Modelling the choice between PP+DE+N and PP+N possessive constructions in Mandarin Chinese: a mixed effects logistic regression approach

2.1 Data

The data utilized in this study were sourced from the Beijing Language and Culture University Corpus Center (BCC) Corpus. This expansive online Chinese corpus boasts a volume of approximately 9.5 billion characters. It encompasses a diverse range of sources, including newspapers (2 billion characters), literature (3 billion characters), comprehensive data (1.9 billion characters), ancient Chinese texts (2 billion characters), and dialogues (600 million characters, derived from microblogs and movie subtitles). Given its vast scale and diversity, the BCC Corpus provides a comprehensive representation of language usage within contemporary Chinese society.

The methodology for extracting targeted data from the corpus is outlined as follows.

1. Extract 16 combinations for each of PP+DE+N and PP+N based on 8 distinct personal pronouns in Mandarin Chinese and 2 registers in BCC.^[4]

2. From the 150 most frequent character strings in each combination, as determined by the corpus, identify nouns appearing more than half the time across the 16 combinations in both constructions.

3. Extract 33 nouns with high frequency in both PP+DE+N and PP+N constructions (refer to Table 1).

4. Query each combination within the corpus and extract five representative sentences for each to ensure a sufficient data pool.

5. Procure a total of 4,050 observations, including 2,349 instances of PP+DE+N and 1701 instances of PP+N.

6. Manually code the observations according to explanatory variables detailed in Section 2.3.

2.2 Mixed effects models

Mixed-effects models, often abbreviated as mixed models, incorporate both fixed and random effects. As delineated by Rolf Baayen (2008: 241), fixed effects are employed for factors with replicable levels, whereas random effects are utilized for ‘factors with levels randomly sampled from a substantially larger population’. Within the context of linguistic experiments and studies, items (words) and subjects (participants) frequently serve as common random effects.

Mixed models have been employed in research pertaining to constructions and metonymic words and expressions, as evidenced in studies by Natalia Levshina et al. (2013), and Zhang Weiwei et al. (2018; 2011]. Furthermore, numerous studies within the realm of second language acquisition have also utilized mixed models. These include, but are not limited to, works by Wu Shiyu et al. (2017), Qi Jing and Wang Hua (2020), Fang Yinjie and Liang Maocheng (2020), and Ji Yinglin (2020), as well as Lu Shiyi et al. (2020).

Mixed models are employed in this study to investigate various factors within PP+(DE)+N constructions due to their ability to quantify the cumulative impact of both fixed and random effects of explanatory variables, thereby ensuring high model accuracy. To facilitate the fitting of mixed models, this study utilizes the glmer() function from the lme4 package in R.

2.3 Variables^[5]

The variables incorporated within the mixed models are delineated as follows. Variables A-H serve as explanatory variables for fixed effects, while Variable I is the explanatory variable for the random effect. Variable J functions as the response variable. From another perspective, Variables A-C and E-G represent syntactic factors, whereas Variables D and I denote semantic factors. Variable H corresponds to the extralinguistic register factor.

3.1 Model overview

The initial model is formulated as follows:

cxn ∼ num_attr + position + syn_comp + sem_rel + person + num_pp + style + (1|Noun)

The analysis reveals that the coefficients for position = NA was excluded due to rank deficiency in the fixed-effect model matrix. Additionally, a quasi-complete separation was observed between the variables position and num_attr. Upon the removal of the variable position, the final model is articulated as:

cxn ∼ num_attr + syn_comp + sem_rel + person + num_pp + style + (1|Noun)

The final model demonstrates a satisfactory fit to the data, as evidenced by a goodness-of-fit measure of C = 0.75. Moreover, the classification accuracy of the final model surpasses the baseline accuracy of 58 % (2,349/4,050), reaching a notable 70.3 %.

The methods for model validation were based on the approach proposed by Fang Yu and Liu Haitao (2021). Initially, the data in this study were randomly divided into two sets, namely the training set (comprising 75 % of the data) and the test set (comprising the remaining 25 % of the data). Subsequently, the final model was fitted to each training set, and its predictions were evaluated using the corresponding test set. The average classification accuracy across the 100 models was found to be 66.41 %, with a maximum accuracy of 68.68 % and a minimum accuracy of 64.33 %. Notably, this average accuracy surpassed the baseline accuracy of 58 %.

To assess the presence of multicollinearity among the explanatory variables for fixed effects in the final model, the variance inflation factor (VIF) was computed. The results revealed that there was no significant multicollinearity, as indicated by a maximum VIF of 1.02.

Furthermore, the importance of each explanatory variable for fixed effects was validated using the Anova() function from the car package in R. The analysis demonstrated that all six variables included in the final model significantly contributed to predicting the choice between PP+DE+N and PP+N. Notably, the three most influential variables were identified as num_attr, syn_comp, and num_pp.

3.2 The analysis of fixed effects

Table 4 displays a comprehensive summary of the fixed effects observed in the final model. It is worth noting that variables exhibiting a positive estimate are associated with an increased likelihood of PP+N, whereas variables displaying a negative estimate suggest a preference for PP+DE+N.

Based on the findings presented in Table 4, it is evident that all the p-values associated with the variables in the final model are below the conventional level of significance (0.05), except for sem_rel = attr, sem_rel = own, and sem_rel = assoc. This implies that, in comparison to the reference level of sem_rel = kinsh, there is no statistically significant distinction between PP+DE+N and PP+N when sem_rel is characterized as attr, own, or assoc.

The intercept estimate in the final model is 0.341, with a significance level of 0.01 ≤ p < 0.05. This indicates a relatively significant difference between PP+DE+N and PP+N when all fixed effects variables are at their reference levels, namely num_attr = 1, syn_comp = s, sem_rel = kinsh, person = 1, num_pp = sing, and style = lit. Since the estimate represents the log odds, it is exponentiated to obtain simple odds. Upon exponentiation, the odds of PP+N compared to PP+DE+N are approximately 1.4. This implies that in the context of literary style, when a first-person singular pronoun functions as an attribute, the head of the phrase is a kinship term, and the entire noun phrase occupies the subject position, PP+N constructions occur approximately 1.4 times more frequently than PP+DE+N. To illustrate this preference, consider example (24) where the symbol ‘ > ’ denotes a higher frequency on the left side compared to the right side:

The shorter form of 我媽媽 wǒ māma ‘my mom’ compared to 我的媽媽 wǒ de māma ‘my mom’ conveys a greater sense of intimacy and aligns better with the conceptual proximity between 我 wǒ ‘I’ and 媽媽 māma ‘mom’.

The estimate for num_attr = 2 is 2.150, which, when exponentiated, becomes 8.6. This indicates that when num_attr = 2, the frequency of PP+N is approximately 8.6 times higher than that of PP+DE+N, in comparison to the reference level of num_attr = 1. This preference is illustrated through examples (25) and (26):

The estimate for num_attr = m is 3.990, which equates to an odds ratio of 54.1 upon exponentiation. This signifies that relative to num_attr = 1, constructions with num_attr = m exhibit a 54.1 times higher frequency of PP+N compared to PP+DE+N. Examples (27) and (28) demonstrate this tendency of favoring PP+N when multiple attributes are present:

Hence, the estimates for num_attr = 2 and num_attr = m provide evidence that an increase in the number of attributives corresponds to a higher occurrence rate of PP+N. This finding aligns with the research conducted by Cheng Shuqiu (2009).

The estimate of syn_comp = o is −0.981 (exponentiated as 0.375), signifing that compared with syn_comp = s, the frequency of PP+N with syn_comp = o is 37.5 % that of PP+DE+N. In other words, PP+DE+N occurs 2.7 (1/0.375) times more frequently than PP+N. Examples (29) and (30) illustrate this preferential selection.

The rationale behind the compatibility of PP+DE+N with the object position has been elucidated in prior scholarly works. For instance, Pan Tingting (2021) posits that de serves a descriptive function by introducing new information, thereby rendering it more congruous in the object position.

The estimate of syn_comp = na is 0.499 (exponentiated as 1.6). This means that compared with syn_comp = s, when syn_comp = na, PP+N appears 1.6 times more frequently than PP+DE+N. Examples (31) and (32) illustrate this preference.

The reason for this preference may be related to sentence length. Sentences with a pivotal construction (SVOV) are typically longer than SVO sentences. The longer the sentence string becomes, the more likely it is for de to be omitted, resulting in a preference for PP+N over PP+DE+N.

The estimated value of sem_rel = whlpt is 0.618, which is exponentiated as 1.9. This indicates that when sem_rel = whlpt, PP+N occurs 1.9 times more frequently than PP+DE+N, compared to when sem_rel = kinsh. This preference is illustrated in examples (33) and (34).

Omitting de is easier for whole-part semantic relationships than for kinship relationships, indicating that body parts are more inalienable while kinship seems more alienable.^[6] First, Chinese culture emphasizes kinship and hierarchy, so de is used to create a sense of distance (respect) between relatives, making kinship seem more alienable. Second, ‘PP+body parts’ commonly appears after prepositions to form prepositional phrases, whereas ‘PP+kinship’ does not have the corresponding usage.

The estimated of sem_rel = nomi is −1.193, which is exponentiated as 0.303. This means that when sem_rel = nomi, the occurrence frequency of P+N is only 30.3 % of the occurrence frequency of PP+DE+N, compared to when sem_rel = kinsh. Another way to express this is that PP+DE+N occurs 3.3 times more frequently than PP+N. Examples (35) and (36) illustrate this preference.

In instances where the head noun is a kinship word, the preceding PP and the noun typically exhibit an attribute-head relationship, thereby reducing the necessity for de to delineate this relationship. Conversely, when the head noun is the nominalization of a verb, the inclusion of de becomes essential. This requirement arises due to the potential overlap between the structure of a PP combined directly with a nominalization of a verb and the subject-verb structure. For instance, the phrase 我 wǒ ‘I, my’ + 建議 jiànyì ‘suggest, suggestion’ could be misinterpreted as the conventional Chinese subject-predicate sentence pattern ‘I suggest something’, rather than the intended ‘my suggestion’. Therefore, de is required to accurately signify the intended attribute-head relationship.

The estimated value for person = 2 is −0.242, which becomes 0.785 when exponentiated. This indicates when person = 2, the occurrence of PP+DE+N is 1.3 times more frequent than PP+N, in comparison to person = 1. Examples (37) and (38) provide a representation of this comparison.

The aforementioned illustration demonstrates that the second person pronoun requires the addition of de to reinforce the speaker-listener distance, in contrast to the first person pronoun. This finding aligns with the concept of linguistic iconicity, as proposed by John Haiman (1983: 782), which suggests that the length of the message corresponds to the social distance between interlocutors, while maintaining equal referential content.

The estimate for person = 3 is −0.179, exponentiated as 0.836. It signifies that when person = 3, the occurrence of PP+DE+N is 1.2 times more frequent than PP+N, in comparison to person = 1. Examples (39) and (40) provide a representation of this tendency.

Similar to person = 2, the aforementioned preference can be attributed to the motivation of linguistic iconicity, wherein the distance between the speaker and the third person pronoun necessitates the use of de to reinforce it.

The estimated value for num_pp = pl is −0.533 (exponentiated as 0.587). Therefore, when num_pp = pl, the occurrence of PP+DE+N is 1.7 times more frequent than PP+N, in comparison to num_pp = sg. Examples (41) and (42) illustrate this preference.

Given that the character string 我們 wǒmen ‘we’ + 媽媽 māma ‘mom’ can convey both an attribute-head meaning, ‘our moms’, and an appositive meaning, ‘we, as moms’, the presence of de between 我們 wǒmen ‘we’ and 媽媽 māma ‘mom’ serves to separate the attribute and the noun head.

The estimated value for style = news is −0.165 (exponentiated as 0.848). This indicates when style = news, the frequency of occurrence of PP+DE+N is 1.2 times higher than the frequency of occurrence of PP+N, in comparison to style = lit. The illustration for such tendency is shown in examples (43) and (44).

In the context of newspapers, where language tends to be more formal, the usage of PP+DE+N is more prevalent. Conversely, in literary works, where language tends to be more colloquial, the usage of PP+N is more common. This observation aligns with John Haiman’s (1983, p. 800) assertion that ‘the more respectful the register, the more syllables in the same message.’

Table 5 provides a comprehensive summary of the fixed effects preference. Upon reviewing Section 2.3, which introduces the variables, the following conclusions can be drawn:

1. The variables position and gender do not demonstrate sufficient relevance to be included in the fixed effects.

2. The effect of num_attr aligns with Cheng Shuqiu’s (2009) findings, which suggest that the omission of de after the possessive attribute is more likely as the number of attribute items increases. The effect of syn_comp is consistent with the research of Lu Bingfu (2011) and Pan Tingting (2021), who propose that NP with de is more likely to be omitted when it functions as a subject compared to when it functions as an object.

3. The effect of style contradicts Li Tianguang’s (2012) claim that the use of de is more frequent in narrative and descriptive works than in news works. Similarly, the effect of sem-rel opposes the findings of Chen Zhenyu and Ye Jingting (2014), who suggest that body parts, as opposed to kinship terms, have a higher tendency to use the possessive marker de due to their higher controllability.

4. The effect of person is not directly related to the ‘potentiality of agency’ or the ‘extended animacy hierarchy’ of personal pronouns mentioned earlier. Instead, it can be explained by the principle of linguistic iconicity, where the distance between the speaker and the second/third person pronoun necessitates the use of de for reinforcement. Similarly, the effect of num_pp is not strongly associated with the differentiation of plural meanings 1 and 2 proposed by Chen Yu-Jie (2008). Rather, it can be interpreted in terms of de marking the attribute-head meaning rather than the appositive meaning between PP and noun.

3.3 The analysis of the random effect

Table 6 provides a summary of the random effect observed in the final model. Among the 20 nouns with positive estimates, the preference is towards PP+N, whereas the 13 nouns with negative estimates exhibit a preference for PP+DE+N. Figure 1 visualizes the summary using a bar plot, where nouns favoring PP+DE+N are depicted in red, and nouns favoring PP+N are represented in blue.

Figure 1:

Bar plot of the random effect.

Nouns that prefer PP+N can be categorized into the following five classes:

1. 房間 fángjiān ‘room’; 信 xìn ‘letter’:

Objects like rooms and letters possess a sense of privacy and are closely associated with the individual. As a result, the omission of de before these nouns is more common.

1. 病 bìng ‘disease’; 心靈 xīnlíng ‘soul’; 心情 xīnqíng ‘mood’; 心 xīn ‘heart’; 臉 liǎn ‘face’; 內心 nèixīn ‘inner heart’; 身體 shēntǐ ‘body’; 精神 jīngshén ‘spirit’:

These nouns possess unique characteristics and are closely tied to one’s own being, leading to a tendency to omit de before them.

1. 手 shǒu ‘hand’; 腳 jiǎo ‘foot’:

Hands and feet are associated with the sense of touch and mobility. They serve as important means for individuals to explore the external world, thus establishing a close relationship with oneself. Consequently, they are more likely to combine directly with PP. Additionally, in Mandarin Chinese, ‘PP+hand/foot’ is commonly used to form prepositional phrases following prepositions, such as 在我手上 zài wǒ shǒu shàng ‘in my hands’ and 在我腳下 zài wǒ jiǎo xià ‘under my feet’.

1. 父母 fùmǔ ‘parent’; 女兒 nǚ’ér ‘daughter’; 母親 mǔqīn ‘mother’; 媽媽 māma ‘mom’; 父親 fùqīn ‘father’; 兒子 érzi ‘son’:

These nouns represent family members and symbolize inseparable kinship, resulting in a tendency to omit de before them.

1. 家庭 jiātíng ‘family’; 國家 guójiā ‘country’:

Both family and country convey a sense of collectivity. Given that Chinese culture places emphasis on close kinship and collectivism, the omission of de between PP and these nouns is more prevalent.

Nouns that prefer PP+DE+N can be classified into the following eight clusters:

1. 爸爸 bàba ‘dad’; 哥哥 gēge ‘elder brother’:

In traditional Chinese culture, the roles of a father and an elder brother are similar. The saying ‘The elder brother is like a father’ emphasizes the need to respect one’s elder brother as one would respect their father. Consequently, dad and elder brother hold more authority and maintain a greater distance from other family members. This sense of distance necessitates the use of de before the nouns 爸爸 bàba’ and 哥哥 gēge.

1. 孩子 háizi ‘child’:

The term ‘child’ is a general reference to both sons and daughters. Addressing someone as a child creates a greater sense of distance compared to directly addressing them as a son or daughter. Additionally, ‘child’ can be used to refer to younger boys and girls who do not share a kinship relationship with the speaker, further emphasizing the distance expressed by the term. These factors explain why de is more commonly used before 孩子 háizi in possessive constructions.

1. 兄弟 xiōngdì ‘brother’:

The term ‘brother’ can be used to address friends, emphasizing personal loyalty. As a result, 兄弟 xiōngdì carries a greater distance from kindred relatives, making it more compatible with the use of de in possessive constructions.

1. 耳朵 ěrduo ‘ear’; 眼睛 yǎnjīng ‘eye’:

People’s perception of sounds and images through their ears and eyes is rather general and does not necessarily reflect individuality or personalization. Therefore, 耳朵 ěrduǒ and 眼睛 yǎnjīng are not representative of a person’s inherent attributes and are more inclined to be preceded by de.

1. 錢 qián ‘money’; 建議 jiànyì ‘suggestion’; 計畫 jìhuà ‘plan’:

These three nouns, regardless of their concreteness, can be transferred to others. As a result, they are distant from oneself and tend to be accompanied by de.

1. 思想 sīxiǎng ‘thought’; 頭腦 tóunǎo ‘mind’:

Thoughts and the mind are more easily influenced than inherent attributes, making them more compatible with the use of de before them.

1. 生命 shēngmìng ‘life’:

In Mandarin Chinese, ‘life’ can appear in phrases such as ‘offering one’s life’ or ‘sacrificing one’s life’. In these contexts, life acts as an object that can be transferred, creating a distance from oneself. Thus, de tends to appear between PP and 生命 shēngmìng to indicate such distance.

1. 人 rén ‘people’:

‘PP+DE+人rén’ expresses subordinates of somebody. These subordinates take orders from their superior, with whom they maintain a greater distance. Therefore, de tends to appear between PP and ‘人rén’ to indicate this distance.

In summary, the random effect analysis of the noun reveals a correlation between the presence or omission of de and the distance between PP and the noun. John Haiman (1983: 782, 783) argues that ‘the linguistic distance between expressions corresponds to the conceptual distance between them’ and that ‘there is a closer conceptual link between a possessor and an inalienably possessed object than between a possessor and an alienably possessed object.’ While Mandarin Chinese may not have a clear distinction between alienability and inalienability in possessed relations, it adheres to the aforementioned principle by adding de to increase the linguistic distance between PP and N, thereby aligning with their greater conceptual distance.