03/31/2022

Martina Miliani

An algorithm for analyzing vaccine tweets

Artificial intelligence also supports the monitoring of the side effects of vaccines against Covid-19. Beatrice Portelli, Simone Scaboro, Edoardo Lenzi, Roberto Tonino, a group of students from the University of Udine, assisted by three international researchers, Giuseppe Serra, Emmanuele Chersoni and Enrico Santus, created a platform capable of analyzing thousands of Twitter post in relation to what users post online about the three most popular vaccines against Covid-19: Pfrizer-Biontech, Astrazeneca / Vaxzevria and Moderna.

The data analyzed

The algorithm underlying the functioning of the portal allows you to view the most used hashtags, the most shared links and the most cited side effects, also in relation to the geographical area of ​​origin of the posts and the sentiment that is conveyed, or the negative polarization or positive of the tweet.

Pharmaceutical companies can thus monitor day by day how public opinion on vaccines changes, and can also collect data on more or less common side effects and possibly orient drug testing in
a certain direction.

The platform also allows you to track which are the most shared links, and therefore have an idea of ​​the main sources of user information, and also observe the spread of any fake news.

The one proposed by the University of Udine is a tool that can therefore also guide public institutions, and especially health institutions, to implement effective communication strategies in order to increase, for example, vaccination coverage.

The idea

The project was born following the participation in a shared-task, a contest in which some data are shared with the scientific community of reference and the model that best manages to
exploit that data to perform a certain task.

The scientific community of reference in this case is related to the field of computational linguistics, a discipline that applies computer techniques to language, including those relating to Natural Language Processing (NLP), also providing for the use of artificial intelligence algorithms.

The aim of the international contest, named SMM4H 2019 (Social Media Mining for Health Applications), was the recognition in the text of the side effects of some drugs, for which AILAB – this is the name of the laboratory of the University of Udine coordinated by professors Giuseppe Serra and Carlo Tasso – achieved the best score among the competing models.

The algorithm

The model proposed by AILAB of the University of Udine employs a neural network, an architecture composed of processing units arranged in such a way as to trace the network structure of neurons.

This network is based on BERT, a computational model of the language capable of capturing also the polysemy of terms within the texts, and which reaches the state of the art in many NLP tasks.

Models such as BERT exploit the so-called attention mechanism, through which the network “focuses” on all the terms of the portion of the text under examination, activating certain units, some “neurons”, which retain information throughout the elaboration process.

This architecture has been combined by researchers with a statistical model called Conditional Random Fields (CRF) which annotates text sequences: the data produced by BERT are then used by the CRF to make decisions, i.e. to establish whether a certain sequence of terms can be noted as a side effect or not.

Upcoming developments

At the moment the system only extracts information from Twitter, but other social networks should soon be included in the monitoring.

“Among the priorities that we have identified – points out Giuseppe Serra – is that of evolving the platform so that it can also serve as a fact-checker: we would like to develop a model that automatically signals the possible untruthfulness of a news”.

But that’s not all: The model used to analyze vaccine posts can also be adapted to monitor the side effects of other drugs.

Beyond the platform

The interests of Emmanuele Chersoni, Assistant Professor at the Hong Kong Polytechnic University also go beyond the possibilities of the platform.

“I’d be curious to see – he explains – how human beings process a certain sentence when they have to perform the same operation that our model does, that is to identify the side effects in a tweet”.

In fact, Chersoni deals with cognitive linguistics, a branch of linguistics that studies the cognitive processes underlying the use of language.

“I would like to use tools like the eye-tracker to track eye movement. Who knows what kind of correspondence exists between the way in which a human being reads to annotate a text and the processing processes implemented by the attention mechanism of our neural network ”.

#social #networks #medical #research

Leave a Reply

Your email address will not be published.