Twitter Sentiment Analysis

This is an entity-level Twitter Sentiment Analysis dataset. For each message, the task is to judge the sentiment of the entire sentence towards a given entity. For example, A outperforms B is positive for entity A but negative for entity B. The dataset contains ~70K labeled training messages and 1K labeled validation messages. It is available online for free on Kaggle.

RETWEET is a dataset of tweets and overall predominant sentiment of their replies. SUMMARY ------ WHAT: Message-level Polarity Classification. GOAL: To predict the predominant sentiment among (potential) first-order replies to a given tweet. IDEA: Mitigate the problem of lacking labeled training data wi treating the unsupervised nature of the problem as a supervised learning case. ### APPROACH: 1. Train a tweet classifier. 2. Automatically label the replies using the classifier trained in the first part. 3. Choose a final label representing the general predominant sentiment of the replies of every tweet. ### DATA COLLECTION To download all of the replies to a tweet, the Search API should be used. However, the Search API is limited to 75000 requests per hour, which causes the mining and downloading process to be slow. Furthermore, using the Twitter API, there is no possibility of downloading absolute random data. Therefore, we try to make the procedure as random as possible by utilizing two different strategies for data downloading and using them in an intermixed manner. 1. Our first strategy is based on a sample of English tweets obtained by filtering the Twitter stream via a list of cultural keywords. This list consists of 147 words that are deemed to play a "pivotal role in discussions of culture and society", covering diverse words such as aesthetics, environment, feminism, power, tourism, or youth. We extracted all tweets in 2019 that have a minimum of 20 first-order replies in the dataset. The data come with an obvious caveat: Both the source tweet as well as all the replies must contain at least one word from the list of keywords. Therewith, it is highly unlikely that the list of replies for any given source is exhaustive, i.e. there might be many more first-order replies to the source tweet that are not in the dataset. 2. As our second approach, we use the GetOldTweets3 library to download all the replies corresponding to every tweet. We define few restrictions to add randomization to the process. Firstly, every tweet and also every reply should contain at least 20 strings. This is due to the fact that our automatic tweet classifier, explsined in the paper, is optimized based on the message-level classification paradigm. Therefore, it operates optimal when the input contains at least a sufficient number of words. The second constraint is that every tweet should contain at least 20 first-order replies. In order to increase randomness, in this strategy, instead of referencing to a list of keywords, we manually choose some keywords, which are most likely to include long discussions, such as Coronavirus and football or the ones, which are most likely to include strong opinions such as birthday, war, or racism in order to account for the easy-to-guess examples. ### MANUAL ANNOTATIONS FOR THE RETWEET (TEST GOLD DATASET) 5,015 tweets with their corresponding replies, collected as a combination of the two different collection strategies, were given to three different students. Each of them had to read all the replies corresponding to every tweet, without observing the original tweet in order to avoid having a prior knowledge, and decide on ONE final sentiment for the replies. The assigned sentiment can only be one of the positive, negative, or neutral labels. Considering the fact that this is a really challenging task for the machine, to prevent human mistakes, we correlated the results of the three annotators and only chose the tweets, in which all of the annotators had the same opinion on the labels, as the final gold standard test data. Therefore, we finally, ended up with a test set consisting of 1,519 human labeled tweets, with the labels being the sentiment of the replies of a tweet and not the tweet itself. DATASET CONTENTS --- 1. Training raw dataset: 34,953 unique tweets in total and individual automatic labels for all of their corresponding replies (1,519,504 total replies). Including, - ./RETWEET_data/train_reply_labels_set1.txt - ./RETWEET_data/train_reply_labels_set2.txt 2. Training autamtically-labeled dataset: 34,953 unique tweets and ONE final automatic label (chosen based on the algorithm 1 of our paper) for every tweet. Including, - ./RETWEET_data/train_final_label.txt 3. Gold standard test dataset (RETWEET): 1,519 unique tweets with their manual labels for replies. ONE final label, which states the predominant overall polarity of all its replies, is assigned to every tweet. Including, - ./RETWEET_data/test_gold.txt NOTES --- 1. Please note that by downloading the Twitter data you agree to abide by the Twitter terms of service, and in particular you agree not to redistribute the data and to delete tweets that are marked deleted in the future. 2. The "neutral" label in the annotations stands for objective or neutral. 3. The distribution consists of a set of Twitter unique tweet IDs with annotations (overall polarity of replies). As for data privacy, the texts of the tweets and replies are not distributed. But as all the utilized resources in this dataset are taken from public tweets, having the tweet unique IDs, you can download the tweet and its replies. You can use the Semeval Twitter data downloading script to obtain the corresponding tweets: https://github.com/seirasto/twitter_download/ 4. The dataset URL: https://kaggle.com/soroosharasteh/retweet/ LICENSE --- The accompanying dataset is released under a Creative Commons Attribution 4.0 International License. SOURCE CODE --- The official source code of the paper: https://github.com/starasteh/retweet ### In case you use this dataset, please cite the original paper: S. Tayebi Arasteh, M. Monajem, V. Christlein, P. Heinrich, A. Nicolaou, H.N. Boldaji, M. Lotfinia, S. Evert. "How Will Your Tweet Be Received? Predicting the Sentiment Polarity of Tweet Replies". Proceedings of the 2021 IEEE 15th International Conference on Semantic Computing (ICSC), Laguna Hills, CA, USA, January 2021. ### BibTex @inproceedings{RETWEET, title = "How Will Your Tweet Be Received? Predicting the Sentiment Polarity of Tweet Replies", author = "Tayebi Arasteh, Soroosh and Monajem, Mehrpad and Christlein, Vincent and Heinrich, Philipp and Nicolaou, Anguelos and Naderi Boldaji, Hamidreza and Lotfinia, Mahshad and Evert, Stefan", booktitle = "Proceedings of the 2021 IEEE 15th International Conference on Semantic Computing (ICSC)", address = "Laguna Hills, CA, USA", pages = "370-373", doi = "10.1109/ICSC50631.2021.00068", url = "https://ieeexplore.ieee.org/document/9364527/", month = "01", year = "2021" } * Dataset DOI: 10.34740/kaggle/ds/736988 * Paper: https://ieeexplore.ieee.org/document/9364527 * Paper DOI: 10.1109/ICSC50631.2021.00068 CONTACT --- E-mail: soroosh.arasteh@fau.de DATA FORMAT FOR ALL THE FILES --- label TAB id where, "label" can be positive, neutral or negative, corresponding to the overall message-level polarity of the replies of the tweet and "id" corresponds to the Twitter unique ID for the tweets.

Sentiment detection remains a pivotal task in natural language processing, yet its development in Arabic lags due to a scarcity of training materials compared to English. Addressing this gap, we present ArSen-20, a benchmark dataset tailored to propel Arabic sentiment detection forward. ArSen-20 comprises 20,000 professionally labeled tweets sourced from Twitter, focusing on the theme of COVID-19 and spanning the period from 2020 to 2023. Beyond tweet content, the dataset incorporates metadata associated with the user, enriching the contextual understanding. ArSen-20 offers a comprehensive resource to foster advancements in Arabic sentiment analysis and facilitate research in this critical domain. The ArSen-20 dataset statistics: | Statistics | Num | |:-------------:|:-----:| | Training set size | 16000 | | Validation set size| 2000 | | Testing set size | 2000 | | Neutral | 17262 | | Positive | 878 | | Negative | 1860 | ## Features The dataset has the following features: | Field | Type | Description | |:-----------:| :--------: |:----------------: | | tweet id | string | The unique identifier of the requested Tweet. | | label | string | Sentiment Classification of this tweet. | | author id | string |The unique identifier of this user. | | created_at | data | Creation time of the Tweet. | | lang | string | Language of the Tweet, if detected by Twitter. | | like_count | int |The number of likes on this tweet.| |quote_count | int | The number of times this tweet has been quoted. | | reply_count | int | The number of replies to this tweet. | | retweet_count| int | The number of retweets to this tweet. | | tweet | string | The actual UTF-8 text of the Tweet. | |user_verified | boolean | Indicates if this user is a verified Twitter User. | |followers_count | int |The number of followers of the author. | | following_count | int | The number of following of the author. | | tweet_count | int | Total number of tweets by the author. | | listed_count | int |The number of public lists that this user is a member of. | |name | string | The name of the user. | | username | string | The Twitter screen name, handle, or alias. | | user_created_at| data | The UTC datetime that the user account was created. | | description | string | The text of this user’s profile description (bio). | ## DownLoad You can download the dataset from here. + ArSen-20_publish.csv - Contains all features. + ArSen-20_id_only.csv - Contains only tweets and their author's id. ## Citation If you use this dataset in your research, please cite the following papers: bibtex @inproceedings{fang2024arsen, title={ArSen-20: A New Benchmark for Arabic Sentiment Detection}, author={Yang Fang and Cheng Xu}, booktitle={5th Workshop on African Natural Language Processing}, year={2024}, url={https://openreview.net/forum?id=GgsRUF5kJt} } bibtex @inproceedings{fang2024advancing, title = "Advancing {A}rabic Sentiment Analysis: {A}r{S}en Benchmark and the Improved Fuzzy Deep Hybrid Network", author = "Fang, Yang and Xu, Cheng and Guan, Shuhao and Yan, Nan and Mei, Yuke", editor = "Barak, Libby and Alikhani, Malihe", booktitle = "Proceedings of the 28th Conference on Computational Natural Language Learning", month = nov, year = "2024", address = "Miami, FL, USA", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2024.conll-1.39", pages = "507--516", } ## contact If you have any questions or comments about the dataset, please contact Yang Fang (20211209024@chnu.edu.cn). Potential cooperation in related fields is also welcome. :)

#chinahate dataset contains a total of 2,172,333 tweets hashtagged #china posted during the time it was collected. It is designed for the task of hate speech detection.

The dataset contains 30 million cryptocurrency-related tweets from 10.10.2020 to 3.3.2021. See https://github.com/meakbiyik/ask-who-not-what for more details.

This dataset was created as part of the Master's thesis titled "Multi-Class Depression Detection Through Tweets Using Artificial Intelligence." It contains tweets labeled for five types of depression (Bipolar, Major, Psychotic, Atypical, and Postpartum) using lexicons verified by psychiatrists. Click to add a brief description of the dataset (Markdown and LaTeX enabled).

SentimentArcs’ reference corpus for novels consists of 25 narratives selected to create a diverse set of well recognized novels that can serve as a benchmark for future studies. The composition of the corpora was limited by the effect of copyright laws as well as historical imbalances. Most works were obtained from US and Australian Gutenberg Projects. The corpora is expected to grow in size and diversity over time. Several dimensions of diversity were considered for inclusion including popularity, period, genre, topic, style and author diversity. The first version of our corpus includes only English, although Proust and Homer are included in translation. SentimentArcs has processed a larger set of novels, including some in foreign languages. The initial reference corpus is in English since performance across all ensemble models was uneven in less resourced languages In sum, the corpora includes (1) the two most popular novels on Gutenberg.org (Project Gutenberg, 2021b), (2) eight of the fifteen most assigned novels at top US universities (EAB, 2021), and (3) three works that have sold over 20 million copies (Books, 2021). There are eight works by women, two by African-Americans and five works by two LGBTQ authors. Britain leads with 15 authors followed by 6 Americans and one each from France, Russia, North Africa and Ancient Greece.