Messages on Twitter about politics in Spain are usually negative. But an analysis of the conversation through 1.3 million tweets since the crisis broke out in the Popular Party until Monday shows that although no one is saved, Isabel Díaz Ayuso receives a slightly more positive sentiment. Until Friday at noon, however, 90% of the tweets were negative, both those that spoke of Ayuso and Pablo Casado. Since the publication of Ayuso’s statement on Friday, the president has slightly but significantly improved her position.
Artificial intelligence models that analyze language and are able to understand or generate it continue to improve in recent years. The analysis of the PP crisis tweets has been done for EL PAÍS by a group of researchers from Cardiff University (Wales). José Camacho Collados and Luis Espinosa Anke, researchers at Cardiff, have created a model that accurately analyzes whether the sentiment of a tweet is positive, negative or neutral. your model, called TweetEval, has become an unprecedented success in recent months: from having a few thousand downloads in January 2021 it has skyrocketed to 15 million in January 2022 alone and competes in the league of models generated by giants such as Google, Meta or OpenAI . Classification can be followed in the company which has become the center of these open source models, HuggingFace. Along with Camacho and Espinosa, two employees of the Snapchat research laboratory, Francesco Barbieri and Leonardo Neves, are the authors of the model.
“We don’t know the reason for this explosion,” says Camacho. “Perhaps it’s people from companies who want to analyze how their brand is doing or politicians on Twitter or pass it through comments on Facebook or Instagram,” he adds. Although the model is optimized for tweets and not comments, it can still work. Basically, it is a new way of measuring opinion states, like a survey, although it has its complexities. “It’s a cheap way to scan Twitter about your brand,” explains Espinosa. With hardly any resources, and with relatively small samples, political parties, sports clubs, film producers, can get an idea of how Twitter breathes in the face of any news or trend.
In Hugging Face they also do not know the reason for the sudden success, beyond giving some context to those millions: “The number of downloads can vary a lot. It does not mean that Cardiff NLP [otro nombre para TweetEval] be more used than other models necessarily”, says Omar Sanseviero, Hugging Face engineer. “It could mean that a few companies are using it very heavily. We can’t really know. Anyway, it’s great and impressive to see such a large use of models that don’t come from big research labs,” he adds. In other words, each download of the Cardiff model does not imply a strictly new use, but rather the same organization can download the model several times to look at similar cases. And each of those uses counts as a download.
The simplicity of using models like this has its dangers: you have to understand what is being done and why. In specific cases, the sample may be biased, perhaps there is artificial conversation caused by specific campaigns or even tweets from other topics may sneak in if the keywords are confusing. Despite the caveats, as in the case of the PP, following a reliable trend is relatively easy if large extrapolations are not made.
The author of the analysis on the PP, Dimosthenis Antypas, saw, for example, until Friday a clear picture: “The results indicated that public sentiment, at least on Twitter, was going in one direction, with more than 90% of the tweets mentioning to Ayuso or Casado classified as negative. There was no significant difference between the two politicians,” he explains. But then something changed and something jumped on Twitter the same Friday that would be reflected in the media, the demonstration on Sunday and the sensations at the beginning of this week: Ayuso, for some reason, improved. “Without knowing what happened at all, I assume that something happened after Friday lunchtime that made public opinion about Ayuso improve. He shot himself around Friday night, ”explains Antypas, who does not closely follow Madrid politics.
The change is not extraordinary, but it is significant. The positive comments about Pablo Casado do not vary; Ayuso’s yes. “Isabel, lots of encouragement and all my support!! You are very big and I invite you to sign for Vox”, “president, always with you. Cheer up and thank you.” The enthusiasm that his figure arouses is another reason that can raise his feelings: he has more fans than Casado.
Politics in Spain (and in other countries as well) is usually negative. In a previous scientific article by these researchers, found that the most popular tweets from MPs in Spain, the UK and Greece tend to receive far more retweets than positive ones. They also detected that politicians who are in the Government write more positive tweets than the opposition: in Spain 84% of Pedro Sánchez’s tweets are positive, those of Casados are 63% negative. It is logical, but the confirmation does not cease to surprise.
Outside of politics, anger also dominates, but much less: “If you take a random tweet in Spain, the distribution is 46% negative, 16% neutral and 39% positive; it is more or less balanced,” says Camacho.
An industry in explosion
This field of natural language processing is exploding. The scientific article that marks almost all research today is one from Google from late 2018: “The basic structure has changed and it has already established the same model that learns from very large collections of texts,” he says, which means enormous changes in a few years: “Our world changes very quickly, now 90% of academic articles use this , which is something that did not exist in 2018”, he adds.
The speed at which the ability of machines to understand language and produce it is changing is enormous. But much remains to be done. Now a model that works well to understand Twitter, is not capable of writing a newspaper article, much less a court ruling. Each field is nourished by its own databases. But that will also change over the years. “There is a limit to what you can do with pattern recognition and with millions and millions of pieces of data. They only imitate what they have seen but do not reason. It is a very different way of learning language. The biggest challenge is for these models to understand the language. For example, if you say that your grandson lives in Granada and the model knows that the Alhambra is there. Now that doesn’t happen”, says Camacho.
The success of their model has led these researchers to want to popularize it. In a few weeks they will post a page where consultations can be made with small samples on cases such as Ayuso and Casado. There are already a lot of scientific articles made based on his model: “The feeling in general has a leg in many other tasks with value”, says Espinosa. ”The sentiment information can be used to improve, for example, a suicide prevention model. We have also seen misinformation about covid or polarization,” she adds. There are those who have used them, for example, to see if successful songs are becoming more and more negative. Turns out yes.
You can follow THE COUNTRY TECHNOLOGY on Facebook and Twitter or sign up here to receive our weekly newsletter.
#analysis #million #tweets #Casado #Ayuso #Madrid #president #slight #advantage