Introduction

There is general agreement that the evolution of language is best studied as a collection of capacities with sometimes independent evolutionary histories, some relating to production and others to comprehension [1–3]. Since language has left no direct traces in the fossil record, a promising way to study its evolution has been to compare the cognitive and communicative capacities across primates, such as the ability to attribute meaning [e.g. 4–7], to produce signal combinations [e.g. 8–11], to communicate intentionally [e.g. 12–15] or to be socially aware [e.g. 16–18]. The logical foundation of this approach, dating back to Darwin, is that more closely related species have more similar brains than more distant ones, suggesting that their psychological capacities are also more similar [1].

Primates have limited control over their vocal output, which effectively precludes them from developing phonologies and vocal learning [19, 20]. Comprehension abilities, however, appear to be more human-like, with evidence for a general primate capacity to extract meaning from each other’s signals [20–22]. In human language, however, utterance meaning is often ambiguous and goes beyond simple signal-referent relations. To this end, humans deploy ‘pragmatics’ to infer the intended meaning of an utterance, which can deviate much from its literal meaning. This process is cognitively (and computationally) challenging and requires assessments of common ground between speakers and listeners, which can include world knowledge, cultural background, current events and socio-cognitive variables [22–24], much of which becomes accessible as the conversation unfolds. According to Grice [25], meaning in language is not just tied to lexical content but emerges as an interaction with the ongoing conversation with interlocutors taking into account a multitude of information, including context and intentions. Pragmatics, in other words, is at the interface between language and social cognition, requiring speakers to reason about other minds, specifically what information needs to be explicitly encoded, and listeners to make inferences about what the speaker has meant, two major processes likely shaped by both biological and cultural evolution [26].

Animal communication is usually seen as free from such complexities. The classic case is the alarm call system of vervet monkeys (Cercopithecus aethiops), with different predator alarm call types closely tied to corresponding predator encounters, allowing recipients to respond to the calls in biologically adaptive ways [4], which also suggests that alarm calls are meaningful to them [2]. This model of relatively rigid call-referent relations has often been used to interpret findings in primate communication, including calls given to food [9, 27–31], predators [e.g. 7, 32] or during within-group aggression [e.g. 33, 34]. Meaning, in this sense, is situated in the ‘functionally referential’ connection between external events and the signals they trigger.

More recently, it has been argued that this model of meaning is insufficient to cover the full range of phenomena in animal communication [35] and it certainly is insufficient to explain human communication. While animal communication always takes place within a context that is shared by signallers and receivers [36], it is not clear how this impacts on whether or how animals encode and extract meaning [20, 37]. In one chimpanzee study, subjects were able to infer the nature of out-of-sight social interactions from listening to call exchanges [34]. When subjects heard sequences of victim and aggressor screams, they looked longer at incongruent sequences (i.e., low-ranking individual’s aggressor screams and high-ranking individual’s victim screams, hence violating existing dominance relations) than congruent ones (low-ranking victim screams and high-ranking aggressor screams, hence in accordance with the hierarchy), suggesting that they made inferences about third-party interactions they could not see. In other studies with monkeys, subjects responded differently to their own and other species’ alarm calls depending on the additional context provided to them [38, 39].

Inferential (or causal) reasoning has been researched in various domains of animal cognition [40], but not usually in communication research. Current theory posits that only humans are able to represent events in terms of an underlying causal structure, which enables them to make inferences about absent entities beyond simple associatively learned stimulus response relations [41]. Human children are thought to possess a ‘causal map’ of the world, which consists of learned representations of the relations among events in the form of a Bayesian network [42]. Whether or to what degree primates and other animals have inferential capacities of this kind is under-researched and controversial. In one study, for example, chimpanzees were unable to use auditory cues to predict the location of a reward dropped through an opaque conductor [43], whereas in another study subjects from four great ape species correctly identified the location of hidden food with the help of both visual and auditory cues [42]. Similarly, across experimental conditions, great apes were able to select boards that were likely to cover food due to their inclined orientation, either due to abstract inferences or simply having learned that inclined orientations predict food [44].

Our goal was to address great ape inferential capacities in the domain of communication, a topic with special relevance for the evolution of language, especially the origins of pragmatics. We use the term ‘pragmatic inference’ in a broad sense, by referring to an ability to incorporate knowledge about a signaller in order to enrich the interpretation of the signal. Importantly, we do not refer to a key aspect of linguistic pragmatics, i.e., the ability to infer communicative or informative intention [45]. Instead, we sought to investigate more broadly whether bonobos were generally able to enrich the meaning of calls by adding knowledge related to call producers, a likely precursor of the human ability to form common ground and other higher forms of social cognition. Common ground refers to the shared mental state that two interlocutors, as signaller and recipient, establish and maintain during a conversation. Here, we only looked at the recipient side, not at the interactive process underlying the formation of common ground.

We chose food calls because this class of vocalisations has already been investigated in previous studies. When interacting with food, bonobos produce sequences of acoustically distinct vocalisations, depending on the perceived quality of food [9, 27]. In particular, sequences to highly preferred foods typically, but not exclusively, contain barks and peeps, whereas sequences to less preferred foods typically, but not exclusively, contain peep-yelps and yelps [27]. The different call types are graded and sequence composition is highly variable [9, 27]. Despite this high level of plasticity in sequence composition, recipients are able to make inferences about the type of food available [9]. In one analysis it was found that duplications of call types within a sequence appeared to be especially influential on how recipients assign meaning [46], but this needs to be tested with targeted experiments. Previous research has also shown that bonobo food calls, like many other primate vocalisations, are individually distinct [9, 27, 47], with experimental evidence that individuals recognise each other by their call sequences over long periods of time [48].

In our study, we first trained two group members to have opposing food preferences for two artificially coloured food items, pink or blue chow, such that KEL preferred blue and DW preferred pink, while a third group member, LNG, served as an untrained control. In previous research we have established that bonobos can develop socially learned preferences for familiar foods with changed visual appearance from artificial colouration, even if they otherwise taste and look the same [49]. After having observed KEL and DW executing their individual preferences for blue and pink chow, respectively, we then asked whether other group members took these individual demonstrator preferences into account when making their own foraging decisions. To test this, we designed a playback study, using recordings of KEL, DW and LNG’s food calls. We predicted that, if subjects’ foraging decisions varied depending on who was the demonstrator, then this would be relevant evidence for evolutionary theories of pragmatic inference. Understanding that other individuals have food preferences that deviate from one’s own is akin to understanding something about their mental states and allowing predictions about their intentions. Importantly, in this sense the current study goes significantly beyond what has been demonstrated, i.e., that chimpanzees and bonobos can make inferences about the presence of generally preferred and generally not-preferred foods [9, 50].

Method

Phase 1: Preference demonstration phase

In a previous study, we established that bonobos could learn individually distinct, arbitrary food preferences of other group members through rapid social learning and by mere observation [49]. In both the previous and the current study, KEL and DW served as demonstrators who preferred either blue (KEL) or pink (DW) chow. During preference demonstrations, artificially dyed blue or pink chow was offered to the demonstrators manually and in full view of the subjects (Fig 1). When demonstrating their preferences before others, both individuals showed very clear preferences for their assigned colours (first choices: KEL–blue 99.6%, DW–pink 100%), providing subjects with an unambiguous learning opportunity.

Download:

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

Fig 1. Layout of indoor and outdoor enclosures for the three experimental phases.

Phase 1 –observational learning: demonstrators (KEL, DW) were held individually in cage 1 (demonstration room) and given a choice of pink and blue chow. Subjects were able to observe this process from across the corridor in cage 2 (observation room). Phase 2 –foraging training: blue or pink chow was provided to the subjects in one of two adjacent food troughs, 4 metres apart and equidistant from the point of entry (trapdoor). Phase 3 –playback experiment: a loudspeaker was placed at midpoint between the two food troughs. The demonstrator whose calls were to be played back was lured away to the large island so as to be out of earshot, while subjects were required to be on the small island within 5 metres of the trapdoor. During each trial, the experimenter (GS) positioned herself at midpoint between the two troughs to document the subjects’ foraging behaviour.

https://doi.org/10.1371/journal.pone.0267574.g001

During the experimental period 24 preference demonstrations were carried out by each demonstrator. An initial demonstration phase, consisting of three demonstration days for each demonstrator, took place in April 2015, a reminder period involving four demonstration days for each demonstrator ensued in June 2015. Following this, refresher preference demonstrations took place approximately every 3–4 weeks until the experiment was over. Preference demonstrations consisted of the experimenter (GS) placing two chow pellets of each colour (i.e., four total) on two identical white plastic trays (20 x 20 cm), placed side by side against the bars of the cage, allowing the demonstrator to select them using either his fingers or lips (S3 Fig). The position (left or right) of blue or pink food items was counter-balanced and randomly determined, with the restriction that a given colour could not be presented on the same side for more than three consecutive trials (for supplementary information see S1.3 Text in S1 Text).

Both the demonstrators and the observers were filmed to document demonstrator choices and observer attention (PANASONIC HC-V100 and PANASONIC HC-V727 full HD cameras). Subjects observing from outside the camera range were included with help from a trained animal keeper. Observer attention was defined as head and eyes oriented towards the demonstrator whilst a choice was made and coded as attending (1) or not attending (0) [49].

Results

As the group was large with individuals free to move around, meeting all required playback conditions was an extremely challenging task, resulting in a total of N = 12 playback trials (N = 6 in the test condition, i.e., KEL’s food calls, indicating blue preference; and N = 6 in the control condition, i.e., LNG’s food calls, indicating no preference) over a period of 13 months (Sept 2015 –Oct 2016). We managed to test N = 10 subjects (3 males; 7 females, age range: 5–46 years; S1 Table) with N = 9 completing at least one trial in both conditions (test: N = 10, control: N = 9). This was because LNG as the call provider could not participate as a subject in the control condition. Two infants (MO, KLS) also took part but were excluded as they were still dependent on their mothers. The average number of subjects per trial was 4.5 (median = 4.5, range = 1–8, N = 12 trials).

For ‘expectation’, we found that subjects generally spent little time looking towards the experimenter, regardless of experimental condition. The full model was not significantly different from the null model (Generalized linear mixed model, likelihood ratio test (LRT): Chi² = 4.20, d.f. = 3, P = 0.2409, S8 Fig, S4 Table), suggesting that (contrary to predictions) the playback stimuli did not trigger obvious and systematic expectation towards the experimenter.

For ‘exploration’ (i.e., peering and probing analysed together), results were in the predicted direction when comparing the full and null models (LRT: Chi² = 6.97, d.f. = 3, P = 0.0728, Fig 2, S5 Table). In the test condition (KEL calls, indicating blue preference), individuals were overall more likely to explore the blue (correct) trough compared to the pink (incorrect) trough, but this was due to three individuals (DV, LNG, NK), who performed in line with predictions. It should be noted that the significance of the statistical model was largely due to one individual, NK (see S1 and S2 Files: removing NK has the strongest effect on the model). The remaining individuals were either more likely to explore the pink trough, or showed the same number of responses to both troughs.

Download:

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

Fig 2. Exploration.

Probability of subjects exploring the blue or pink food troughs in test and control conditions. Circles represent the average probability to explore for each subject (10 subjects participated in the 6 test trials and 9 subjects participated in the 6 control trials). Squares with bars represent model estimates with 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0267574.g002

In the control condition (LNG calls, indicating no food preference), individuals were overall equally likely to explore both the blue and the pink trough. The majority of subjects (6 of 9) were in line with our predictions by showing similar interest in both troughs.

We then analysed the components of exploration separately and found that ‘probing’ was generally rare, with only 5 of 10 subjects exhibiting the behaviour (see S1 and S2 Files), which prevented us from carrying out further analyses. For ‘peering’, however, we found a significant (undirected) trough bias in the test condition relative to the control condition (Fig 3): Seven of nine subjects showed a trough bias (peering more in the blue or pink trough) after hearing KEL’s calls, while only three of nine subjects showed a trough bias after hearing LNG’s calls (KEL: blue: N = 3, pink: N = 4; LNG: blue: N = 1, pink: N = 2, see S9 Fig; Wilcoxon signed rank test, N = 6 (2 ties), V = 21, P = 0.036, two-tailed).

Download:

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

Fig 3. Peering frequency.

Mean frequency of peering across individuals at the blue and pink food troughs in test and control trials. Circles represent mean frequencies of peering across individuals (10 subjects participated in the 6 test trials and 9 subjects participated in the 6 control trials). Squares with bars represent medians with 25% and 75% quartiles. Subjects in the test condition show higher bias for one of the two troughs than subjects in the control condition.

https://doi.org/10.1371/journal.pone.0267574.g003

We encourage the reader to further explore our raw data (S1 Data) and analyses with the online interactive tool provided in the supplemental information (S1 and S2 Files).

Discussion

The aim of this study was to investigate whether bonobos were able to draw inferences from knowledge about other individuals’ food preferences previously acquired by observational learning. The design was based on the species’ demonstrated capacity for individual vocal recognition [48] and their documented propensity to vocalise when encountering highly valued food, such as chow [9, 27]. We predicted that, when hearing the calls of a demonstrator with a specific and previously demonstrated food preference, subjects should explore the location where the demonstrator’s preferred food was usually available (see Fig 1: left trough for blue chow; right trough for pink chow). We also predicted that, when hearing the calls of another individual, with no known food preference (LNG), subjects should not show such a bias.

To test our hypothesis, we provided subjects with social learning opportunities. First, subjects observed that two of their group members (KEL, DW) had opposing preferences regarding coloured chow while a third individual (LNG) served as a no preference control. Subjects also learned that blue and pink chow were available in fixed locations, but never simultaneously. Once this was established, we simulated food was available in inside by playing back food calls by KEL (blue preference) or LNG (no preference).

We monitored subjects upon entry and did not find a clear preference for approaching one trough over the other, neither in the experimental nor the control condition. Upon arrival at a trough, we measured foraging behaviour, exploration and expectation, and found an effect for exploration in the predicted direction (Fig 2), although there was considerable inter-individual variation. When analysing peering and probing separately, we found that for most subjects peering was significantly biased towards one trough (blue or pink) when responding to KEL’s calls, but not when responding to LNG’s calls (Fig 3). Interestingly, the highest peering rates were in the experimental condition, i.e., when responding to blue-trained KEL’s food calls, whereas peering rates to untrained LNG’s food calls were substantially lower. This may be due to the fact that subjects could make a prediction with KEL’s calls, and may have approached the troughs in a more determined and anticipatory way, whereas this was not possible in response to LNG’s calls.

From these data we concluded that at least some subjects recognised KEL as someone with a chow colour preference and LNG as someone without such a preference. Three subjects (DV, LNG, NK) behaved as if they not only knew that KEL had a colour preference but also that it was a preference for blue chow. Clearly, the task in this experiment was complex, requiring subjects to form various associations about the food preferences of others and knowledge about where and when to find the corresponding foods. First, subjects had to associate the identity of a demonstrator to his preferred chow colour, then to associate the two chow colours to two trough locations, before finally being able to make an inference that a demonstrator’s food calls were indicative of the availability of a specific food at a specific location. Although subjects were provided with extensive opportunities to learn these associations, our data suggest that not all animals were able to form them or to recall them during the playback experiments. For example, rather than remembering KEL as someone who preferred blue chow, some subjects may have formed a less sophisticated representation, i.e., that KEL had some colour preference, but not for which one (see results). It is also possible that some individuals simply did not remember where the blue and pink foods were provisioned. Although they were given reminders (4–5 per month, food colour and location) it is not certain that they were able to track these locations over time. To address this hypothesis, it would have been necessary to carry out another test, but this would have further increased the complexity of the experimental design. In any case, individual differences in cognitive capacities are well documented in great apes [72] and this may have led to some individuals not learning the colour/location association. Naturally, there may have been other reasons to explain our finding, such as differences in attention during the demonstration phase or carryover effects of the refresher baiting routine. Nevertheless, the fact that KEL’s food calls caused a foraging bias, whereas LNG’s food calls did not, is in line with the hypothesis that bonobos are able to draw basic inferences about others’ food preferences, although there appear to be individual differences in the complexity of the underlying mental representations.

How do these results compare to other research on primate call comprehension? As mentioned, field experiments with Old World monkeys have demonstrated that non-human primates can take context into account when responding to each other’s calls, across call types [38, 39, 73]. However, in all previous studies the context was always the same for the caller and the recipients, in contrast to the current study. This is also true for the previous playback study on bonobo food calls [9], where there was no disagreement between call providers and their audiences regarding the referents, since all subjects preferred kiwi over apples.

Conclusions

While we acknowledge that the task was complex, our data suggest that at least three individuals performed in line with the hypothesis that bonobos are capable of some pragmatic inference, or at least an essential component of this capacity. Other subjects also behaved as if they recognised the call provider as someone with a known food preference (KEL), which differed from their responses to another call provider with no known food preference (LNG). We take these results to suggest that—at a very basic level—bonobos take others’ individual preferences into account, even if they differ from their own, although our study does not reveal much about the content of this knowledge.

If these results can be confirmed by future work, they have clear implications for current theories of language evolution and the debate about the origins of Gricean communication. Data also bear relevance to the question of whether pragmatic competence was a precursor to language or co-evolved with it. The coevolution hypothesis rests on the assumption that language and social cognition, particularly mindreading, emerged together as suggested by simulation experiments [74]. The precursor hypothesis is that only humans possess socio-cognitive abilities required for Gricean communication and that they operated as precursors for language evolution and ontogenetic pre-requisites for language development [75, 76]. A more moderate view is that the socio-cognitive abilities necessary for language evolved gradually, reaching beyond the last common ancestor. Human pragmatic abilities, in this view, are mere improvements of more basic socio-cognitive abilities, which now operate in conjunction with significantly increased processing power to enable language [77]. On the production side, other-mind awareness, partially demonstrated by this study, is ostensive communication (pointing, showing), which is difficult to demonstrate in great apes [78]. Our data are in line with the notion that Gricean pragmatics evolved gradually, starting with an early ability to make inferences about how others operate, an ability that must have emerged before chimpanzees, bonobos and humans split into separate lines.

Supporting information

Acknowledgments

We thank “La Vallée des Singes” and its director Emmanuel Le Grelle as well as “Le Conservatoire pour la Conservation des Primates” and its president Jean-Pascal Guéry for allowing access to the study site and subjects. We thank Carole Michelet, Franck Alexieff, Lise Morel, Jérémy Mergault and Alexandre Albert for their valuable help in carrying out the experiments. We thank Josep Call, Redouan Bshary, Heidi Lyn and Maggie McGonigle for useful comments on the study design, data analyses and interpretation of results.