Artificial Intelligence & Machine Learning
Lexplore’s technology for analyzing eye movements and assessing the ability of children’s reading levels is based on artificial intelligence (AI) and machine learning (ML). Artificial intelligence generally refers to computers and computer programs which have the ability to resolve complex tasks and to recognize connections that normally require intelligence or cognitive ability on a human level. Research into AI over the past 5-10 years has made major strides which have, for example, enabled improved automated systems for speech and language comprehension, image recognition and decision-making support within health and medical care (e.g. medical diagnosis). The number of new application areas and industries that use AI in their technology development is steadily growing.
Machine learning is often used synonymously with artificial intelligence, but it was originally a sub-area of and a methodology within AI. It aims to develop programs or systems that resolve complex tasks through experience and available data rather than based on explicit rules and instructions. The methods or algorithms (procedures) used within machine learning are thus primarily data-driven and based on statistical calculations. This means that for systems developed with ML, the more available data there is with which to train the algorithms, the better equipped the systems are to resolve tasks.
So how is Lexplore’s screening technology based on machine learning? In simple terms, we use large amounts of eye movement data to automatically train ML algorithms to recognize correlations between children’s eye movements and reading ability. By tracking how the eyes move while reading, we obtain immediate data that reflects the cognitive processes in the brain. Our ML models were developed based on data from 3,000 anonymous children in grades 1-3. When discussing screening tests, standardization is sometimes addressed (i.e. standardization of the collected results). You could say that our test has been standardized based on 3,000 children aged 7-10 years.
It is very important to note that our ML models are based on more than just eye movement recordings. They are also based on results from a battery of other common tests, such as rapid automatized naming, word strings and reading of non-words and real words. Based on these tests, we have subsequently been able to estimate the general reading ability (or more specifically word decoding ability) of students in relation to their grade level. The task which ML algorithms are faced with during the training process consists of learning to perform statistical mapping between the results of the tests and the children’s eye movements as they read short grade-level texts. When Lexplore measures an individual child’s reading ability, the assessment pertains not only to the child’s eye movements, but also indirectly evaluates how the child would perform on other tests that assess word-decoding ability — tests which consume much more time and resources to administer and which ultimately arrive at the same analysis.
It is also important to point out that there is no single parameter or trait in the eye movements that determines how our models assess a child’s reading ability. Rather, it is the combination of several parameters and how they reflect the cognitive reading process that determines the child’s results. Since our screening tool can be more quickly administered than the various tests upon which it is based (no correction is required, for example), time and resources can be more effectively used to better help the students who need the most support in their reading development.