SenWave: COVID-19 Multilingual Sentiment Dataset

• SenWave is a COVID-19 sentiment-monitoring system that provides a multilingual, fine-grained dataset for analyzing pandemic-specific emotions.
• It combines over 105 million posts from Twitter and Weibo, featuring manually annotated and translation-derived data across several languages.
• The system employs transformer-based classifiers and a domain-specific affect taxonomy to enable robust longitudinal, cross-lingual, and topical sentiment analysis.

SenWave is a COVID-19 sentiment-monitoring system and fine-grained multi-language sentiment dataset designed to analyze how public emotions evolved during the pandemic across languages, countries, and topics. It combines 105+ million collected tweets and Weibo messages with manually annotated multi-label sentiment data and transformer-based classifiers, with the explicit aim of moving beyond coarse positive/neutral/negative polarity toward pandemic-specific affective categories such as optimistic, anxious, denial, official report, and joking (Yang et al., 2020). In later work, it is also presented as a benchmark dataset comprising 10,000 annotated tweets each in English and Arabic, 30,000 translated tweets in Spanish, French, and Italian, and over 105 million unlabeled tweets collected during various COVID-19 waves (Yang et al., 9 Oct 2025).

1. Conceptual scope and problem setting

SenWave was introduced in response to a specific limitation of conventional sentiment analysis: pandemic discourse is inherently multi-label and event-conditioned. A single post may simultaneously express sadness and pessimism, joking and annoyance, or gratitude and optimism, making a three-class polarity scheme analytically inadequate (Yang et al., 2020). SenWave therefore frames COVID-19 affect analysis as fine-grained multi-label classification rather than ordinary single-label sentiment prediction.

The resource has two closely related identities. First, it is a monitoring system for longitudinal social sensing over a very large multilingual corpus. Second, it is a benchmark dataset for supervised fine-grained sentiment classification. The original monitoring formulation spans English, Spanish, French, Italian, Arabic, and Chinese, whereas the later dataset-centered formulation emphasizes five Twitter languages—English, Arabic, Spanish, French, and Italian—and treats Chinese Weibo separately in the original system because it uses a different annotation scheme (Yang et al., 2020, Yang et al., 9 Oct 2025).

This design places SenWave at the intersection of affective computing, crisis informatics, multilingual NLP, and social media epidemiology. A plausible implication is that its enduring value lies less in generic sentiment modeling than in providing an event-specific affect taxonomy for large-scale longitudinal inference under rapidly shifting public-health conditions.

2. Corpus construction and multilingual composition

SenWave’s large-scale collection layer was built from Twitter and Sina Weibo. For Twitter, the authors used Twint, an open-source Twitter crawler, with search terms including “covid-19,” “coronavirus,” “covid,” “corona,” and Arabic equivalents; data collection covered March 1 to May 15, 2020, and used 12 Twint instances on 24 cores (Yang et al., 2020, Yang et al., 9 Oct 2025). The original paper reports a collection rate averaging slightly over one million tweets per day and states that the combined multilingual corpus exceeds 105 million posts when Weibo is included (Yang et al., 2020).

The labeled and translated portions are organized as follows:

Resource	Scale	Notes
English tweets	10,000	Manually annotated
Arabic tweets	10,000	Manually annotated
Spanish, French, Italian tweets	10,000 each	Translated from annotated English tweets
Chinese Weibo posts	21,173	Separate 7-category annotation in the original system

The translation-based extension to Spanish, French, and Italian was produced with Google Translate and was intended to bootstrap training data for those languages rather than replace native annotation (Yang et al., 2020). The later dataset paper reports a BLEU score of 0.33 for the translation quality check and characterizes that quality as acceptable/good relative to a cited SOTA tied-transformer BLEU4 of 0.39 (Yang et al., 9 Oct 2025).

Preprocessing in SenWave removes URLs, emojis/emoticons, noisy symbols, line breaks, tabs, redundant spaces, and the retweet token RT, while retaining hashtags because they often carry the tweet topic (Yang et al., 2020, Yang et al., 9 Oct 2025). The original system uses NLTK for English, Spanish, French, and Italian, PyArabic for Arabic, and Jieba for Chinese segmentation (Yang et al., 2020). The later dataset paper also notes removal of usernames, gender, and @-mentions for privacy and noise reduction (Yang et al., 9 Oct 2025).

3. Annotation scheme and label semantics

The core SenWave label space consists of ten sentiment categories: optimistic, thankful, empathetic, pessimistic, anxious, sad, annoyed, denial, official report, and joking (Yang et al., 2020). These categories were developed after repeated refinement and, in the later dataset paper, are described as having been defined with input from four domain experts in public health and epidemiology (Yang et al., 9 Oct 2025).

Semantically, the categories are intended to isolate affective and discourse functions that are characteristic of pandemic communication. “Official report” captures factual government/WHO-style reporting, while “joking” captures irony, humor, memes, and ridicule; both are categories that standard sentiment corpora typically do not encode (Yang et al., 2020, Yang et al., 9 Oct 2025). The inclusion of “denial” is likewise domain-specific, targeting conspiracy-theory rejection or skepticism toward official narratives, depending on the formulation in the respective paper.

Annotation is explicitly multi-label. In the original paper, each tweet was labeled by at least three experienced annotators and the final label decision was made by majority voting under strict quality control; the authors state that they recruited more than 50 experienced annotators (Yang et al., 2020). The later dataset paper specifies that the annotation work was handled by Lucidya, that tweets were labeled by three annotators each, that 52 experienced annotators were recruited, and that tweets without overlap were re-annotated (Yang et al., 9 Oct 2025). It also reports Average Inter-rater Agreement and Cohen’s kappa: for English, $\iota = 0.904$ and $\kappa = 0.381$; for Arabic, $\iota = 0.931$ and $\kappa = 0.549$ (Yang et al., 9 Oct 2025).

SenWave’s multi-label character is not marginal. The original paper reports that more than 70% of English tweets have multiple labels, whereas about 20% of Arabic tweets have multiple labels (Yang et al., 2020). This makes English classification particularly challenging and helps explain why evaluation emphasizes set-overlap and ranking metrics rather than only exact-match accuracy.

One interpretive complication appears in later reuse. VADEC describes SenWave as containing 10K English tweets split into 8,000 train, 1,000 validation, and 1,000 test, with 11 binary emotion labels, but it does not list all 11 names in the main text (Mukherjee et al., 2021). By contrast, the original SenWave papers define a 10-category sentiment taxonomy (Yang et al., 2020, Yang et al., 9 Oct 2025). This suggests that different downstream papers may operationalize the benchmark slightly differently or refer to different preprocessed versions.

4. Modeling framework and benchmark performance

The original SenWave system uses the simpletransformers framework for multi-label sentiment classification by fine-tuning pretrained LLMs on the labeled data (Yang et al., 2020). Model choice is language-specific: XLNet for English, AraBERT for Arabic, BERT for Spanish, French, and Italian, and ERNIE for Chinese Weibo (Yang et al., 2020). Training uses batch size $16$, learning rate $4 \times 10^{-5}$, and 20 epochs, with 5-fold cross-validation on the labeled datasets before applying the trained classifiers to the unlabeled corpus (Yang et al., 2020).

Evaluation uses a multi-label suite including Accuracy, Jaccard Accuracy, F1-macro, F1-micro, LRAP, and Hamming loss (Yang et al., 2020, Mukherjee et al., 2021). The original paper explicitly defines Jaccard accuracy as a set-overlap measure and weak accuracy as a label-wise multi-label accuracy (Yang et al., 2020).

Reported performance in the original system includes weak accuracy about 0.847 and Jaccard about 0.495 for English; weak accuracy about 0.905 and Jaccard about 0.589 for Arabic; weak accuracy about 0.823 for Spanish, 0.824 for French, and 0.827 for Italian; and ERNIE accuracy about 0.88 for Chinese (Yang et al., 2020). The paper notes that F1 scores are only around 0.5 due to class imbalance, even though LRAP and Hamming loss indicate practical utility (Yang et al., 2020).

A later benchmark-oriented paper revisits the dataset with a different modeling stack: BART for English, AraBERT for Arabic, and BERT for Spanish, French, and Italian, with a custom classifier consisting of two MLP layers and binary cross-entropy loss (Yang et al., 9 Oct 2025). It reports validation results for English of $0.498 \pm 0.008$ Accuracy, $0.535 \pm 0.012$ F1-Macro, $0.580 \pm 0.008$ F1-Micro, $0.548 \pm 0.007$ LRAP, and $\kappa = 0.381$0 Hamming Loss, and concludes that BART performed best overall among the compared baselines (Yang et al., 9 Oct 2025). This suggests that SenWave has evolved from a single-system contribution into a reusable multilingual benchmark.

5. Longitudinal, cross-lingual, and topical findings

SenWave’s main empirical use is longitudinal monitoring of pandemic discourse. Across languages, the original paper reports a common temporal structure of rapid rise followed by slow decline, with sentiment curves highly correlated across languages (Yang et al., 2020). It further states that the global conversation reacts strongly negatively to herd immunity strategies and that optimistic and positive sentiments increased over time (Yang et al., 2020).

The language-wise analyses are more differentiated. In English, optimistic, thankful, and empathetic first rose and then fell, while sad and annoyed stayed relatively high; anxious and joking declined over time (Yang et al., 2020). In Arabic, optimistic rose, anxious, denial, and joking fell, and sad kept increasing in several Arabic-speaking countries (Yang et al., 2020). In French, thankful increased, joking declined, and denial rose sharply around conspiracy theories about the virus origin (Yang et al., 2020, Yang et al., 9 Oct 2025). In Spanish, optimistic and thankful rose while pessimistic and annoyed fell (Yang et al., 9 Oct 2025). Italian showed mostly weak trends, with thankful and empathetic exhibiting the clearest changes (Yang et al., 2020, Yang et al., 9 Oct 2025).

Country-level analyses cover locations such as the USA, UK, Spain, Argentina, and Saudi Arabia (Yang et al., 9 Oct 2025). The USA is characterized by negative emotions generally exceeding positive ones; the UK by negative reactions around panic buying and financial fear followed by more positive reactions after policy measures; Spain by applause for healthcare workers alongside anger and sadness around alarm extensions and death counts; Argentina by persistently high negative sentiment; and Saudi Arabia by stronger positive sentiment associated with government measures and Ramadan (Yang et al., 2020, Yang et al., 9 Oct 2025).

Topic-wise studies address stock market, oil price, herd immunity, economic stimulus, drug/medicine/vaccine, employment/job, and working from home in the original paper (Yang et al., 2020). Herd immunity is the clearest example of a strongly negative reaction pattern: denial and joking rose sharply when the UK government initially considered herd immunity on March 13, discussion intensified on March 14–15, annoyed increased later, and another denial spike appeared in early April (Yang et al., 2020). Working from home is described as the most positive topic, with optimistic sentiment above 40% and a clear upward optimistic trend over time (Yang et al., 2020).

A later analysis also isolates “joking” as a distinct and culturally variable class, reporting that Spanish had a higher proportion of joking than English or Arabic and that topic-specific co-occurrence patterns differ by language and theme (Yang et al., 9 Oct 2025). This reinforces SenWave’s central premise that pandemic discourse cannot be reduced to a single scalar polarity axis.

6. Reuse, transfer learning, and limitations

SenWave has been reused as a benchmark and transfer-learning resource in later work. In VADEC, it serves as the multi-label emotion-classification dataset for a multi-task framework that jointly trains emotion classification on SenWave and Valence-Arousal-Dominance regression on EMOBANK using a shared BERTweet encoder (Mukherjee et al., 2021). On the SenWave benchmark, VADEC improves the original baseline from 0.847/0.495/0.517/0.573/0.745/0.153 to 0.877/0.560/0.563/0.620/0.818/0.123 across Accuracy, Jaccard Accuracy, F1-Macro, F1-Micro, LRAP, and Hamming Loss, which the authors summarize as an 11.3% performance gain averaged over all six metrics (Mukherjee et al., 2021).

In HP-BERT, SenWave plays a second-stage transfer-learning role. A BERT model is first fine-tuned on a newly curated Hinduphobic COVID-19 X (Twitter) Dataset to build HP-BERT as a Hinduphobia detector, and then the same model is further fine-tuned on SenWave “to enable sentiment analysis” (Singh et al., 7 Jan 2025). This stage uses SenWave’s COVID-era sentiment taxonomy to bridge binary Hinduphobia detection and downstream multi-label sentiment analysis; the resulting sentiment model is evaluated with Hamming Loss $\kappa = 0.381$1, Jaccard Score $\kappa = 0.381$2, LRAP $\kappa = 0.381$3, Macro F1 $\kappa = 0.381$4, and Micro F1 $\kappa = 0.381$5 (Singh et al., 7 Jan 2025).

These reuses show that SenWave functions as more than a descriptive social-sensing dataset. It also serves as a sentiment-domain adaptation corpus for tasks that require pandemic-specific emotional labeling. A plausible implication is that its taxonomy is especially valuable when the target discourse is COVID-conditioned and affectively heterogeneous.

The dataset also has explicit limitations. The later paper notes that the 105M unlabeled tweets are smaller than billion-scale datasets such as BillionCOV, that the focus is mainly on the outbreak period, that post-COVID sentiment is not explored, that only the top five languages were covered in the Twitter component, that Twitter API collection may introduce sampling bias, that data may not represent the whole population, and that label imbalance remains a challenge (Yang et al., 9 Oct 2025). In addition, the Spanish, French, and Italian labeled sets are translation-derived rather than manually annotated from scratch (Yang et al., 2020, Yang et al., 9 Oct 2025). Finally, the ChatGPT compatibility study reports substantially lower performance for GPT-3.5 in zero-shot and few-shot settings—Accuracy 0.137 and 0.190, respectively—than for fine-tuned transformer models, indicating that SenWave’s label space is structured enough for prompt-based experimentation but still nontrivial for general-purpose LLMs without task-specific adaptation (Yang et al., 9 Oct 2025).

Taken together, these properties establish SenWave as a multilingual, multi-label COVID-19 sentiment resource whose main significance lies in its event-specific label design, large-scale longitudinal deployment, and continued reuse as a benchmark for affective modeling and domain transfer in social-media NLP.