About ExtremeSentiLex
Two people can disagree with the same thing online and sound completely different doing it. ExtremeSentiLex is a research collection of public social media posts, labelled for whether the sentiment is positive or negative and also for how strong it is. Everything here was publicly visible when we collected it. Below you'll find what the collection contains, where it came from, what the labels mean (and what they don't), and how we handle the fact that real people wrote these posts.
What this is
Most sentiment analysis sorts text into three groups: positive, negative and neutral. "I didn't love it" and "this is an absolute disgrace" both end up as negative, even though they say very different things. For anyone studying polarisation, online hostility or reactions to an event, that difference usually matters.
This collection uses a five-point scale instead, with separate labels for extreme positive and extreme negative sentiment. Each post has a polarity value, a language code and one of twenty topic categories. We cover many topics, from politics to food, because the behaviour we study shows up in all of them.
Where it came from
The corpus and the collection pipeline behind it, SocialNetCrawler, are outputs of Project MOVES - Virtual Monitoring of Crowds in Smart Cities (project description, PDF) - carried out at Universidade da Beira Interior (UBI) within HULTIG - Human Language Technology and Bioinformatics Group, led by Sebastião Augusto Rodrigues Figueiredo Pais and funded by Fundação para a Ciência e a Tecnologia (FCT), project PTDC/EEI-AUT/28918/2017 (1 July 2018 – 30 June 2021).
MOVES set out to model how collective behaviour forms on social media around an emerging event. To do that, the project needed a large, multilingual set of annotated posts, and building it took considerable work in itself. That set is what you're searching here. It records intensity as well as direction because the MOVES research needed to measure intensity.
Collection has continued since the funded project ended. We publish the corpus separately from the project's results because it is also useful for research that has nothing to do with crowds but needs annotated, multilingual social media text.
Where the data comes from
Posts arrive through SocialNetCrawler, a collection pipeline built at the University. It requests publicly visible content that matches a set of research topics from each platform. Where a platform offers suitable public access, it uses that platform's API: the AT Protocol for Bluesky and the public APIs of Mastodon, Tumblr, Reddit and YouTube. For the other platforms it reads publicly visible pages with an automated browser, seeing only what anyone can see without logging in. It currently covers seven networks:
Public posts only. We only collect a post if anyone could read it without logging in, following the author or joining a group.
That rules out protected accounts, direct messages, friends-only and follower-only posts, closed or invite-only groups, anything behind a paywall or age gate, and anything that would have required signing in as a member of a restricted space to see.
Collection has been running since 2018. If a post was public when collected and its author later made it private or deleted it, see Data protection for how to ask us to remove it from public view.
What we store
The database stores more fields than the search page displays. The full list is below, and the last column shows which fields are published.
| Field | What it is | Shown here |
|---|---|---|
Post |
The text of the post, as published. | Yes |
CreatedAt |
Date and time the post was published on the source platform. | Yes |
URL |
Permanent link to the post on the source platform. | Yes |
Language |
ISO 639-1 code, detected automatically. | Yes |
Polarity |
Sentiment value from −2 to +2, assigned automatically. | Yes |
Category |
Topic category (see below). Used as a filter. | Filter only |
SNetwork_ID |
Which platform the post came from. | Filter only |
Original_ID |
The platform's own identifier for the post. Used internally to avoid storing the same post twice. | No |
Author reference |
The platform account associated with the post, kept so that duplicates and threads can be resolved. | No |
Media links |
URLs of images or video attached to the post. Media files themselves are not downloaded or stored. | No |
Research tag |
The internal search term that caused the post to be collected. | No |
What is not collected
- Private or restricted content of any kind (see Where the data comes from). Only publicly visible posts are stored.
- Email addresses, phone numbers, IP addresses or device identifiers of post authors.
- Follower graphs, friend lists, or any reconstruction of a person's social network.
- Media files (images, video, audio) - only the links to them.
- Any data about you, the visitor, beyond what is needed to run the service: this site sets no tracking cookies and uses no analytics. Search terms are not associated with visitors.
The polarity scale
Every post gets one of these five values. The colours match the charts on the search page.
Strongly favourable language - enthusiasm, praise, celebration, often with intensifiers or emphatic markers.
Mildly favourable language - approval or satisfaction, without marked intensity.
No clear sentiment detected: factual statements, questions, links, or text the model could not resolve either way.
Mildly unfavourable language - criticism, disappointment, disagreement.
Strongly unfavourable language - hostility, outrage, condemnation, often with intensifiers or profanity.
How the value is assigned
Automatically, by the ExtremeSentiLex method (Pais, S., Tanoli, I., Albardeiro, M., Cordeiro, J. "Unsupervised Approach to Detect Extreme Sentiments on Social Networks." 2020 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 651–658. 2020. DOI: 10.1109/ASONAM49781.2020.9381420), which combines lexical resources with machine learning to place each post on the scale above. No one reads the posts before they are added; with this many posts, manual review isn't possible.
Automatic labelling is what makes a corpus of this size possible. It is also the source of the limitations listed below.
The labels are estimates and are wrong more often than you might expect. The errors also follow patterns, so they don't cancel each other out. The main problems are:
- Sarcasm often fools it. "Oh, brilliant work everyone" is scored as praise.
- It can't tell when someone is quoting. A person who repeats an insult in order to condemn it is scored as if they meant it.
- Accuracy varies by language, roughly according to how many lexical resources exist for each one. Treat results in less-resourced languages with more caution.
- Neutral is partly a catch-all. It contains genuinely neutral text as well as text the method couldn't classify.
- Swearing pushes scores towards the extremes. Friends joking with each other in strong language can end up as Extreme Negative.
A pattern across a few thousand posts is more reliable than the label on any single post. If your argument depends on one post, open it and read it yourself.
Categories & languages
Each post is assigned one category. The categories are deliberately broad: they're meant for narrowing a search, not as a detailed classification. Chatter is everyday conversation with no particular topic, and Other covers anything that doesn't fit elsewhere.
Languages
The language is detected automatically. This works well for a full paragraph and less well for a few words and an emoji. Mistakes are most common with very short posts, posts that switch language mid-sentence and posts made up mostly of hashtags.
How search works
Enter a topic and narrow the results with any combination of filters: date
range, language, category, network and polarity. At least one filter is
required, but it doesn't have to be the topic. Searches match whole words
only, so climate won't find "climatology".
The results show how posts are split across polarity levels, how average polarity changed day by day, and the posts themselves. Clicking a slice of the pie chart shows only the posts with that polarity; click the × on the filter label to see all of them again. Each post links to the original.
Reading a result
Above each post's text you'll see its polarity, source network, publication date and language. Keep an eye on the network: the same wording can mean something different on different platforms, and results that come mostly from one network say little about social media as a whole. If your search covers several networks, the summary at the top shows how many posts came from each.
Result limits
Each search returns at most a few thousand posts so the service stays usable for everyone. If more posts match, a notice tells you so. In that case, narrow your search, because what's on screen is only part of the results.
Exporting
You can export results as CSV, JSON or PDF. Exports include the fields shown on screen plus the source network. CSV opens directly in Excel. JSON also includes the filters you used and the summary counts. PDF is meant for reading, not for processing, and because of a font limitation it leaves out posts in non-Latin scripts. Use CSV or JSON for those.
Before you republish anything, remember that the original platforms' terms still apply to this content. Being able to download a post here doesn't mean you're allowed to redistribute it. The annotations are a different matter: you can reuse them freely under the terms in Licence.
Limitations & responsible use
Before using this data in your work, be aware of the following limitations.
- It is not a representative sample and can't be turned into one. Posts were collected around specific research topics, through platform search tools that have their own ranking and rate limits. The proportions you see describe this collection only. They are not estimates for a platform's users, and even less for the general public.
- Changes in volume over time are partly caused by how the data was collected. Platforms changed their API terms several times over these years, sometimes drastically. If posts from one network drop sharply in a given month, check for a change in our access before assuming public interest fell.
- The labels are automatic and have the weaknesses described under the polarity scale.
- It's a snapshot. Posts deleted or made private after we collected them may still appear here, which is usually why a link doesn't open.
- It describes discourse, not people. Use it for aggregate analysis. It cannot support claims about any individual's views, character or intentions. For that reason there is no way to search by author or view someone's posting history. Using exported data to profile a named person, or to target, contact or expose anyone, is outside the purpose for which the data was collected.
If you publish work based on this collection, explain how the data was gathered and what its limitations are, as is standard for observational social media research.
Data protection
Everything here was public when we collected it, meaning anyone could read it without logging in, following the author or joining a group. We don't store anything private or protected.
Being public doesn't mean the data can be used freely. These posts were written by real, identifiable people, so under the GDPR they are personal data, however openly they were posted. This section explains what that means in practice and what you can ask us to do.
Controller and Data Protection Officer
The data controller is Universidade da Beira Interior (UBI), Convento de Sto. António, 6201-001 Covilhã, Portugal.
For any question or request about your data, including removal requests, please contact the researcher responsible for the project, Sebastião Augusto Rodrigues Figueiredo Pais, at sebastiao@ubi.pt.
As the GDPR requires, we also give the contact details of the University's Data Protection Officer, Pedro Ricardo Morais Inácio: dpo@ubi.pt or +351 275 319 022 (ext. 2022). See the University's data protection page for its full statement.
Purpose and legal basis
We collect this data from social media platforms, not directly from the people who wrote it. In that situation, Article 14 GDPR requires us to publish this information ourselves instead of waiting to be asked. We don't contact authors before their posts are added to the corpus. Collection runs continuously across seven networks, so finding and messaging the author of every post would be disproportionate to the size of the collection and would, in practice, slow down or stop the research. Article 14(5)(b) GDPR allows this exception when individual notice would take disproportionate effort, as long as appropriate safeguards are in place. Instead of direct notice, we publish this page as a permanent public notice, limit the data we store (see What we store) and offer a simple removal process (see Asking us to remove a post below).
We process this data on the basis of the performance of a task carried out in the public interest (Article 6(1)(e) GDPR). Scientific research is one of the legal duties of the University as a public higher education institution (Article 8(1)(c) of Portuguese Law 62/2007 of 10 September), and processing for research purposes is further regulated by Article 31 of Portuguese Law 58/2019 of 8 August.
Personal data in this corpus is used only for scientific research on sentiment, discourse and collective reactions in social media text, as described under What this is. We don't use it for advertising, profiling, automated decisions about individuals or any commercial purpose, and we don't sell or license it to third parties. We don't supply it to law enforcement or security services, and we don't run any alerting or operational monitoring service from it.
The posts in this collection, with their links to the original platform, can be viewed and exported by anyone who visits this site.
The corpus covers many topics, including politics, religion, health and current affairs. Posts on these topics may reveal political opinions, religious or philosophical beliefs or information about health, which Article 9 GDPR treats as special categories of personal data. We process this data under the research provisions of Article 9(2)(j) GDPR, together with Article 89 GDPR and Article 31 of Portuguese Law 58/2019, with the safeguards described below.
All processing follows the safeguards for scientific research in Article 89 GDPR and in Portuguese Law 58/2019 of 8 August, which implements the GDPR in Portugal. These safeguards include data minimisation and limiting the use of the data to research.
Retention
We keep the corpus for as long as it is needed for the research purposes described on this page, including allowing published results to be checked and reproduced. Because the data is processed only for scientific research, with the safeguards of Article 89(1) GDPR, it may be kept for longer than would otherwise be allowed, but not in a form that identifies the authors for longer than these research purposes require (Article 5(1)(e) GDPR).
Your rights
If something you wrote is in the corpus, you can ask us what data we hold about you, ask us to correct or delete it, ask us to stop using it, or object to the processing altogether. If you have any concern about how your data is handled, please contact Sebastião Augusto Rodrigues Figueiredo Pais first (see Contact), so that we can try to resolve it. You also have the right to lodge a complaint with the Portuguese supervisory authority, the Comissão Nacional de Proteção de Dados (CNPD) (Article 77 GDPR).
Article 89 GDPR and Article 31 of Portuguese Law 58/2019 allow some of these rights to be limited for research when exercising them would seriously harm the research. If we ever rely on this, we will tell you so and explain why, instead of simply refusing your request.
Asking us to remove a post
Email sebastiao@ubi.pt with the link to the post and mention ExtremeSentiLex so your message reaches the right team. You don't need to give a reason. We'll confirm that we received your request, remove the post from public view on this site, so that it no longer appears in searches, results or exports, and reply within the time limits set by the GDPR.
Where the law allows data to be kept for scientific research (Article 17(3)(d) GDPR), the post may remain in the research records under the safeguards described on this page, but it will no longer be shown or made available to anyone through this site.
About your visit to this site
This site uses no tracking cookies, advertising or analytics. Your searches aren't linked to you or stored as a profile. There are two exceptions. The web server keeps standard access logs, and the search service briefly counts recent requests from each address to prevent overload. Neither is used to identify you. The page also loads some fonts and scripts from external providers instead of hosting them itself, so your browser connects to those providers when the page loads, as it does on any site that uses shared font or script libraries. This information is provided under Article 13 GDPR, because this data is collected from you directly. We process it on the basis of our legitimate interest in keeping the service secure and available (Article 6(1)(f) GDPR and Recital 49). Access logs and request counts are kept only for as long as needed for that purpose and are then deleted. Where a provider processes data outside the European Economic Area, the transfer relies on an adequacy decision of the European Commission or on other appropriate safeguards under Chapter V GDPR (Articles 45 and 46). You can ask us for more information about these safeguards using the contact details above.
Further reading
- Data protection at Universidade da Beira Interior (UBI) - the University's statement and DPO contacts
- Legal notice and privacy policy of the University website
- Regulation (EU) 2016/679 - General Data Protection Regulation, full text on EUR-Lex
- Law 58/2019 of 8 August, which implements the GDPR in Portuguese law
- Law 62/2007 of 10 September, the legal framework for higher education institutions in Portugal
- Comissão Nacional de Proteção de Dados (CNPD), the Portuguese supervisory authority
How to cite
If you use this collection in your work, please cite:
- Pais, S., Cordeiro, J., Martins, R., Albardeiro, M. "SocialNetCrawler: Online Social Network Crawler." Proceedings of the 11th International Conference on Management of Digital EcoSystems (MEDES '19). ACM, 2019. DOI: 10.1145/3297662.3365805
- Pais, S., Tanoli, I., Albardeiro, M., Cordeiro, J. "Unsupervised Approach to Detect Extreme Sentiments on Social Networks." 2020 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 651–658. 2020. DOI: 10.1109/ASONAM49781.2020.9381420
Licence
The annotations added by this project (polarity, topic category and detected language) are made available under a CC BY 4.0 licence. You may copy, adapt and redistribute them, including for commercial purposes, as long as you cite the works above and say whether you changed anything.
This licence covers only the annotations. The text of the posts belongs to their authors and remains subject to the terms of the platform where it was published, so we cannot license it to you. If you need to share a dataset derived from this collection, the safest option is to share the post identifiers or links together with our annotations, and let others retrieve the text from the original platforms.
Contact
For questions about the collection, the method, collaborations or research access, write to sebastiao@ubi.pt. If you find an error in the data, please let us know.
For data protection, access and removal requests, contact the researcher responsible for the project, Sebastião Augusto Rodrigues Figueiredo Pais, at sebastiao@ubi.pt. See Data protection for more details.