Turkish Twitter Dataset on "Sessiz İstila" (Silent Invasion) and Syrian Immigration (2021–2022)

Overview
This dataset contains Turkish tweets collected between 2021 and 2022 surrounding the topic of "Sessiz İstila" (Silent Invasion) and public discourse regarding Syrian immigrants and refugees in Turkey. The dataset is designed to support research on the discourse surrounding the “sessiz istila” phrase, on emotion analysis for Turkish, and on social media dynamics. Matching the keyword does not by itself make a post hostile to refugees, and the corpus should not be read as a collection of xenophobic speech: in a random sample of 400 distinct normalized posts, 56.2% endorse or amplify the framing, 24.2% mention it without evaluation, 13.5% argue against it or defend refugees, and 6.0% are too fragmentary to code.

Dataset Structure & Files
The dataset includes the following files:
`silentinvasionv4_2021_2022_with_emotions_v14.csv`: Main dataset containing scraped tweets from 2021 to 2022, enriched with emotion annotations/labels and sentiment analysis.
`silentinvasionv4_2021_2022_with_emotions_v14_EN.csv`: Main dataset containing scraped tweets from 2021 to 2022, enriched with emotion annotations/labels and sentiment analysis. (English translation)
`emojis_emoticons_Turkish.csv`: A reference/mapping file for Turkish emojis and emoticons used in text preprocessing and sentiment processing.
`TREMODATAcalibrationApr2024.csv`: Calibration data utilized for Turkish emotion classification and model validation (April 2024 update).
`TREMODATAcalibrationApr2024_EN.csv`: Calibration data utilized for Turkish emotion classification and model validation (April 2024 update). (English translation)
`softmax_scores_corpus.csv`: Per-post softmax probability distribution over the six emotion categories for all 47,024 posts of the main corpus, row-aligned with `silentinvasionv4_2021_2022_with_emotions_v14.csv` through the `created_at` field. Columns: `created_at`, `prob_happy`, `prob_fear`, `prob_sadness`, `prob_disgust`, `prob_surprise`, `prob_anger`, `max_prob`, `label_at_0.6`. The file lets users apply a confidence threshold other than the 0.6 used for the released labels without re-running the model; at 0.6 it reproduces the released label column exactly, giving 40,880 labelled and 6,144 ambiguous posts.
`model_tremo_v13`: The fine-tuned checkpoint that produced the released emotion labels, saved with torch.save(model.state_dict(), ...). Load it into a BertForSequenceClassification built on dbmdz/bert-base-turkish-cased with num_labels=6; label ids are 0 = happy, 1 = fear, 2 = sadness, 3 = disgust, 4 = surprise, 5 = anger.

Key Features & Fields
Temporal Scope: 2021–2022
Language: Turkish (tr) & English (en)
Annotations: Emotion labels.

Code Availability
The code that produced these files is openly available at https://github.com/dwicak/TwitterScrapping-for-Sezzis-Istila: two R Markdown scripts covering the retrieval call, the normalization rules and the emoji substitution, and a Python notebook that fine-tunes the emotion recognition model and runs the inference that produced the emotion labels. The Turkish emoji lexicon released here is maintained at https://github.com/dwicak/Turkish-Emoji.

Usage & Ethical Considerations
This dataset is provided strictly for academic, educational, and research purposes.
Re-users must comply with Twitter/X Terms of Service and data protection/privacy regulations regarding social media data distribution.