21 Jun 2021 |
Research article |
Information and Communications Technologies
Tackling Fake News with Artificial Intelligence
Purchased on Istockphoto.com. Copyright.
In the field of artificial intelligence, a classifier is a mathematical algorithm based on a model that assigns a class to an image, audio or text. The number of possible classification types is infinite and depends on the application. For example, a classifier could analyze images and indicate the probability that the represented object is a cat. Each classifier is a kind of specialist that performs well under some conditions and not so well under others.
Combining several specialist classifiers, instead of using one very complex generalist classifier, increases the performance of artificial intelligence systems while decreasing computing time. This way, each classifier is assigned a simpler model and can be implemented on a different computer. The overall performance of a classifier set will depend on two factors: the choice of classifiers and the way queries are distributed among them.
At ÉTS, we are conducting several research projects to address both aspects. We are working on approaches to generate sets of complementary classifiers, so that there is always a competent specialist for each case that arises in an application. However, we are primarily interested in developing methods that can select the best specialist among classifiers in a given set.
Dynamic Ensemble Selection
When the same query is submitted to different classifiers, the answer is not necessarily the same: they may have diverging opinions. Instead of asking all the classifiers to give their opinion to then combine all the answers, we can place an upstream dispatcher in charge of routing the queries only to the most qualified specialists, as the case may be. The dynamic ensemble selection approach uses past performances to identify these specialists. It analyzes case details and compares them to other already processed cases to find similarities. Then, using statistical models, it routes the new case to the classifiers most likely to give the correct answer.
Dynamic ensemble selection is a general method that is beginning to show promising results in image, sound, and text processing. Below are some of the research projects underway at ÉTS, addressing dynamic ensemble selection while seeking to make it more efficient and reduce its computing requirements.
Fake News and Hate Speech: Flagging Inappropriate Texts
There are several ways to represent a text: semantics, grammar, natural language… Each model can be integrated into a classifier that becomes a specialist of this representation. Dynamic ensemble selection can either choose the model most likely to respond well or weight each of the classifiers to arrive at an overall score. Systems like this can be used to verify whether the author to whom the text is attributed actually wrote it, and where the text was published.
The same method can be used to flag hate speech. By analyzing a text from several perspectives, each represented by a specialist classifier, it is possible to highlight certain characteristics associated with this type of text, such as poor grammar.
A TV channel that targets Arabic-speaking listeners who speak dialects somewhat similar to classical Arabic wished to offer a subtitling machine translation service that adjusts to the dialect used in each program. It therefore launched a competition to obtain interesting solutions for this application.
Designing a model that would take into account all the variations of Arabic dialects would be extremely difficult and expensive and would risk resulting in inaccurate transcriptions. It is easier to develop as many models as there are dialects in order to obtain a set of specialized classifiers. All that remains is to route the audio files to the corresponding classifier using dynamic ensemble selection.
In video surveillance, several cameras of various types (standard and infrared) offering a view from different angles are often combined to provide the most complete view of a location. Each camera can be associated with a classifier and a dynamic ensemble selection system can decide which camera offers the best view or is the most efficient depending on lighting conditions.
The Future of Artificial Intelligence
Results obtained so far are very encouraging. Dynamic ensemble selection can be applied in many seemingly complex contexts and allows a quick adaptation to the inevitable changes that eventually occur in the environments to be analyzed. In the long run, it promises very important gains for many applications in artificial intelligence.
Rafael Menelau Oliveira e Cruz
Rafael Menelau Oliveira e Cruz is a professor in the Department of Software Engineering and IT specializing in artificial intelligence.
Research laboratories : LIVIA – Imaging, Vision and Artificial Intelligence Laboratory