Bilingual Dataset of Multinational Brand Posts and Consumer Replies from Platform X (Turkish and English)

Overview This dataset consists of posts and consumer replies from a curated selection of Fortune 500 companies that actively maintained Platform X (formerly Twitter) accounts in both English and Turkish. The sample includes 33 multinational brands spanning multiple industries, including fast-moving consumer goods (FMCG), fast food, technology, automotive, apparel, retail, finance, and logistics. All Platform X messages posted by these brands, together with corresponding consumer responses, were collected over a five-year period (June 2016 – June 2021). The dataset captures both brand-initiated communication and public engagement, enabling large-scale analysis of emotional framing, and audience response patterns. The full raw dataset contains: English: 8,101,034 messages 2,708,874 brand posts 5,392,160 consumer replies Turkish: 735,459 messages 245,422 brand posts 490,037 consumer replies The matrices released here are derived representations of these messages, designed to support reproducible analysis while complying with platform redistribution policies. What Is Released This Zenodo deposit provides sparse frequency matrices representing: Emoji usage per message Seed-word usage per message These representations enable interpretable analyses of: emoji-based emotional signaling, lexical emotion framing, emoji–word co-usage patterns, cross-language comparisons. Tweet texts and tweet identifiers are not embedded in the matrices. Instead, deterministic alignment is ensured through explicit row identifiers. Data Contents The dataset is organized by language (tr, en). For each language, the following files are provided. 1. Frequency Matrices {lang}_emoji_matrix_X.npzSparse emoji frequency matrix(rows = messages, columns = emoji features) {lang}_seedword_matrix_X.npzSparse seed-word frequency matrix(rows = messages, columns = lexical features) Matrices are stored in SciPy sparse CSC format for memory efficiency and scalable matrix operations. 2. Metadata Files {lang}_emoji_matrix_meta.npz {lang}_seedword_matrix_meta.npz Each metadata file contains: feature_names – ordered list of emoji or seed-word features corresponding to matrix columns row_id – sequential integer identifiers (0…N−1) defining row alignment The row_id ordering is identical across emoji and seed-word matrices within the same language. 3. Tweet Identifier Mapping {lang}_tweet_ids.csv Contains: row_id – row identifier used in frequency matrices tweet_id – Platform X message identifier This file provides the only linkage between frequency matrices and original messages and allows tweet rehydration where permitted. File Structure . ├── tr_emoji_matrix_X.npz ├── tr_emoji_matrix_meta.npz ├── tr_seedword_matrix_X.npz ├── tr_seedword_matrix_meta.npz ├── tr_tweet_ids.csv ├── en_emoji_matrix_X.npz ├── en_emoji_matrix_meta.npz ├── en_seedword_matrix_X.npz ├── en_seedword_matrix_meta.npz ├── en_tweet_ids.csv Seed Words and Emotion Lexicons Seed words were selected from crowd-sourced emotion lexicons. For each language, the top-scored 90 words per emotion were retained to construct the seed-word frequency matrices. English Seed Words Derived from the NRC Emotion Lexicon: Mohammad, S., & Turney, P. (2013).Crowdsourcing a Word-Emotion Association Lexicon.Computational Intelligence, 29(3), 436–465. Turkish Seed Words Derived from the Turkish Emotion Lexicon: Toçoğlu, M. A., & Alpkoçak, A. (2019).Lexicon-based emotion analysis in Turkish.Turkish Journal of Electrical Engineering and Computer Sciences, 27(2), 1213–1227.https://doi.org/10.3906/elk-1807-41 These lexicons enable language-specific yet conceptually comparable emotion analysis. Loading the Data (Python) Requirements numpy scipy pandas Load Emoji Frequency Matrix import numpy as np import scipy.sparse as sp X_emoji = sp.load_npz("tr_emoji_X.npz") emoji_meta = np.load("tr_emoji_meta.npz", allow_pickle=True) emoji_features = emoji_meta["feature_names"] row_id = emoji_meta["row_id"] Load Seed Word Frequency Matrix X_seed = sp.load_npz("tr_seed_X.npz") seed_meta = np.load("tr_seed_meta.npz", allow_pickle=True) seed_features = seed_meta["feature_names"] Load Tweet Identifier Mapping import pandas as pd df_tweets = pd.read_csv("tr_tweet_ids.csv", dtype={"tweet_id": str}) # Optional integrity check assert (df_tweets["row_id"].to_numpy() == row_id).all() Emoji–Word Co-Frequency via Dot Product To quantify co-usage patterns between emojis and seed words, a dot product can be applied between the emoji and seed-word frequency matrices. Let: X_emoji ∈ ℝⁿˣᵉ (messages × emojis) X_seed ∈ ℝⁿˣʷ (messages × words) The co-frequency matrix is computed as: C=Xemoji⊤XseedC=Xemoji⊤Xseed Where: C[i, j] represents the total number of messages in which emoji i and seed word j co-occur. Compute Co-Frequency Matrix C = X_emoji.T @ X_seed Interpret Results import pandas as pd cofreq_df = pd.DataFrame( C.toarray(), index=emoji_features, columns=seed_features ) # Example: seed words most frequently co-occurring with a given emoji cofreq_df.loc["🙏"].sort_values(ascending=False).head(10) This representation supports emotion framing analysis, emoji-anchored discourse interpretation, and interpretable CSR communication studies. Code Availability All preprocessing, analysis, and modeling code used in the associated paper is available at: GitHub repositoryhttps://github.com/atanaltay/beyond_words_emoji_patterns.git The repository includes scripts for: data preprocessing and filtering, matrix construction, emoji–word co-frequency analysis, downstream statistical and visualization analyses. Reproducibility Notes All matrices are raw frequency counts (not TF-IDF). Feature ordering is explicitly stored. row_id ensures deterministic alignment across files. Tweet identifiers are distributed separately to comply with Platform X policies. No implicit assumptions about row ordering are required. Citation If you use this dataset, please cite the associated paper and Zenodo record: Citation details will be updated upon publication. License and Usage This dataset is released for academic research and non-commercial use.Users are responsible for complying with Platform X content redistribution and rehydration policies.