AI Readability Assessment in Education

This article explores how Automatic Readability Assessment (ARA) leverages artificial intelligence to accurately grade educational texts. It examines the integration of syntactic parse depth to measure sentence structure complexity and lexical psycholinguistic features to evaluate cognitive word difficulty. Finally, it details how machine learning models combine these advanced linguistic metrics to outperform traditional readability formulas and improve educational content matching.

Automatic Readability Assessment (ARA) is the computational process of determining how difficult a text is to read and understand. Historically, educators relied on basic formulas like Flesch-Kincaid, which only count word length and sentence length. Modern ARA uses artificial intelligence to analyze the deeper cognitive load required to process a text, ensuring students receive materials suited to their exact reading comprehension levels.

To evaluate structural complexity, AI systems measure syntactic parse depth. Using Natural Language Processing (NLP), the AI generates a parse tree for every sentence, mapping out the grammatical relationships between words. The "depth" of this tree indicates how many layers of embedded clauses and phrases exist before reaching the core meaning. A sentence with a high parse depth forces the reader to hold multiple grammatical dependencies in their working memory at once. By calculating the average and maximum parse tree depths of a text, the AI directly quantifies the structural strain placed on a student's memory.

Beyond sentence structure, ARA evaluates word-level difficulty through lexical psycholinguistic features. Instead of merely counting syllables, AI cross-references the text against psycholinguistic databases to assess how the human brain processes specific words. Key features include:

• Age of Acquisition (AoA): The average age at which a person typically learns a specific word.
• Concreteness: How easily a word can be experienced through the senses (e.g., "apple" is concrete, while "justice" is abstract).
• Imageability: How quickly and easily a word evokes a mental picture.
• Word Frequency: How often the word appears in everyday language versus specialized academic texts.

Texts containing words with a high Age of Acquisition and low concreteness or imageability require significantly more cognitive effort to decode, even if the words themselves are short.

Artificial intelligence brings these syntactic and lexical elements together through machine learning algorithms, such as Random Forests, Support Vector Machines, and deep learning transformers. During training, the AI is fed thousands of educational texts that have already been graded by human experts. The system extracts the syntactic parse depths and lexical psycholinguistic scores as numerical features, learning the complex patterns that correlate with specific grade levels.

Unlike static formulas, AI models recognize how different features interact. For example, an AI can determine that a sentence with a deep syntactic parse tree is still readable for a middle schooler if the vocabulary has high concreteness and a low Age of Acquisition. Conversely, it will flag a structurally simple sentence as advanced if it is packed with abstract, low-frequency terminology. By synthesizing these linguistic dimensions, AI-driven ARA provides educators and publishers with a precise, objective tool for tailoring learning materials to student capabilities.