, −0.5 = Dutch) as fixed terms, as well as their interactions (see Table 3). We chose to dummy‐code Sentence Pair Type for this model, as we expected that different‐agent pairs would elicit higher difference scores than same‐agent events, and by setting same‐agent pairs as the reference level, we could directly test this prediction. We kept the Group variable sum‐coded, as in the previous model, to assess how the two cultural groups differed from the overall mean, without choosing one as the reference group. The model also included random intercepts for chain, participant, and item. To improve model convergence, we set the optimizer to “bobyqa” and increased the maximum number of function evaluations to 100,000 using the glmerControl function (lme4 package; Bates et al., ²⁰¹⁵). As expected, the model showed a significant effect of Sentence Pair Type (β = 4.886, se = 1.178, z = 4.147, p < .001), indicating that the difference score was higher in the different‐agent sentence pairs compared to the same‐agent sentence pairs. No significant main effect of Generation or Group was found, and no significant interaction terms emerged as significant.

Table 3.

Outcome of logistic generalized mixed‐effects regression model for the collapsed difference score

	Estimate Std	Error	z value	Pr (>|z|)
(Intercept)	−0.582	0.586	−0.995	0.320
Generation	0.004	0.072	0.057	0.955
Sentence Pair Type	4.886	1.178	4.147	3.36e‐05 ***
Group	0.933	0.614	1.519	0.129
Generation × Sentence Pair Type	0.569	0.380	1.499	0.134
Generation × Group	−0.003	0.134	−0.026	0.979
Sentence Pair Type × Group	−1.925	1.813	−1.062	0.288
Generation × Sentence Pair Type × Group	0.058	0.761	0.076	0.939

Model: Difference Score ∼ Generation ^* Sentence Pair Type ^* Group + (1|Chain) + (1|Participant) + (1|Item). *** p <.001, ** p <.01, * p <.05

3.1.3. Interim discussion

The quantitative results presented above largely replicate the findings by Motamedi et al. (2021). Yet, they do not provide many insights beyond the conclusion that in this type of task both Balinese and Dutch participants manage to communicate successfully by developing referent distinction strategies using different gestures to indicate a switch in agent. To analyze our data in the context of a cross‐cultural comparison, we need to qualitatively examine participants’ behaviors to determine whether the developed strategies differ across cultural groups and, if so, how they differ exactly.

3.2. Qualitative results

Our second objective was to examine the strategies that emerged for referent distinction, that is, how participants distinguished agents in different‐agent sentence pairs. To achieve this, we start from the four parameters distinguished above (agent type, agent location, verb location, verb path). Motamedi et al. (2021) distinguished three different strategies based on different combinations of parameters. Here, we apply the same strategy to characterize our data and compare it to Motamedi et al. (2021) data.

In what follows, we provide detailed descriptions and visualize how changes in different parameters produce distinct strategies (see Table 4 for an overview). We first briefly revisit the strategies discovered by Motamedi et al. (2021) and then introduce the strategies developed by the Dutch participants to illustrate their similarities. Subsequently, we present the strategies produced by the Balinese participants. A detailed cross‐cultural comparison of strategies and an in‐depth theoretical interpretation of the role of space and body attributed across strategies can be found in the Discussion.

Table 4.

Referent distinction strategy types by which parameter indicates an agent change

	Agent type	Agent location	Verb location	Verb path
Lexical strategies	X
Body shift strategies		X	X	X
Body mapping strategies	X	X
Entity classifier strategies		X
Indexing strategies		X

Note. The rows indicate different strategies, and the columns indicate features that were coded; for each strategy type, the required features are indicated. This is compiled from the strategies observed across the Balinese and Dutch chains in this study and the UK chains from Motamedi et al. (2021).

3.2.1. UK strategies (Motamedi et al., 2021)

As discussed in Section 1.4, Motamedi et al. (2021) found that three distinct strategy types emerged in their UK participant sample. Out of the five chains in their study, two chains encoded an agent change on the agent type parameter and were classified as lexical strategies. These strategies employed different gestures for the agents in different‐agent sentence pairs. Two other chains encoded an agent change in agent location, verb location, and verb path. These strategies were grouped together as body (shift) strategies (to avoid confusion with another strategy type, we added “shift” to the original label by Motamedi et al., ²⁰²¹). Participants in these chains indicated an agent change by shifting the orientation and location of their bodies. Lastly, one chain developed a strategy in which the change of agent was encoded in the agent location. Categorized as indexing strategies, participants using this approach employed pointing gestures to distinct locations, indexed in abstract space, to refer to different agents.

3.2.2. Dutch strategies

3.2.2.1. Lexical strategies

Out of the five chains run with Dutch participants, two chains (chains 3 and 5) developed strategies that fall under lexical strategies. When communicating a different‐agent sentence pair, participants used two distinct gestures for the agent type parameter (see Fig. 7a). Specifically, a 1‐handshape (using the index finger) and a 2‐handshape (using index and middle finger) were used to distinguish between the two agents (see Fig. 8a). Initially, first‐generation participants in both chains tried to assign the handshape to one of the agents similarly to a name sign. This resulted in cases where the first agent of the sentence pair was represented by the 2‐handshape. However, by the second generation, the 1‐handshape consistently represented the first agent, while the 2‐handshape signaled a different agent in the second part of the sentence pair.

Fig. 8.

Stills showing the strategies developed by the Dutch chains: (a) lexical strategy, (b) body shift strategy, (c) entity classifier strategy. All stills were taken from the last generation of the respective chain.

3.2.2.2. Body shift strategies

Two chains (chains 2 and 4) employed a change in body orientation and position to signal a switch in agent. These types of body shift strategies employed a difference in agent location, verb location, and verb path to indicate a different‐agent sentence pair (see Fig. 7a). Specifically, participants pushed themselves on their chair to one side of the booth, sideways to the camera, and gestured straight ahead for the first agent. This position was maintained if there was no change in agent. If the agent changed for the second sentence in the sentence pair, they moved to the other side of the booth, again sideways to the camera gesturing straight ahead (see Fig. 8b). Instead of merely indicating a different agent by changing their location, participants changed their orientation to also signal a change in perspective, specifically, facing the perceived recipient or destination on the opposite side of the booth.

3.2.2.3. Entity classifier strategy

One chain (chain 1) developed a strategy encoding an agent change on the agent location parameter (Fig. 7a). This type of strategy employs so‐called entity classifiers (sometimes also called “static size and shape specifiers” or “body part classifiers”), which are morphological tools that use handshapes to represent whole entities, their movement, and their spatial relations (Perniss, 2007; Zwitserlood, ²⁰¹²). The index finger of each hand represented an agent, and a different‐agent sentence pair was indicated by switching the hand used for the classifier (see Fig. 8c). First‐generation participants used a mix of entity classifiers and index points, sometimes pointing upward (classifier) and sometimes downward, indexing a location for the referents in space (i.e., “abstract deixis,” McNeill et al., ¹⁹⁹³; or referential loci in sign languages, Engberg‐Pedersen, ¹⁹⁹³; Perniss, ²⁰¹²). Subsequent generations consistently produced only entity classifiers. In terms of parameters used to encode an agent switch, entity classifier strategies and indexing strategies (as observed in Motamedi et al., ²⁰²¹) look identical. Both employ a location change of the agent gesture to distinguish referents. Based on this, they could be classified as belonging to the same general strategy type. However, we chose to treat them as different strategy types due to the roles they attribute to body and space, respectively. We will address this in more detail in the Discussion.

3.2.3. Balinese strategies

3.2.3.1 Body mapping strategies

Four Balinese chains (chains 2–5) all developed slightly varying versions of the same strategy type. Participants in these chains signaled an agent switch by changing the gesture signifying the agent and the location where this gesture was executed (see Fig. 6a). Participants mapped the agent onto their body by tapping their chest and signaled the other agent somewhere opposite their body, either directly or diagonally in front or to the side of the body.⁷ The location of the agent outside the body was most frequently indicated by a double‐handed pointing gesture, with both hands open and palms facing upward (as in Fig. 9b, top). Only a few participants used a double‐handed index‐finger point (as in Fig. 9b, bottom) and even fewer used a single‐handed index‐finger pointing gesture.

Fig. 9.

Stills showing the strategies developed by the Balinese chains: (a) an entity classifier strategy and (b) different versions of the body mapping strategy type. All stills were taken from the last generation of the respective chain.

3.2.3.2. Entity classifier strategy

The remaining chain (chain 1) developed a strategy nearly identical to the entity classifier strategy observed in one of the Dutch chains. A different agent was indicated by a change in the location of the agent gesture (see Fig. 6a). This was achieved by switching which hand was used as the entity classifier representing the agent and by pointing toward it with the other hand. Similar to the Dutch participants, the Balinese first‐generation participants used outstretched index fingers as the classifier. However, subsequent generations switched to using the whole hand in an open, vertical position (see Fig. 9a).

4. Discussion

Languages employ various strategies that enable their users to differentiate between thematic roles and indicate a change in referent. Yet, it is difficult to trace how these strategies evolved and why they differ across languages. The present study examined the role of cultural biases, such as gestural and spatial conventions, in the emergence of referent distinction strategies in a manual communication system developed by hearing participants. To do so, we methodologically replicated the study by Motamedi et al. (2021), which combined a silent gesture paradigm with interaction and iterated learning to simulate pressures exerted on communication systems during their evolution. Restricting communication to the manual modality in this paradigm allows one to limit the influence of established linguistic conventions. Additionally, manual communication is gestural and spatial; therefore, if conventions in these domains influence referent distinction strategies, we should be able to discover them there. We conducted our study in Bali and the Netherlands, two cultures that differ in some gestural behaviors and, more significantly, predominantly conceptualize space differently.

4.1. Communicative success and referent distinction

Our first objective was to assess whether we could replicate the findings by Motamedi et al. (2021) with Balinese and Dutch participants. Like in Motamedi et al. (2021), our participants (in both cultural groups) could communicate successfully and developed referent distinction strategies that allowed differentiation of agents in different‐agent sentence pairs, resulting in a high overall rate of correct responses. We found an interaction effect between sentence pair type and cultural group, caused by the Balinese participants having a slightly higher success rate for the different‐agent sentence pairs and the Dutch participants performing slightly better in the same‐agent pair trials. However, when analyzed separately, for neither of the cultural groups was there a significant difference in success according to sentence pair type.

Referent distinction strategies were developed across all chains irrespective of which cultural group they belonged to. While Motamedi et al. (2021) did observe an interaction effect between sentence pair type and generation, we did not observe this in our data. The lack of it in our data may be indicative of a ceiling effect: participants across cultural groups communicated successfully and naturally developed strategies to differentiate sentence pair types, even within the first generation from improvisation only.

4.2. Strategy types

For our second objective, we investigated the differences in the referent distinction strategies that evolved across the two cultural groups. Qualitative analyses of the strategies (see Section 3.2) showed that Balinese participants and Dutch participants largely developed different referent distinction strategies, as the body mapping strategy was dominant in the Balinese data but did not occur at all in the Dutch data. Establishing that most of the Balinese chains and the Dutch chains developed different strategies, however, is not enough to suggest that different cultures necessarily develop different strategies. We proposed that if cultural biases were at least partly the cause for referent distinction strategies emerging, we would find differences aligning with different cultural conventions regarding gesture and space. We predicted that the Dutch chains should develop strategies similar to those strategies seen in Motamedi et al. (2021), because in the cultural context of the Netherlands and the UK (as well as in many other Western cultures) egocentric spatial conceptualization is dominant. This is exactly what we found, as four of the Dutch chains developed the same strategies (i.e., body shift and lexical strategies) as four chains in Motamedi et al. (2021).

Merely one Dutch chain did not evolve a strategy observed in the UK data, but instead developed an entity classifier strategy also present in the Balinese data. The entity classifier strategies in the Dutch and Balinese data differed slightly. Participants in the Dutch chain used their hands balled into a fist with only the index finger extended, while the Balinese participants used an open hand, vertically outstretched. This difference is most likely due to Balinese cultural conventions, as open hand gestures are considered polite, especially in situations where the social hierarchy is unclear (Imelda & Harisal, 2024). It should be noted, though, that like the Dutch participants, and despite this convention, Balinese participants employing this strategy used index‐finger pointing to point at the entity classifier with their other hand.

This culturally conventional use of specific gestures can also be observed in the body mapping strategies. While these strategies vary in which axis the “other” agent is placed on (laterally to the side, or diagonally or directly in front), they all indicate the agent mapped on the body by placing one or both open hands on the chest. Imelda and Harisal (2024) state that this is a signal of politeness, whereas an index‐point to the chest would only be used in conversation with a younger interlocutor. Furthermore, open hand gestures are also used to indicate the agent not mapped on the body. Most frequently, this was done with both hands, open and facing upward. While some participants used index‐finger pointing, this was rarely done with one hand alone. Therefore, these findings seem to suggest that culturally specific gestural behaviors that usually appear alongside speech are borrowed in silent gesture, influencing behaviors even when there is no speech present. Unlike structural patterns from spoken languages, whose influence is limited in silent gesture (Goldin‐Meadow et al., 2008), the carryover of gestural conventions may be facilitated by the shared modality.

Thus, we have established that there are differences in the referent distinction strategies between those that have emerged in the Balinese chains and the ones developed by the Dutch and the UK chains. What follows is a discussion on how the developed strategies make use of space and body in different ways and how different conceptualization of space favors certain strategies over others.

4.3. The role of space and body

When communicating information in the visual‐manual modality, body and space function as communicative resources. However, the roles they are assigned may differ depending on what strategy is employed. Specifically, space could just be the physical arena within which gesture is produced (or acted out), or it could be used in a grammaticalized way in which arbitrary locations in space are indicated and meaning, for example, roles such as “agent” and “patient,” are assigned to them to express grammatical relationships (McNeill et al., 1993). Likewise, the body could either simply be the articulator of gestures, or it could adopt the role of one or more animate agents. This means that the body is no longer just an articulator but becomes a grammatical subject, a strategy prevalent across many sign languages (see “body as subject,” Meir et al., ²⁰⁰⁷). The body (of the gesturer or signer), due to its iconicity, can be used to embody grammatical roles. However, to be used in that way, the understanding of gesturer or signer as the literal first person, has to be abstracted to simply represent any grammatical person (Montemurro et al., 2019; Motamedi et al., ²⁰²¹; Padden et al., ²⁰¹⁰). In a similar vein, space can be used in an abstract way to express grammatical relationships rather than literal spatial locations (where the indicated location relates to actual real‐life spatial information). Therefore, we take grammaticalization to mean the trajectory from literal meaning to abstract representation of grammatical properties.

To summarize, the usage of both space and the body can vary gradually between being just the physical context for gesture and being used in a grammatical way. With this in mind, in the following paragraphs, we will systematically discuss the role of space and body across strategies from our experiment and the ones observed by Motamedi et al. (2021). Fig. 10 presents a schematic visualization of their roles. In Fig. 10, we placed the indexing strategy at the top left, to reflect the fact that space is used grammatically, but the body is largely used only as an articulator: participants used referential loci established in the gesturing space in front of them to refer to the arguments. This strategy uses space in a fully grammaticalized way, and the body, when indicating the referent distinction, simply represents the first person, here the gesturer, producing silent gesture.

Fig. 10.

Schematic illustration of the role of body (x‐axis) and space (y‐axis) in the various referent distinction strategies observed in the current study and in Motamedi et al. (2021). Each axis shows whether space and body are used purely as the physical context, in a grammaticalized way, or somewhere in between.

On the other hand, the body mapping strategies produced by participants do not make use of space in a grammaticalized way beyond using it as the physical context for gesturing. The participant's body is abstracted from its literal role as the gesture producer to represent a grammatical concept, often the subject (the agent), but sometimes also the object (e.g., the recipient), of the sentence pair. This strategy is reminiscent of what has been described in sign linguistics as “body as subject” where the signer's body is used to represent the grammatical subject (Meir et al., 2007). To communicate a switch in agent, the other agent is indicated in space opposite the gesturer, but rather than being distinguished by the location in space, it is distinguished by “not being mapped on the body.”

Some strategy types we observed have spatial as well as body‐related properties. For example, body shift strategies employ the body as the grammatical subject, as the participants use their own bodies to represent the agents. At the same time, to indicate a switch in referent for different‐agent sentence pairs, participants moved within space to shift their orientation. In entity classifier strategies, parts of the participants’ bodies, that is, their hands or fingers, were used to represent the agents, but the body as a whole was not used to its full iconic potential as grammatical subject. A contrast in space was used to distinguish referents, but not by using arbitrarily indexed locations in the signing space. Instead, the different agents were anchored to opposing hands. So, although it is in essence a matter of hands pointing upward instead of downward that differentiates the indexing and entity classifier strategies, the latter does not use space in the same abstract way. Therefore, it uses neither space nor the body in a fully grammaticalized way. Because in entity classifier strategies space is used more strongly in a grammatical way than the body, we reflected this in the placement, by placing it closer to the indexing strategy. In a similar vein, we placed body shift strategies closer to body mapping strategies because the body is used more clearly grammatically than space is.

The lexical strategies developed by chains in the Netherlands and the UK do not make use of modality‐specific affordances. For referent distinction, neither body nor space are employed beyond their role as the physical context of gesture. Motamedi et al. (2021) discussed that while the 1‐ and 2‐handshapes could be similar to the aforementioned list buoys, they are not numerical markers for the events in a sentence pair but represent the agents present in the sentence pair.

4.4. The body can replace space to carry grammatical weight

We can now use our conceptualization of the role of body versus space in grammar to analyze the difference between spatial and body‐related strategies. The contrast between indexing and body mapping strategies is particularly interesting. Motamedi et al. (2021) observed the one chain in their experiment that initially used a body mapping strategy switch to an indexing strategy within a few generations. This suggests a developmental shift from prioritizing the body, and its iconic affordance, to prioritizing the potential of a grammatical use of space. This observation is in line with generational changes seen in some natural sign languages (Montemurro et al., 2019; Padden et al., ²⁰¹⁰). However, four out of five Balinese chains developed a body mapping strategy,⁸ yet none underwent this shift despite having the same number of generations as in Motamedi et al. (2021). In other words, body‐based strategies are presented in the literature as a prestage of spatial strategies, but our Balinese participants, despite having the possibility to do so, never shifted in this way. In what follows, we will discuss a potential explanation.

We propose that the persistence of body‐as‐grammar in our data is due to the predominant spatial conceptualization in Balinese culture: people predominantly conceptualize space geocentrically, relying on concrete environmental landmarks rather than egocentric, abstract spatial conventions. As a possible result, our Balinese participants did not develop strategies that require the abstraction of space to a grammaticalized feature. These observations are in line with the absence of spatial modulation in Kata Kolok (de Vos, ²⁰¹⁴), a sign language that evolved in the same cultural context and conceptualizes space in the same way. The absence of a requirement to use space in a grammaticalized way challenges the assumption that spatial modulation has simply not yet developed in Kata Kolok. Given that sign languages younger than Kata Kolok have developed this feature (Meir et al., 2012; Sandler et al., ²⁰⁰⁵), it is perhaps more plausible that the geocentric conceptualization of space inhibits its grammaticalization altogether.

For the four chains in our Balinese data set employing body mapping strategies, instead of using space in a grammaticalized way, the role of the body goes beyond being an articulator. It often represents the grammaticalized subject, similar to the concept of “body as subject” (Meir et al., 2007). However, we suggest that for some of the body‐mapping chains, the use of the body goes beyond representing the subject, taking on the role of a “landmark.” As mentioned above, the Balinese spatial system is anchored on environmental landmarks such as the mountains and seashore (Aryawibawa et al., 2018; Covarrubias, ¹⁹⁵⁰; Dasen & Mishra, ²⁰¹⁰; Wassmann & Dasen, ²⁰⁰⁶). In the absence of any information provided about the agents’ locations within that system, the Balinese participants map one on the body and juxtapose the other. Therefore, the body provides a stable reference point within the communication system, similarly to how environmental landmarks provide fixed reference points in the Balinese spatial system. This raises the question of whether the observed reliance on the body is driven primarily by the lack of spatial cues in the task. If salient environmental landmarks were present in the stimuli, participants might instead have recruited space for referent distinction. However, Kata Kolok, which emerged with the same culturally embedded spatial system and where landmarks are continuously available, has not (yet) developed spatial modulation. This suggests that the availability of geocentric cues may not override the preference for body‐based strategies. Instead, the body itself may take on the role of a landmark within communication.

In earlier generations of the Balinese chains, the mapping broadly resembles “body as subject” with the gesturer's body representing the (first) agent of the sentence pair. Three out of the four chains, however, between the third and the last generation, start to consistently map the same person on the body and use this anaphorically throughout the experiment, even when the person is not the agent in either of the sentences of the pair. This means that, for example, Sarah might be consistently mapped on the body, even for sentence pairs where Hannah is the agent throughout. In these cases, the body no longer represents the agent but instead the patient or recipient. We propose that the body is now no longer simply the subject but also a landmark, anchoring the referents in these strategies. One referent is consistently mapped on the body (the landmark); it is syntactically foregrounded, regardless of its role in the sentence. In some cases, this leads to passive‐like constructions where the body‐mapped referent receives the action rather than performing it (e.g., receiving a book). Importantly, we understand the two functions of “body as subject” and “body as landmark” as not mutually exclusive or in competition, but see the latter as a more generally grammaticalized version where the body can take on different roles while serving as a landmark.

We suggest that this function of the gesturer's body as a landmark may be directly linked to the geocentric conceptualization of space in Balinese culture. In the absence of information on absolute spatial relationships with regard to the fixed landmarks of the Balinese system, the body is recruited as a fixed referential landmark and its grammatical role is adjusted accordingly. Crucially, regardless of whether the body consistently represents one referent, the body mapping strategy developed by the majority of the Balinese chains, in which the body rather than space is recruited as a grammatical device, is effective in distinguishing referents. Therefore, the development from physical (geocentric) space to abstract, grammaticalized space to convey referential information may be unlikely when space is conceptualized geocentrically, and a viable alternative strategy has developed. We propose that this may serve as an alternative explanation for why abstract spatial modulation has not developed in Kata Kolok, a sign language that emerged in the same cultural context.

5. Conclusion

What is the role of the cultural context and biases connected to this in the emergence of referent distinction strategies? Our study addressed this question by providing a comparison of the emergence of referent distinction strategies across two cultural groups. To summarize our results, all our participants (regardless of cultural group) were successful in conveying information about different referents. However, while previous work using silent gesture has highlighted universal patterns across speakers of different languages (Goldin‐Meadow et al., 2008; Hall et al., ^{2013; 2014}), the strategies that emerged in our study look different across chains and cultures. The core differences we observed coincide with the different spatial conceptualizations used across cultural groups.

This is perhaps not surprising, given that space serves as the physical context within which manual communication takes place. Beyond this basic physical role, space can be grammaticalized and used in an abstract way in which agents become associated with different locations in space. However, our findings suggest that Balinese participants who culturally predominantly conceptualize space in an absolute, geocentric way are less likely to develop strategies that use space in such a grammaticalized way. This is also evidenced by the lack of spatial modulation in Kata Kolok, a sign language that emerged in the north of Bali and shares the same spatial conceptualization as our Balinese participants (de Vos, ²⁰¹²). Moreover, our findings indicate that this avoidance of grammaticalization of space could lead to greater reliance on the grammatical potential of the body. Referents are distinguished by whether they are mapped on the gesturer's body or not, creating the contrast of “body” and “not body.” In this way, the grammaticalized use of the body assumes the function that spatial modulation holds in many sign languages, rendering further developmental steps of the use of space, such as its grammaticalization, unlikely in this cultural context.

Furthermore, when the same agent is consistently applied to the body within dyads and even across dyads, then the mapped agent (and the gesturer's body) can be considered an anchor or landmark, similar to landmarks integral to the Balinese spatial system. This, in turn, imparts some of the geocentric spatial information according to which agents can be arranged within the gesturing space in a task that otherwise does not provide any geocentric information. The findings from our study thus suggest that the way space is conceptualized cognitively influences how information in a visual‐manual communication system is conveyed.

Finally, our findings underscore the importance of cultural context in shaping how communicative systems emerge and evolve. They suggest that cross‐linguistic variation may, at least partially, reflect deeper differing cultural conventions. Future work may replicate this paradigm with participants from geocentric cultures outside Bali. It would also be valuable to investigate sign languages other than Kata Kolok that are embedded in geocentric communities to further test the relationship between conceptualization of space and the grammatical recruitment of body and space. In addition, systematically manipulating the availability of geocentric cues within the experiment would help disentangle immediate contextual affordances from more deep‐seated cultural biases that may be shaping these strategies. Together, such work can further improve our understanding of the emergence and diversity of human communication.

Ethics statement

This study was approved by the TSHD Research Ethics and Data Management Committee.

Supporting information

Supporting Information