PaperlayerView exact revision on arXiv

Source-linked AI summary

Reduplicative constructions in Mandarin: Socio-emotional profiling through distributional semantics

Chaoyi Wu, Yu-Hsiang Tseng, R. Harald Baayen

arXiv:2609.16860v1cs.CL

TL;DR

This study asks whether distributional semantics can capture the varied meanings of Mandarin reduplication and clarify how its patterns relate to their bases. It profiles embeddings and compares base-word and reduplication spaces, finding systematic pattern differences and largely preserved overall organization with local mismatches.

  • Problem

    The study asks whether embeddings can capture Mandarin reduplication’s varied semantic and pragmatic properties, distinguish AABB and ABAB, and clarify semantic shifts and transparency relative to bases.

    Source: §1 · pp. 1–2

  • Method

    The study profiles reduplicative and base-word embeddings, models base-to-reduplication shifts, and uses Procrustes analysis to compare their semantic spaces.

    Source: §6 · pp. 23–24

  • Results

    AABB and ABAB differ systematically in semantic space and profiles, while Procrustes analysis finds that base-word space organization is largely preserved in reduplication space.

    Source: §7 · pp. 27–28

  • Takeaways & Limitations

    Distributional semantics provides a quantitative perspective on the systematicities and semantic versatility of Mandarin reduplication and may prove useful for studying other word-formation processes.

    Source: §7 · p. 28

    Static embeddings blend a form’s context-dependent meanings and discourse functions, so the reported semantic profiles are approximate.

    Source: §7 · p. 28

Abstract

from arXiv · show

Mandarin Chinese has two productive reduplicative constructions that repeat either two-character base words or their constituents (e.g., `in good health', `discuss a bit'). Their varied meanings have been described as realizing plurality, valence coloring, sound symbolism and pragmatic functions. The aim of this study is twofold. A first goal is to clarify whether it is possible to come to a more precise understanding of the variegated semantics of Mandarin reduplication by using word embeddings from distributional semantics. A second goal is to explore how useful embeddings are for understanding the details of a semantically complex word-formation process. We show that the embedding space recovers the semantic and grammatical properties of reduplications previously identified in the literature, validating Tencent embeddings for morphological investigation. Semantic profiling revealed that reduplicative constructions are often strongly represented on multiple dimensions. The two patterns exhibit clear semantic and pragmatic differentiation in distributional space. Procrustes analysis clarified that the overall organization of the base-word space is largely preserved in the reduplication space, with local mismatches highlighting regions of discourse-pragmatic reorganization. Taken together, these results show that high-dimensional word embeddings can recover established linguistic generalizations, and capture the semantic versatility of Mandarin reduplication and constructional transparency.

1 Introduction

This study asks whether distributional semantics can recover the varied semantic and pragmatic properties attributed to Mandarin reduplication and clarify how AABB and ABAB constructions relate to their bases. It uses embeddings to profile the patterns, examine semantic shifts, and compare base-word and reduplication spaces.

1 Introduction

  • The study represents CCL Corpus reduplications and disyllabic bases with 200-dimensional Tencent embeddings, profiles their properties, models base-to-reduplication shifts, and compares the spaces using Procrustes analysis.Mandarin non-reduplicative words serve as a reference baseline.Source: §1 · pp. 1–2
  • Earlier descriptions attribute plurality, intensification, affective coloring, and pragmatic force to reduplication, with tendencies differing between AABB and ABAB [Chao 1968, Hua 2003].AABB has been associated with nominal plurality, verbal event iteration, and adjectival intensification, while verbal ABAB often has tentative or suggestive interpretations [Zhang 2015, J. Zhu 2003, Melloni & Basciano 2018, Lu & Müller 2026].Source: §1 · p. 2
  • The study investigates whether embeddings recover reduplications’ semantic and pragmatic properties, distinguish AABB from ABAB, characterize shifts from bases, and assess semantic transparency.The questions address whether the patterns form distinct regions in semantic space and how their properties differ.Source: §1 · pp. 2–3

2 Data

The dataset comprises four-character AABB and ABAB reduplications extracted from the CCL corpus, paired where possible with corresponding disyllabic bases and represented using Tencent embeddings.

2 Data

  • The CCL corpus yielded 1,010 four-character reduplications: 520 AABB forms and 490 ABAB forms, represented with Tencent embeddings based on word2vec algorithms.The embeddings came from the Tencent repository.Source: §2 · p. 3
  • Researchers identified a corresponding base word for each reduplication, usually the underlying disyllabic AB form.For example, 健康 is paired with 健健康康, and 讨论 with 讨论讨论.Source: §2 · p. 3
  • Some reduplications lack an existing or semantically plausible base, leaving 955 reduplication–base pairs involving 916 distinct bases.For example, 盆盆罐罐 has no existing disyllabic base 盆罐, while 林总 is unrelated in meaning to 林林总总.Source: §2 · p. 3
  • Thirty-nine base words give rise to both AABB and ABAB constructions.热闹, for example, corresponds to both 热热闹闹 and 热闹热闹.Source: §2 · p. 3

3 Clusters of reduplicative constructions

Mandarin reduplications occupy identifiable regions in the broader lexical semantic space and form differentiated clusters. The clusters are associated with distinct grammatical, semantic, and interactional tendencies.

3 Clusters of reduplicative constructions

  • 98.7% accuracy, versus a 68.3% majority-class baseline, distinguished reduplications from reference words under leave-one-out validation.Figure 1 situates reduplications in several identifiable regions rather than scattering them randomly among reference words.Source: Figure 1 · p. 4, §3 · p. 4
  • 90.50% accuracy, versus a 16.24% majority baseline, supported selecting 10 k-means clusters; Figure 2 shows their two- and three-dimensional maps.The LDA predicted the k-means clusters under leave-one-out cross-validation.Source: §3 · p. 4, Figure 2 · p. 5, Table 1 · p. 6
  • Cluster 2 is dominated by ABAB attenuative or delimitative verbs, while Cluster 3 mainly contains AABB forms, including verbal and nominal formations bringing opposites together.Cluster 6 mixes AABB and ABAB forms associated with quantitative increase or intensified degree.Source: §3 · pp. 6–7, Table 2 · p. 6
  • Cluster 7 consists entirely of ABAB interpersonal expressions, while Clusters 4 and 6 include intensified adjectives and Cluster 9 gathers many onomatopoeic forms.Cluster 5 contains many adverbial expressions, and Clusters 1 and 10 include incremental and completive meanings.Source: Table 2 · p. 6, §3 · p. 7

4 Semantic profiling

The study profiles reduplications across grammatical and semantic properties, using embedding-based measures to characterize clusters and compare AABB with ABAB. The profiling procedure represents properties with anchor-word embeddings.

4 Semantic profiling

  • Profiling examines lexical category, valence, self-relatedness, and sound-relatedness to characterize clusters and differentiate the two construction types.The properties range from categorical distinctions to gradient cognitive dimensions.Source: §4 · p. 7

4.1.1 Profiling method

  • The category-defining vector method represents a property by the centroid of prototypical anchor-word embeddings and estimates a construction’s association from its position relative to that centroid.Unipolar profile scores use cosine similarity; bipolar properties use an axis between opposing centroids, with the projection indicating alignment with one pole or the other.Source: §4.1.1 · pp. 7–8

4.1.2 Anchor-word selection

  • For each property, the study selected 100 high-frequency anchor words representing the relevant semantic pole or category, retaining only words with Tencent embeddings.Emotion anchors referenced the Chinese Emotional Lexicon Ontology, and lexical categories referenced A Frequency Dictionary of Mandarin Chinese.Source: §4.1.2 · p. 8

4.2 Construction: AABB vs. ABAB

AABB and ABAB constructions are unevenly distributed across semantic clusters and are highly predictable from embeddings. Their separation is not the only source of cluster structure.

4.2 Construction: AABB vs. ABAB

  • AABB and ABAB patterns have markedly different distributions across clusters: Cluster 5 contains only AABB items, while Cluster 7 contains only ABAB items; Clusters 2, 6, and 8 favor one type.Table 3 presents the distribution of AABB and ABAB patterns across the 10 clusters.Source: §4.2 · p. 8, Table 3 · p. 9
  • 96.44% accuracy (majority-class baseline: 51.49%) identifies AABB versus ABAB constructions from embeddings under leave-one-out cross-validation.The classifier correctly identified 506 of 520 AABB items and 468 of 490 ABAB items.Source: §4.2 · p. 8
  • Despite their separability, clusters are not determined by construction pattern alone, with other linguistic predictors such as lexical category also contributing.The clusters differ in how strongly they favor either pattern, while most contain both.Source: §4.2 · p. 8

4.3 Lexical category

Lexical categories show distinct tendencies across the semantic clusters and are related to construction pattern. Continuous lexical axes capture these gradients, but lexical-category scores alone do not account for all cluster structure.

4.3 Lexical category

  • 84.06% accuracy (majority-class baseline: 45.05%) predicts lexical category from embeddings under leave-one-out cross-validation.Table 4 distributes verbs, nouns, adjectives, and adverbs across the 10 clusters.Source: §4.3 · p. 9, Table 4 · p. 10
  • Continuous Noun-Verb and Adjective-Adverb axes profile gradient category tendencies, reflecting that Mandarin words can serve multiple lexical categories depending on context.The axes are used to project reduplications and examine category tendencies in different regions of semantic space.Source: §4.3 · p. 9
  • Clusters 10, 5, and 1 tend noun-like, while Clusters 6 and 2 tend verb-like; adjectives cluster especially in 5 and 6, and adverbs mainly in 3.Cluster 6 also shows adverb-like tendencies; Table 4 reports lexical-category counts by cluster.Source: §4.3 · pp. 9–10, Table 4 · p. 10
  • More verb-like tendencies are mainly associated with ABAB, whereas noun-like tendencies are mainly associated with AABB; adjectives and adverbs occur in both, more often in AABB.Source: §4.3 · p. 10
  • 32.97% accuracy (majority-class baseline: 16.24%) predicts clusters from lexical-category axis scores, but some clusters remain unrecoverable from those scores alone.Clusters 2 and 8 were identified with moderate success, whereas Clusters 3, 4, and 6 were not recoverable.Source: §4.3 · p. 11, Table 5 · p. 11

4.4 Valence

Reduplications occupy different semantic-space regions according to their scores on three emotion contrasts. Those emotional scores relate to cluster membership but only partially explain it.

4.4 Valence

  • High anger-fear scores are more common on the left and high surprise-disgust scores on the right; low scores on these axes tend toward the opposite regions.For the surprise-disgust axis, low scores—closer to disgust—are concentrated on the left, especially in Cluster 8.Source: §4.4 · pp. 11–12
  • High happy-sad scores cluster in the left-central region, while lower scores are more common in the middle and right regions of the semantic space.Figure 4 maps Happy-Sad, Anger-Fear, and Surprise-Disgust scores across the three panels.Source: §4.4 · p. 11, Figure 4 · p. 12
  • 27.03% accuracy (majority-class baseline: 16.24%) predicts cluster membership from the three emotional axes under leave-one-out cross-validation.Clusters 1, 2, and 5 were recovered with moderate success, but Clusters 3, 4, and 10 were not recoverable from emotion scores alone; Table 5 reports per-cluster accuracies.Source: §4.4 · p. 12

4.5 Self-relatedness

Cluster 7 contains many high self-relatedness items used to express a speaker’s stance or attitude. Self-relatedness scores predict clusters only partially.

4.5 Self-relatedness

  • Cluster 7 has particularly high self-relatedness scores, including interactional expressions used to convey the speaker’s stance or attitude.Figure 5 colors formations by cosine similarity with the centroid of self-related anchor words.Source: §4.5 · pp. 12–13, Figure 5 · p. 13
  • 28.12% accuracy (majority-class baseline: 16.24%) predicts clusters from ego scores under leave-one-out cross-validation.Cluster 7 contains 19 of the 20 highest-ego reduplications, but 6 of 10 clusters are not recoverable from ego scores alone.Source: §4.5 · p. 13

4.6 Sound-related meaning

Sound-related meanings occupy recognizable regions of the reduplication space, with physical sound-source categories concentrated especially in several clusters and human sounds distributed more broadly. LDA recovered Clusters 7, 8, and 9 from sound-source scores with 49.11% accuracy, above the 16.24% majority-class baseline.

4.6 Sound-related meaning

  • Figure 6 maps cosine similarity to four sound-source categories, with physical-source categories concentrated around Clusters 9, 1, and 8.Cluster 9 is especially associated with gases; liquids and vibrating solids also occur in Clusters 9 and 1.Source: §4.6 · pp. 13–14, Figure 6 · p. 14
  • Cluster 9 includes vibrating-solid sounds such as jingling bells and cracking, while Clusters 10 and 2 include examples orthogonal to that category.The category is one of three everyday physical sound-source types, alongside gases and liquids.Source: §4.6 · pp. 13–15
  • Human-sound formations span more of the semantic space than physical sound sources, including vocal expressions in Cluster 7 and action-related sounds in Cluster 8.Clusters 10 and 6 include examples orthogonal to the human-sound category.Source: §4.6 · p. 15
  • 49.11% accuracy (majority-class baseline: 16.24%) was achieved when LDA predicted clusters from sound-source scores under leave-one-out cross-validation.Clusters 7, 8, and 9 were recovered with moderate to high accuracy.Source: §4.6 · p. 15

4.7 Predictability of clusters from semantic profiling

Semantic profiling categories combined with construction type predict cluster identity well above the majority-class baseline. Recovery varies by cluster, and the investigated features do not fully define the clusters.

4.7 Predictability of clusters from semantic profiling

  • 71.78% accuracy (majority-class baseline: 16.24%) was achieved when LDA predicted cluster identity from semantic profiling categories and construction type.The analysis used leave-one-out cross-validation; Table 6 presents the misclassification results.Source: §4.7 · p. 15, Table 6 · p. 15
  • Cluster 2 (0.88), Cluster 7 (0.87), and Cluster 1 (0.71) were predicted with high accuracy, while Clusters 3 and 4 were not well recoverable.Source: §4.7 · p. 15
  • Ten clusters were supported by LDA, and their identities were predictable from linguistic features, but those features did not fully define individual clusters.The features included construction type, word category, valency, self-relevance, and sound type; the authors infer further systematic variation beyond these variables.Source: §4.7 · p. 16

5 Shift vectors

Shift vectors represent how meaning changes from a disyllabic base to its reduplicated form, and their clusters reveal both recurring shifts and substantial variation. AABB and ABAB forms show distinct shift profiles, while some clusters contain several different kinds of change.

5 Shift vectors

  • A shift vector represents the direction and extent of semantic change, computed from the reduplication vector relative to its base-word vector.The analysis asks whether these changes are systematic across reduplicative patterns.Source: §5 · p. 16

5.1 Semantic transparency: base word similarity in shift space

  • Among 955 reduplications with identifiable bases, shift-space clusters preserve some structure, but others disperse or overlap, including Clusters 1 and 10.Clusters 2, 4, and 5 remain relatively compact; Clusters 3, 8, and 9 become more dispersed.Source: §5.1 · pp. 16, 19, Figure 7 · p. 17
  • Cluster 2 softens and makes event meanings more tentative, while Cluster 4 heightens sensory perceptibility, including visual, tactile, and color effects.Examples include 讨论讨论 ‘have a brief discussion’ and 高高大大 ‘burly and imposing’.Source: §5.1 · pp. 17–18
  • Cluster 3 spans all four shift-space quadrants, grouping repeated actions, entity quantification, vertical iteration, and frequent or comprehensive movement.For instance, 进进出出 is more prevalent with human subjects, while 来来去去 extends more readily to abstract subjects and global situations.Source: §5.1 · pp. 18–19

5.2 Types of base-to-reduplication shifts

  • K-means clustering used k = 10; LDA classified those clusters with 87.33% accuracy versus a 16.65% majority-class baseline, although some clusters overlap or contain varied shifts.Cluster 6 overlaps substantially with Cluster 7, and the clusters differ in how coherent their members’ shifts are.Source: §5.2 · pp. 19–20
  • LDA classified AABB versus ABAB shift vectors with 92.88% accuracy, well above the 50.89% majority-class baseline.Figure 8 marks AABB with circles, ABAB with crosses, and cluster centroids with red crosses.Source: Figure 8 · p. 20, §5.2 · p. 20
  • Shift types include scalar amplification in Cluster 1, bounded and tentative ABAB events in Cluster 3, and positive-affect AABB forms in Cluster 5.Cluster 5 forms can give a more positive evaluative stance than their bases; reduplicated scalar forms also show emphatic contexts.Source: §5.2 · pp. 20–22
  • Some clusters resist a single shift label: Cluster 6 includes spatially distributed configurations, repeated sound depiction, and habitual event construals.For example, 高高低低 presents height contrasts distributed in space, while 唱唱跳跳 construes singing and dancing as habitual.Source: §5.2 · pp. 22–23

6 Procrustes analysis

Procrustes analysis compares the semantic spaces of Mandarin bases and reduplications, finding strong global alignment alongside local differences that reflect varying semantic transparency and pragmatic reorganization.

6 Procrustes analysis

  • Because 5.4% of reduplications lack a base word, Procrustes analysis compares independently clustered spaces using Hungarian-matched centroids and aligns reduplication centroids to the base-word reference.The resulting transformation is applied to all reduplication vectors.Source: §6 · pp. 23–24
  • The base-word and reduplication centroid configurations share highly similar global geometry (r = 0.9336, residual sum of squares m2 12 = 0.1284, p = 0.001).Significance was assessed with 999 permutations.Source: §6 · p. 24
  • Figure 9 relates cluster alignment to transparency: better Hungarian matching accompanies smaller Procrustes residuals (r = −0.72, p = .028).Cluster 2, the outlier with Cook’s distance > 2, was excluded from the regression.Source: §6 · pp. 24–25, Figure 9 · p. 25
  • Cluster 3 has the smallest residual (0.323), while Cluster 6 has the largest (0.630); their centroid cosine similarities are 0.87 and 0.80, respectively.The aligned vectors are well mixed in Cluster 3 but more dispersed in Cluster 6.Source: §6 · p. 25
  • Cluster 3 preserves core meanings while amplifying plurality, event iteration, or scalar degree, whereas Cluster 6 includes ABAB directives less predictable from base meanings.Examples include 枝枝叶叶 ‘many branches and leaves’ and 睡觉睡觉 ‘time to sleep’.Source: §6 · p. 26
  • Cluster 4 has an intermediate residual (0.520): reduplications generally preserve lexical meaning but favor adverbial use, unlike adjective-like bases.Reduplicated 高高兴兴 normally resists the degree modifiers that can accompany 高兴.Source: §6 · p. 26
  • Overall, close alignment marks regions where reduplication elaborates base meanings, while larger residuals characterize less transparent forms or different pragmatic functions.Elaboration includes scalar degree, referential plurality, and event iteration.Source: §6 · pp. 26–27

7 General discussion

The study uses distributional semantics to profile Mandarin AABB and ABAB reduplications and compare their semantic spaces with those of their bases. Embeddings recover distinct pattern profiles and show that base-space organization is largely preserved, with local pragmatic shifts; static embeddings nevertheless blend context-dependent meanings.

7 General discussion

  • AABB and ABAB occupy partly different semantic regions: AABB is more associated with affective, plural, and sound-related meanings, while ABAB is more verbal, interactional, and self-relevant.Both embeddings and shift vectors distinguish the patterns with high accuracy.Source: §7 · p. 27
  • Reduplications combine multiple semantic dimensions rather than fitting one category, so literature descriptions such as intensification, plurality, affect, and stance can capture different facets.The example 开开心心 is strongly positive, relatively adjective- and adverb-like, moderately self-relevant, and weakly sound-related.Source: §7 · p. 27
  • Semantic shifts distribute differently across patterns: tentative meanings occur only in ABAB, while positive affect is dominant in AABB.The shift types discussed include intensification, iteration, tentativeness, positive affect, plurality, stance reinforcement, and onomatopoeic depiction.Source: §7 · pp. 27–28
  • Procrustes analysis finds that base-space organization is largely preserved, although local alignment ranges from elaborated core meanings to discourse-pragmatic changes such as a directive use of 走开走开.The study combines semantic profiling, shift vectors, Hungarian matching, and Procrustes analysis to examine these patterns and their transparency.Source: §7 · pp. 27–28
  • Static embeddings blend a form’s context-dependent senses and functions, making the reported semantic profiles approximate; contextualized embeddings are identified as a way to improve them.The blends are dominated by the most frequent senses and functions.Source: §7 · p. 28
  • The study concludes that distributional semantics offers a quantitative perspective on the systematicity and versatility of Mandarin reduplication and may be useful for studying other word-formation processes.Source: §7 · p. 28

Related papers

Loading 2609.16860v1…