Machine learning and digital pragmatics: Which word category influences emoji use most?

Mohammed Q. Shormani; Yehia A. AlSohbani; Mohammed Q. Shormani

Source-linked AI summary

Machine learning and digital pragmatics: Which word category influences emoji use most?

Mohammed Q. Shormani, Yehia A. AlSohbani, Mohammed Q. Shormani

arXiv:2608.21975v1Aug 22, 2026cs.CL

TL;DR

The study asks how lexical and structural features relate to emoji occurrence in Colloquial Arabic discourse on X and whether MARBERT can predict these categories. It combines MARBERT with logistic regression and finds that verbs show the strongest positive association with emoji occurrence, while hybrid analysis links emoji use to linguistic structure and contextual communication.

Problem
The study investigates relationships between lexical or structural features and emoji occurrence, their statistical associations, and MARBERT’s predictive performance in Arabic X discourse.
Method
Using 15856 emoji-containing X texts, the study combines MARBERT classification with binary logistic regression to analyze category identification and associations with emoji occurrence.
Results
Verbs show the strongest positive association with emoji occurrence (β = 0.821, p = .001; 95% CI [0.332, 1.309]) and the strongest MARBERT performance (F1 = .85).
Takeaways & Limitations
Emoji use in Arabic X discourse is systematically related to linguistic structure and contextualized communication rather than lexical frequency alone.
Takeaways & Limitations
Automatic annotation may be inaccurate for morphologically complex Arabic and informal social-media language, while restricting data to X limits generalizability to other platforms.

Abstract

from arXiv · show

This study examines the performance of the state-of-the-art MARBERT model in identifying the lexical/pragmatic category associated with emoji use on X within a digital pragmatics approach (DPA). A net corpus of 15856 Colloquial Arabic (CA) posts containing emojis was collected from X using Python. The texts were tokenized and normalized into 4 lexical categories, namely noun_norm, verb_norm, adj_norm, and adverb_norm, and 2 pragmatic/structural categories, question_norm and exclamation_norm. MARBERT was finetuned and optimized to identify which category scores standard metrics more, hence associated with emoji use, while binary logistic regression was used to examine which category is statistically associated with emoji occurrence. Findings unveil that nouns dominate the corpus in normalized frequency (M = 0.675, SD = 0.161), followed by verbs (M = 0.083, SD = 0.100). However, verbs have the strongest influence of emoji use indicated by verb density (\b{eta} = 0.821, p = .001, 95% CI [0.332, 1.309]). The study concludes that in digital pragmatics of CA on X, emoji use association with lexical/pragmatic category can be explained by a hybrid approach of computational, statistical, and pragmatic methods, reflecting the interaction among machine learning, linguistic/lexical features, contextual representation, and pragmatic communication.

1. Introduction

The study examines how lexical and structural features relate to emoji occurrence in Colloquial Arabic digital discourse on X, combining statistical analysis with MARBERT modeling. It addresses whether specific linguistic categories are associated with emoji use and whether MARBERT can identify those categories computationally.

1. Introduction: Emojis function as meaningful visual resources that complement written language in multimodal digital communication.Digital communication increasingly combines textual language with visual symbols that contribute to message interpretation.
1. Introduction: X’s word-limit affordances encourage users to employ emojis alongside linguistic, pragmatic, and multimodal resources.The platform’s limited post length is presented as one reason users may resort to visual signs.
1. Introduction: Digital pragmatics treats emoji use as part of the interaction between textual pragmatics and visual communication.Emojis may reinforce, clarify, or alter meanings that are underspecified without tone, facial expression, or prosody.
1. Introduction: The study combines linguistic analysis of lexical and structural features with MARBERT modeling of emoji occurrence.The statistical component estimates associations, while the computational component models category identification from the same textual data.
1. Introduction: The research asks which features are statistically associated with emojis and whether MARBERT can predict relevant categories using standard classification metrics.These questions address both interpretable linguistic associations and computational predictive performance.

2. Theoretical foundations and previous studies

The paper situates emoji use within machine learning, digital discourse, and digital pragmatics, emphasizing context-sensitive multimodal meaning in Colloquial Arabic. It motivates a hybrid framework in which computational models detect patterns while statistical and pragmatic analysis interpret them.

2.1. Machine learning: Machine learning infers patterns from data to make predictions or decisions, while deep learning uses multilayered neural networks to learn representations.Traditional methods remain relevant for interpretability and hypothesis testing, whereas transformer architectures support language-related tasks.
2.1. Machine learning: Combining machine learning with statistical models supports both pattern prediction and explanation of relationships in digital discourse.The approaches serve complementary objectives: computational detection and theoretically interpretable analysis.
2.2. Digital discourse and pragmatics: Digital discourse is interactive, multimodal, rapid, and shaped by platform affordances such as character limits and interface design.These conditions encourage efficient strategies that combine written language with visual elements such as emojis.
2.2. Digital discourse and pragmatics: Digital pragmatics studies how visual elements convey and shape pragmatic meaning in context, especially when online communication lacks face-to-face cues.Its interpretation is anchored in local sociolinguistic contexts, dialectal practices, and community conventions rather than fixed universal meanings.
2.4. Computational modeling and digital pragmatics: The study uses MARBERT to assess contextual category identification and logistic regression to estimate associations between linguistic features and emoji occurrence.This integration connects computational learnability with pragmatic interpretation.
2.3. Colloquial Arabic: Lexical categories represent propositional content, whereas question and exclamation forms are treated as indicators of interactional meaning.Colloquial Arabic is conceptualized as Arabic dialect acquired at home rather than Standard Arabic.
2.5. Previous studies and research gap: The study addresses a gap through a corpus-based examination of whether word category influences emoji use in colloquial digital discourse.It contributes empirical evidence concerning interaction between textual and visual communication.

3. Methods

The study analyzes Colloquial Arabic posts from X through a supervised pipeline combining linguistic preprocessing, MARBERT modeling, and statistical evaluation. Texts were categorized morphosyntactically and evaluated with complementary classification metrics.

Data collection: 15,856 of 28,778 collected Colloquial Arabic posts contained emojis and formed the study corpus.Data were retrieved from X using Python-based tools via API v2.
Preprocessing: The preprocessing pipeline removed URLs, mentions, hashtag-identification parts, non-linguistic symbols, duplicate posts, empty posts, and irrelevant entries.Cleaning used Python libraries including re and pandas to improve consistency and protect privacy.
Preprocessing: Texts were tokenized and normalized into noun_norm, verb_norm, adj_norm, adverb_norm, question_norm, and exclamation_norm categories before training and testing splits.The first four categories were lexical, while the final two were pragmatic categories.
Data annotation: CAMeL Tools assigned part-of-speech and morphological information at token and sentence levels, producing the normalized feature variables.Annotations covered nouns, verbs, adjectives, adverbs, interrogative forms, and exclamatory forms.
Modeling: MARBERT was fine-tuned on normalized textual representations, with validation guiding optimization and an unseen evaluation dataset assessing generalization.Training-set parameters were used exclusively for feature standardization.
Evaluation: Accuracy, precision, recall, F1-score, and weighted F1-score evaluated category-level classification, with weighted F1 accounting for category support.Because category distributions were substantially imbalanced, per-category metrics were reported alongside weighted F1 rather than accuracy alone.

4. Results

The results distinguish category frequency from emoji association and classification performance. Nouns are most frequent, whereas verbs show the strongest positive association with emoji occurrence and the best MARBERT performance.

Descriptive statistics: Nouns had the highest mean normalized frequency (M = 0.675, SD = 0.161), followed by verbs (M = 0.083, SD = 0.100).Adjectives, adverbs, questions, and exclamations occurred substantially less frequently.
Logistic regression: Verb density showed a significant positive association with emoji occurrence (β = 0.821, p = .001, 95% CI [0.332, 1.309]).Adjective and question densities were also significantly positive, while noun and adverb densities were significantly negative.
MARBERT performance: MARBERT performed best for verbs (precision = .74, recall = .99, F1 = .85), followed by nouns (precision = .88, recall = .44, F1 = .59).Adverbs had F1 = .05, while adjectives, questions, and exclamations received zero precision, recall, and F1.

5. Discussion

The study combines logistic regression and MARBERT to examine how linguistic categories relate to emoji occurrence and how reliably those categories can be identified from contextual text. Results show that statistical association, lexical frequency, and computational predictability capture distinct aspects of emoji-related language use.

Statistical associations: Logistic regression found significant negative associations for noun and adverb density and significant positive associations for verb, adjective, and question density.Exclamation density was positive but statistically non-significant.
Statistical associations: Verbs had the strongest positive association with emoji occurrence, with β = 0.821, p = .001; 95% CI [0.332, 1.309].Noun density was negatively associated with emoji occurrence, with β = −1.477, p = .001; 95% CI [−2.356, −0.599].
MARBERT classification: MARBERT performed best for verbs, achieving accuracy = .76, precision = .74, recall = .99, and F1 = .85.It also identified nouns with accuracy = .58, precision = .88, recall = .44, and F1 = .59, while weaker performance occurred for less frequent categories.
MARBERT classification: Adverbs achieved precision = .76, recall = .02, and F1 = .05, while adjectives, questions, and exclamations received zero precision, recall, and F1.These results indicate difficulty identifying sparsely represented categories from textual input.
Digital-pragmatic interpretation: The findings support an associational, multimodal account in which emojis function as pragmatic resources linked to linguistic structure and discourse function.The analysis does not establish that any linguistic category causes emoji use.
Hybrid interpretation: The study treats MARBERT and logistic regression as complementary methods rather than interchangeable evidence.Logistic regression estimates interpretable associations, whereas MARBERT evaluates category identification from contextualized text.

6. Conclusions and limitations

The study concludes that emoji occurrence in Arabic digital discourse is systematically related to linguistic structure and discourse function, and is best examined through a hybrid computational and statistical framework. Its conclusions are bounded by annotation, platform, category-sparsity, and contextual-complexity limitations.

Conclusions: The study uses MARBERT and logistic regression to investigate whether six linguistic categories systematically relate to emoji use.The dataset contains 15856 tweet-emoji texts used for both analyses.
Conclusions: Emoji occurrence is systematically related to linguistic structure and discourse function rather than randomly distributed across textual contexts.The conclusion frames emoji use as a phenomenon shaped by linguistic contexts and pragmatic communication.
Conclusions: MARBERT captures contextual and semantic patterns, while logistic regression provides transparent evidence of associations between linguistic structures and emoji occurrence.The hybrid framework connects contextual learning with interpretable lexical-statistical modeling.
Model performance: MARBERT performs comparatively well for verbs with precision = .74, recall = .99, and F1 = .85, but less frequent categories perform substantially worse.Nouns achieve precision = .88, recall = .44, and F1 = .59; adverbs achieve precision = .76, recall = .02, and F1 = .05.
Limitations: Automatic linguistic annotation may be inaccurate because Arabic is morphologically complex and social-media language is informal and variable.This limitation concerns the reliability of the linguistic features used in the analyses.
Limitations: Because the dataset is restricted to X, the findings may not generalize to platforms with different communicative norms, user practices, and technological affordances.The authors also identify sparse-feature recognition and incomplete capture of contextual complexity as limitations.