2.1 Oceanic Possession

Oceanic possession is widely discussed in the areal literature (Lichtenberk 1985; 2009; 2018; Ross 2004; and Lynch et al. 2002, among others);¹ we give a brief overview of the possessive constructions found in the six sample languages in our study. These languages show a split in the class of nouns: some allow suffixation of a possessor index directly, whereas others cannot be directly suffixed. This two-way split is commonly termed ‘direct’ and ‘indirect possession’ in the Oceanic literature. In the typological literature it is called the alienability split (Bugaeva et al. 2022) or obligatory/optional possession (Bickel & Nichols 2013).

The rationale for this differing terminology is that nouns that are directly possessed tend to be semantically inalienable (including body parts, kin terms, part-wholes, and personal possessions). Nouns that are indirectly possessed are more likely to be easily transferable from one possessor to another. The following example from Iaai shows a direct possessive construction ((3)a) with a kin term directly suffixed by a possessor suffix. Example ((3)b) shows an indirect possessive construction where the possessor suffix does not attach directly to a possessed noun, but instead to one of a set of possessive classifiers.

It is the possessive classifiers, as in ((3)b), which are our focus. The number of such possessive classifiers varies considerably; the number found in each of our sample languages is shown in Table 1.

The possessive classifiers in Oceanic languages are often called relational classifiers (Lichtenberk 1983), since they encode the relation (intended or actual) between the possessor and possessed, rather than a particular salient attribute of the referent of the possessed noun. Hence, they are informative. Nouns can often occur with different classifiers, depending on the relationship that the speaker highlights. Example (4) from Iaai shows that the noun wââ ‘fish’ can occur with three different possessive classifiers depending upon the interactional context. This same noun can occur with the food classifier if the speaker intends to eat the fish; with the catch classifier if they want to highlight that they caught the fish; or with the general classifier if they want to highlight another use not specified by the available set of classifiers in their repertoire, for example if they wish to sell the fish.

Both the New Caledonian languages in our sample, Nêlêmwa and Iaai, have more ‘open’ systems of classifiers; speakers are able to create new classifiers when a particular need arises. This is a productive process in Iaai and Nêlêmwa. New classifiers can emerge from existing directly possessed nouns, or from verbs by attaching possessor suffixes. To take an example from Iaai, the verb tûö ‘to smoke’ can be used as a classifier for possessions which you can smoke by the addition of possessor suffixes, e.g., tûe-n sigââ ‘his/her cigarette (smoke.cl-3sg cigarette)’ or tûe-n paep ‘his/her pipe (smoke.cl-3sg pipe)’. The Vanuatu languages in our sample do not allow the creation of new classifiers; their class of classifiers is closed.

In contrast to the widely held view that Oceanic possessive classifiers are relational classifiers, the North Ambrym language has a much stricter relation between noun and classifier, with the majority of nouns only being able to occur with just one classifier. For example, we ‘water’ in (5) can occur only with the drink classifier, regardless as to whether the possessor intends to drink the water or not:

Across the Oceanic languages, the directly possessed nouns tend to be a closed class, disallowing new members. In contrast, the class of indirectly possessed nouns is open, allowing new words, such as borrowings, to be possessed. Across the sample languages, the Vanuatu languages fit this description well. However, the two New Caledonian languages appear to have a semi-open class of directly possessed nouns. In Iaai, indirectly possessed nouns and verbs can be a source of new directly possessed nouns. Equally, directly possessed nouns act as a source for new possessive classifiers, which are an open class in Iaai.

Possessive classifiers are not homogenous within and across languages. We make a distinction between major and minor classifiers. Major classifiers occur with a large number of nouns and can allow neologisms or borrowings. Minor classifiers typically occur with only a handful of nouns and are essentially closed to new ones. North Ambrym has two minor classifiers: fire and basket. The latter classifier allows just a handful of nouns that refer to traditionally made baskets, whereas borrowings such as bak ‘bag’ are disallowed, and instead occur with the major general classifier. As minor classifiers occur with only limited numbers of nouns, there is a possibility that these are more likely to diminish in use. For example, speakers of North Ambrym are happy to possess arrbol ‘basket’ with either the basket or the general classifier with no change in semantics. Similar examples can be found in Iaai and Nêlêmwa, which have varying numbers of minor classifiers, with a shift in use to major classifiers.

The possessive classifiers in the languages of Vanuatu in our sample show signs of heterogenous semantic domains. Thus, the labels given to the classifiers are mnemonic devices, which refer to commonalities in the semantic domain across languages. For example, the food classifier in North Ambrym can also occur with nouns referring to tools, utensils, crockery, units of time and some kinship terms. The drink classifier in North Ambrym, Vatlongos and Lewo also occurs with nouns referring to buildings and mats. In Vatlongos and Lewo the drink classifier also encodes some items of clothing. Generally, the possessive classifiers in the languages of New Caledonia in our sample cover more homogenous semantic domains, due to the large number of classifiers that have carved up the semantic space in more fine detail. However, there are some classifiers which do have heterogenous semantic domains. The animal classifier in Iaai is typically used for domesticated animals, though some speakers also use it for domesticated trees and plants they look after.

Both North Ambrym and Nêlêmwa also have a further indirect construction which uses prepositions to link possessor and possessed. We include this here, since these indirect prepositional possessive constructions occur in speakers’ responses to some of the storyboards. In North Ambrym the general associative preposition ne can encode a looser connection between possessor and possessed. Schneider (2011: 385) discusses similar constructions across other languages of Vanuatu, describing them as “where two NPs are juxtaposed in a difficult to define, not-quite-possessive relationship.” Often these constructions can be used when the ‘possessor’ is inanimate, such as rrem ‘yam’ in example (6), where it is not an intrinsic part of saaroan ‘story’, but it is what the story is about:

Furthermore, the general associative preposition in North Ambrym can be used for some part-whole relationships as well, such as internal body parts (7).

Schneider (2011) discusses the potential difference between indirect and associative constructions as being defined by control, where associative constructions encode a lack of control of the possessor over the possessed.

Nêlêmwa, unlike the other sample languages, does not have a general possessive classifier. Instead, Nêlêmwa³ uses an indirect linker i shown in example (8):

In Nêlêmwa, we have found that some directly possessed nouns can also be possessed using the general linker preposition (though this is speaker-dependent), thus making the use of the linking preposition not restricted to a particular class of noun:

2.2 Reference Tracking

In the sample languages, possessive classifiers can function as reference tracking devices. In these instances, possessive classifiers can be used without a possessed noun. The possessed noun is either recoverable from prior discourse or evident from context. In example (10) from Vatlongos, a section of discourse between a pigeon and a heron on the reef, the first instances of the a- food classifier occur with the possessed noun tiei ‘clam’ ((10)a). However, the second occurrence of the food classifier occurs on its own without the possessed noun ((10)c). Here, the classifier functions as a reference tracking device, and refers back to the first mention of tiei ‘clam’.

In our sample, five of the six languages allow full possessed noun elision when the referent of the possessed noun in a possessive classifier construction is recoverable from prior discourse or from pragmatic/external context. However, North Ambrym allows only partial elision of the possessed noun of a classifier construction. In example (11), instead of the possessed noun, the partitive marker ge (ptv) must occur. The antecedent of the possessive phrase is bwehel ‘bird’ mentioned first in the sentence. The possessive phrase is interpreted as ‘his (food) one’.

In the available corpora of Vatlongos and North Ambrym, the use of classifiers as reference tracking devices is rare. As both languages have limited documentation, they have small corpora: Vatlongos has a total corpus of 51,302 tokens and North Ambrym 31,788 total tokens. Table 2 shows that the classifier constructions only make up a small part of the total number of indirectly possessed noun tokens in both corpora.

3.1 Participants

The Participants were 125 adults over the age of 18 from the language communities in Vanuatu and New Caledonia. A power analysis was conducted using G*Power (v3.1.9.7), ensuring the sample is appropriate for detecting small-medium effects (Faul et al., 2009). The sample included 68 (54.4%) male and 57 (45.6%) female participants. Their ages ranged between 18 and 80 years (M = 44.37, SD = 15.46). The level of education was measured in years of formal education and ranged between 0 and 23 years (M = 8.89, SD = 4.00). Further participant information broken down by language is shown in Table 3. Participants were recruited using convenience and snowball methods via contacts in each language community. Favourable ethical opinion was obtained from the University of Surrey prior to the study.

3.2 Experimental Setup

The storyboard experiment was conducted alongside two other experiments – a possessive labelling task and an eye-tracking task. Only the storyboard experiment will be discussed here; the other two experiments will be reported elsewhere. The order of experiments was randomised across participants. Testing took place across 56 days in the six language communities during July to October 2022. One to four participants were tested on each of the data collection days. All participants were provided with information about the study and gave informed consent. For participants in Vanuatu, the information sheet and consent form were read aloud in Bislama (the lingua franca of Vanuatu) and oral consent recorded. Participants in New Caledonia read the information sheet and signed written consent forms in French (the lingua franca of New Caledonia). For Iaai and Nêlêmwa, the experiments were conducted either at a participant’s home on a veranda or underneath a shelter, or in the building where the experimenter was staying. For North Ambrym, Vatlongos, Lewo, and Merei, the experiment was set up in a house on a table. Participants were reimbursed for their time with small monetary payments (Vanuatu) or gifts (New Caledonia), following the different cultural norms.

3.3 Materials and Design

Eight storyboards were designed to investigate how different interactional contexts affected the use of classifiers as anaphoric reference tracking devices across stretches of discourse. Each storyboard consisted of four images that represented different interactions related to a target noun (see Figure 1).

Figure 1

Storyboard targeting ‘fish’ with different interactions: catching (top left), carrying home (top right), eating (bottom left), and selling (bottom right).⁵

A representative sample of nouns were chosen to test across all six languages, combining variation in frequency and variation in cognitive salience. Corpora exist for North Ambrym and Vatlongos (c.f., Table 2); nouns with similar semantic definitions that appeared in both corpora were chosen as a starting point, resulting in 116 nouns. These nouns were cross checked with results from a prior free-list experiment (see Franjieh et al. 2022 for methodology). The free-list experiment elicited the most salient nouns associated with each classifier in each language. Cognitive Salience Index (CSI) scores (Sutrop 2001) were calculated to give an indication of the psychological saliency of the association of each noun to a particular classifier. CSI scores combine the frequency and mean position of a term within a list, across all participants. They range from 0 to 1, with 1 denoting maximum salience. The maximum CSI score of each noun across the sample languages was used to divide the list of 116 nouns into four ranks (>0.75 = 1; >0.5 and <=0.75 = 2; >0.25 and <=0.5 = 3; <=0.25 = 4). Token frequency was normalised per 10,000 words and averaged across both the North Ambrym and Vatlongos corpora. The list of 116 nouns was then divided into four frequency quartiles.

Abstract nouns, nouns referring to large entities (e.g., home, cliff), superordinate terms (e.g., animal, thing, fruit), and nouns relating to people and their relationships (e.g., old woman, teacher) were removed. This was because abstract nouns, superordinate terms and large entities are hard to create concrete images for. In some languages, people and relationship terms are possessed using the direct construction, rather than with a classifier. Equally, these nouns are hard to fit into different interactional contexts.

The remaining list of 66 nouns was used as a basis for developing the eight storyboards. Two nouns from each CSI rank were chosen, which were mapped onto being either high frequency (top two frequency quartiles) or low frequency (bottom two frequency quartiles) nouns. However, the third CSI rank nouns all corresponded to high frequency nouns (top frequency percentile) so it was not possible to fully choose a high and a low frequency noun for the third CSI rank. Nevertheless, this technique enabled us to select eight nouns ranging in cognitive salience and frequency. The order of the eight storyboards (c.f. Table 4) was randomised for each participant.

3.4 Procedure

Prior to data collection, experimental protocols were translated into Bislama and French and were checked by a French and a Bislama native speaker. The protocols were then back translated into English to validate the translations. Participants were first read the instructions below in either Bislama or French. Participants were encouraged not to repeat the possessed noun in the stories, so as to elicit classifiers used anaphorically without a possessed noun. Participants were shown the story on a laptop screen. The story was presented one picture at a time and the Bislama/French text read out. Then the storyboard was presented again picture by picture and the participant was asked to retell the story in their language. Often the French or Bislama version was repeated during this time as a prompt. Answers were recorded directly onto the laptop underneath each picture.

Instructions

There are eight short stories to look at. I will explain each story first in Bislama/French and then I would like you to retell the story in your language. Each story has four pictures. For each picture I want to know how you say that the object in the picture belongs to the man or the woman in the picture.

After the first picture you do not need to repeat the name of the object, just the word meaning ‘his’ or ‘hers’. For example, if the first picture is of a man and a woman making coffee, you can say ‘they are making their coffee’. If the next picture shows the man drinking the coffee, you should say ‘he is drinking his one’ but not ‘he is drinking his coffee’.

Some stories will sound a bit strange, but I want you to translate them even though they sound strange.

I will write down what you have said and ask you to help translate it. I will record what you say on the audio recorder.

The instructions explicitly asked that, after establishing the referent of the possession in the first picture using a noun, the participant was to avoid further uses of that noun and instead use just the possessive classifier, if possible. If a participant gave a response that deviated from expected, they were prompted to reformulate, using a classifier without a possessed noun. If a participant only used a noun as the referring device in the second, third or fourth picture, they were asked if it was possible to replace the noun with just a possessive classifier. Similarly, if a participant gave a full possessive construction consisting of classifier and possessed noun, they were asked if it was possible to say the sentence again, without the possessed noun. The goal of the experiment was not to uncover whether possessive classifiers can occur on their own without a possessed noun as a reference tracking device, but whether the classifiers used as reference tracking devices would vary according to the interactional contexts shown by the different pictures of the storyboards. All responses were recorded from the participants.

4.1 Classifiers used Anaphorically

We isolated all responses where participants used the possessive classifiers anaphorically without a possessed noun. We did this to track the use of classifiers across discourse units to see whether they matched or differed from their antecedent. The following types of responses were excluded: (i) no possessive construction, (ii) a directly possessed noun, (iii) a preposition, or (iv) a classifier with a possessed noun. This is a strict view of the use of classifiers as reference tracking devices, since we excluded classifiers that occur with a possessed noun. Then in section 4.2 we offer a less strict analysis, which includes all possessive constructions.

A minimum of two classifiers used anaphorically per storyboard were required in order to assess whether participants used a matching or non-matching strategy. The first picture of each four-picture storyboard sets the scene and introduces the target noun. Participants were not required to use a possessive phrase to introduce the target noun. However, some participants did use a classifier construction, which we counted in our analysis. In the subsequent three pictures (pictures 2, 3 and 4), only anaphoric uses of classifiers were counted. The amount of data required to determine if the anaphoric uses of classifiers matched an antecedent was either:

1. A minimum of two anaphoric occurrences of classifiers for pictures 2, 3 and 4. (see Table 4 for the interactional contexts for each picture).

2. A minimum of one anaphoric occurrence of a classifier for pictures 2, 3 and 4 and a classifier used for picture 1.

We give two glossed examples from the experiment below from the fish storyboard. Example (12), from North Ambrym, shows the introduction of the noun myaalo ‘fish’ in the first picture. The possessive classifiers in pictures 2–4 are all used anaphorically within the text to refer back to the antecedent myaalo ‘fish’. The North Ambrym example shows a matching strategy, as all classifiers are the same, regardless of the contextual interactions.

Example (13), from Iaai, shows the noun wââ ‘fish’ being introduced for the first picture along with the food classifier. The further three instances of classifiers are all used anaphorically without a possessed noun and refer back to the antecedent wââ ‘fish’. The Iaai example shows a non-matching strategy as different classifiers are used across an individual storyboard.

Table 5 shows different permutations of dummy data to explicate the permissible responses accepted for analysis detailed above, where CL indicates ‘classifier with possessed noun’; ANA indicates ‘anaphoric use of a classifier’; and NO indicates ‘no possessive construction’. A classifier with a possessed noun was only included for analysis if it occurred as a response for the first picture as this was acceptable to introduce the target noun and function as the antecedent for future anaphoric uses of classifiers. Any further instances of classifiers with possessed nouns were deliberately omitted in this analysis (but see section 4.2 for an analysis where they are included).

Table 6 shows the different classifiers used anaphorically across the sample languages. One important by-product of this methodology is that we were able to describe three different classifiers that had previously not been documented (marked by an asterisk in Table 6).⁷

Responses were coded into a dichotomous variable to indicate one of two strategies: either matching classifiers used anaphorically across a storyboard, or non-matching (different) classifiers used anaphorically across a storyboard.

Figure 2 shows the mean percentage of matching classifiers used anaphorically across all storyboards by language. North Ambrym, with over 90% mean matching responses, shows a different pattern to the other languages in the sample. This supports previous research illustrating that North Ambrym has largely fixed assignment of classifiers (Franjieh 2018), in contrast to the other five languages. We investigate this further to establish whether these five languages show statistically significantly different patterns to North Ambrym.

Figure 2

Bar plot showing the mean percentage of matching anaphoric strategies for each sample language (total number of classifiers across all participants for each sample language given in brackets). Error bars show the 95% confidence intervals.

There were fluctuations in the number of matching strategies according to storyboard across languages (c.f. Figure 7). To account for these differences, we ran a binomial generalised linear mixed model with random intercepts.⁸ The dependent variable was the matching/non-matching strategy for a storyboard, with language as the fixed effect. The model had two random effects – individual storyboard and participant. Models were created in R using the glmer function from the lme4 package (Bates et al. 2015). Table 7 gives an overview of the results. R² values were calculated using the MuMIn package in R (Bartoń 2024).⁹ There were 870 observations, 124 participants and 8 storyboards. Iaai was used as the reference level for the fixed effects of language.

The model shows that the dependent variable is more often 1 (a matching strategy) in North Ambrym than in Iaai, and there is no significant difference between any other languages and Iaai. The model shows that North Ambrym is a significant predictor of matching reference tracking strategies. The model’s R² values show that — when including the random effects of storyboard and participant — the model can account for 43.3% of the matching strategies.

4.2 All Possessive Constructions

Although we asked participants to give an anaphoric use of a classifier after the target noun was introduced, not every participant did so, even after prompting. This resulted in some instances where the minimum amount of data for analysis was not met (c.f. section 4.1), especially for the two New Caledonian languages (Iaai and Nêlêmwa) shown in Table 8. These missing data were due to several reasons. First, some participants saw each picture within a storyboard as an individual scene to be described, rather than as part of a continuous story. In these instances, participants preferred to give a classifier with a possessed noun, rather than use a classifier anaphorically without a possessed noun. Second, some missing data stemmed from participants using a directly possessed noun or avoiding a possessive phrase by giving just the noun without a classifier. Finally, in the case of Nêlêmwa, there is no general classifier, instead a general preposition is used, which must occur with a possessed noun (c.f., example (8)).

In Nêlêmwa we found examples where different participants could use a particular noun with either possessive classifiers or with the general linking preposition. For the yam storyboard, one participant used kuvic ‘yam’ only with the linking preposition: e.g., kuvi i ye (yam link 3sg), whereas another participant varied in choice between the ââ- plant classifier or the caa- starch classifier for kuvic ‘yam’. Another example from Nêlêmwa involves wi ‘water’, which could be either directly possessed wi-n ‘water-3sg’ or referred to anaphorically with the drink classifier, kêâ-, even by the same participant.

In order to account for the missing data, we investigated the use of possessive constructions across the four pictures in each storyboard regardless of their type. The following types of constructions were included for comparison:

• Possessive classifiers used anaphorically without a possessed noun

• Possessive classifiers with a possessed noun

• Directly possessed nouns

• Prepositional possessive constructions

The analysis sought to discover whether the same strategy (matching vs. non-matching) was used across the four pictures of a storyboard. As we include all types of possessive constructions, this led to an increase in the number of categories used for the comparison for the following languages, in addition to those given in Table 6:

• For Iaai, participants also used directly possessed nouns and two previously undescribed classifiers: wash and bathe.¹⁰

• For Nêlêmwa, participants also used directly possessed nouns, and the general prepositional possessive construction (example (8)).

• For North Ambrym, participants also used an associative preposition (example (6)).

Table 9 shows the most frequent possessive construction used across each storyboard. Speakers of Merei have a preference for the general and consumable classifiers — the two main classifiers in the language. Speakers of the three languages of Central Vanuatu — Lewo, Vatlongos and North Ambrym used the food or drink classifier more than other classifiers. Speakers of the two New Caledonian languages — Nêlêmwa and Iaai — who have access to larger inventories of classifiers, used a mix of classifiers. Note that these are just the most frequent classifier used for each storyboard, and this does not show whether participants used a matching or a non-matching strategy, which will be discussed below.

Figure 3 shows the mean percentage of matching classifiers used anaphorically across all storyboards by language. North Ambrym’s pattern of matching responses is greater than all other languages in the sample. In contrast, both Nêlêmwa and Iaai have the fewest mean matching responses, compared to all other languages.

Figure 3

Bar plot showing the mean percentage of matching strategies for each sample language (total number of possessive strategies across all participants for each sample language shown in brackets). Error bars show the 95% confidence intervals.

The results show both similarities and differences to the analysis given in section 4.1, which looked at the strict anaphoric use of classifiers. In both analyses, the results show that North Ambrym has a higher number of instances of matching classifiers compared to the other languages. When considering all possessive strategies, used anaphorically or non-anaphorically, the two New Caledonian languages (Iaai and Nêlêmwa) have much lower matching responses across the storyboards (c.f. Figure 3). This indicates that they have more variable classification systems compared to the Vanuatu languages. The three Vanuatu languages other than North Ambrym, that is Merei, Vatlongos and Lewo, have intermediate systems, which have a slight preference for using the same possessive strategy across the storyboards.

Multiple linear regression was conducted to investigate whether the percentage of non-matching possessive strategies used by each participant was affected by the number of different possessive constructions used across the storyboards and the language spoken. Homoscedasticity was confirmed with a Breusch-Pagan test result of x² = 1.175 p = .278. However, the Shapiro-Wilk test showed the residuals were non-normal, W = .970, p = .006. Despite this, given the nature of our data, transformation of the data or exclusion of valid data points that presented as outliers, were deemed more problematic as approaches than cautionary interpretation of the findings (see Pek, Wong & Wong 2018 for a review). The model was significant, F(6,118) = 12.02, p < .001, indicating at least one predictor significantly affects the percentage of non-matching strategies. The model explains 35% of the variance in number of non-matching strategies, with an adjusted R² of .348. The number of possessive constructions that a participant used across storyboards was a significant predictor t(3.449) = 7.461, p < .001, indicating a strong positive effect on the percentage of non-matching possessive strategies. Only the North Ambrym language was a significant predictor t(–2.819) = –27.826, p = .006, indicating a strong negative effect on the percentage of non-matching possessive strategies.

Similar to the analyses detailed in section 4.1, we ran a GLMM with random intercepts. The dependent variable was the matching/non-matching strategy for a storyboard for all possessive constructions, with language as the fixed effect. The model had two random effects — individual storyboard and participant. Table 10 gives an overview of the results. There were 988 observations, 125 participants and 8 storyboards. Iaai was used as the reference level for the fixed effects of language.

The model shows that the dependent variable is more often 1 (a matching strategy) in North Ambrym, Merei, Vatlongos and Lewo than it is in Iaai. There was no significant difference between Nêlêmwa and Iaai. This model gives a nuanced picture as all the Vanuatu languages are significant predictors of matching strategies, with North Ambrym the most significant. The R² values in the model show that when including the random effects of storyboard and participant, the model accounts for 42% of the matching strategies.

To investigate if the three Vanuatu languages of Lewo, Merei and Vatlongos function as intermediate systems different to North Ambrym on the one hand, and Iaai and Nêlêmwa on the other hand, we re-ran the GLMM with Vatlongos as the reference level. Table 11 shows the results for the fixed effects of language.

The model shows that the dependent variable is more often 1 (matching strategy) in North Ambrym than in Vatlongos and that the dependent variable is more often 0 (non-matching strategy) in Iaai and Nêlêmwa than in Vatlongos. There is no significant difference between Lewo and Vatlongos or Merei and Vatlongos. The model confirms that Merei, Lewo and Vatlongos function as intermediate systems.

To visualise the difference between participants and languages, Figure 4 shows the cumulative percentage of participants using a matching strategy across storyboards. The eight storyboards are represented on the x-axis, where the number 8 represents the percentage of participants who used matching strategies for all eight storyboards. The number 7 represents the cumulative percentage of participants who gave matching strategies for either all eight storyboards, or seven out of eight of the storyboards. Figure 4 gives a clear representation of the differences between the languages, with three clusters of clines. North Ambrym, with the most matching strategies, has over 35% of participants using a matching strategy for all eight storyboards, with 90% of participants using a matching strategy for either all storyboards, or seven out of eight storyboards. The three other Vanuatu languages (Merei, Vatlongos and Lewo) all have smaller percentages of participants using matching strategies for all eight storyboards (between 9% and 14%), rising to between 22% and 32% for all eight storyboards and for seven out of eight storyboards. It is important to note therefore, that some speakers in these languages do prefer to use a matching strategy regardless of interactional context. There were no speakers of the two New Caledonian languages, Iaai and Nêlêmwa, who gave a matching strategy for either all eight storyboards or seven out of eight storyboards. Around 16% of speakers of Iaai and Nêlêmwa used matching strategies for six out of eight of the storyboards. These three clusters confirm the mixed model analysis above.

Figure 4

Line plot showing the cumulative percentage of participants using matching strategies across the storyboards.

4.3 Response patterns

In this section, we provide a brief overview of the use of classifiers across each language and storyboard to get a better understanding of how speakers of each language use their system. To get a more in-depth understanding, we then abstract away from the semantic/mnemonic labels used to compare across languages in a more equal way.

For understanding the move towards fixed assignment in classification system, we have to investigate what types of matching and non-matching strategies were chosen by the participants across the storyboards, and whether certain strategies are preferred. The storyboards followed a similar four picture pattern of (i) finding or catching a particular object, (ii) returning home with it, (iii) doing something natural with it (e.g., eating, drinking, making fire), and (iv) doing something less obvious, such as selling or washing with it. As each language has a varying inventory of classifiers, which embody different semantic domains, and a number of other possessive constructions, it is not a simple task to compare like for like. In order to create a fair comparison across languages, we abstracted away from the semantics of each possessive construction and instead focussed on whether a particular construction was used and then if it was repeated later in the story. There are five logical values based on the four possible responses for a storyboard:

• 0: no possessive strategy given. For example, if only a noun, a pronoun or zero anaphor was used by a participant in a particular response.

• 1: the first possessive construction used. For example, a particular possessive classifier, such as food or drink, or a directly possessed noun or prepositional construction. Any further uses of the same strategy will also receive this code.

• 2: the number given to the second different possessive construction used and any uses of that same construction thereafter.

• 3: the number given to the third different possessive construction used and any uses of that same construction thereafter.

• 4: the number given to the fourth different possessive construction used.

Table 12 gives some fictitious data to show how the patterns are coded. Fictitious data is shown for eight different participants and their responses to the fish storyboard. Different classifiers are shown in small caps – gen ‘general classifier’, food ‘food classifier’, burden ‘burden or carry classifier’. Some responses have both a classifier and a possessed noun, other cells have anaphoric use of classifiers without a possessed noun, and yet other cells have just a noun which does not appear in a possessive construction. The pattern row gives a four-digit code corresponding the pattern used by the participant, outlined above. Participants 1–3 used matching strategies, and participants 4–8 used non-matching strategies.

Descriptive statistics for the number of pattern responses for each language are shown in Table 13.

We conducted an analysis of variance using the rstatix package (Kassambara 2023). The ANOVA shows a significant difference between number of patterns given by participants across languages F(5, 119) = 12.579, p < .001, ηp² = .346. Pairwise comparisons comparing individual languages indicated that North Ambrym had significantly lower numbers of patterns than all other languages (largest p < .002). Iaai participants gave significantly more patterns than Lewo, Merei, Vatlongos and North Ambrym (largest p = .012). Nêlêmwa participants gave significantly more pattern responses than North Ambrym and Vatlongos (largest p = .005). Figure 5 shows the mean number of patterns across the language sample. The pattern closely mirrors that shown in Figure 3, revealing that languages with more matching possessive strategies used lower overall patterns. This is expected as there were only 9 matching possessive strategies (patterns that either have 0 or 1), lower than the twenty patterns for non-matching possessive strategies (patterns that have 2 or higher in them).

Figure 5

Bar plot showing the mean number of patterns for each sample language (total number of possessive strategies across all participants for each sample language shown in brackets). Error bars show the 95% confidence intervals.

The stacked bar plot in Figure 6 illustrates the different patterns used across the languages. In order to simplify the data, patterns where there were insufficient numbers of classifiers to be counted as either matching or non-matching are not included. Patterns such as 0000, 0100 or 0010, where only one or no classifiers were given for a storyboard amounted to 3.95% of the data for Iaai, 2.63% for Nêlêmwa, and 1.14% for Vatlongos.

Figure 6

Stacked bar plot showing matching and non-matching strategies across the sample languages.

Table 14 shows a summary of the highest frequency patterns for both matching and non-matching strategies. Out of the nine fixed patterns, the 0111 pattern had the highest frequency across all languages. Out of the twenty variable patterns, 0121 had the highest frequency for Merei, Nêlêmwa and Iaai. 0112 had the highest frequency for Lewo and Vatlongos, and 0122 was the highest percent for North Ambrym.

Individual languages were analysed using paired samples t-tests comparing the total number of unique matching and non-matching patterns. North Ambrym showed significantly more matching patterns than non-matching patterns (t = 6.736, p < .001). All other languages had no significant difference between the frequency of matching and non-matching patterns (lowest p = .752).

4.4 Individual Storyboard Analysis

The storyboards investigate eight different target nouns that differ both in noun frequency and saliency. Figure 7 details the percentage of matching possessive strategy across individual storyboards (same criteria as section 4.2) for all participants across each sample language. The x-axis shows each individual storyboard in left to right descending order according to average matching possessive strategy percentages across all six languages (i.e., ‘firewood’ has the highest mean agreement percentages across the six languages, and ‘water’ has the lowest mean agreement percentages).

Figure 7

Matching possessive strategies (%) across each storyboard and language. The brackets after each storyboard name indicate the CSI rank followed by the frequency quartile.

Chi-squared tests with Yates’ continuity correction showed that there were significant differences between storyboards and whether participants used a matching or non-matching strategy in their response (either anaphorically or with a possessed noun). There was a significant difference for Merei, with a moderate association (X² = 17.485, p = .015, V = .323¹¹). Both Lewo (X² = 32.273, p < .001, V = .428) and Vatlongos (X² = 21.016, p = .004, V = .428) showed significant differences, with a relatively strong association. For North Ambrym, expected values were less than 5 for non-matching strategies, hence a Fisher’s test of exactness was conducted, which resulted in a significant difference, but with a moderate association (p = .038, V = .278). Nêlêmwa has a moderate association (X² = 14.459, p = .044, V = .313). For Iaai, a Chi-squared test was non-significant (p = .179).

For the languages with significant differences and a relatively strong association (Lewo and Vatlongos), Bonferroni corrected pairwise comparisons (adjusted p = .003) were conducted by comparing adjusted standardised residuals. Lewo showed significantly fewer matching strategies for the pig storyboard and significantly more matching strategies for the yam storyboard. Vatlongos showed significantly more matching strategies for the firewood storyboard.