AI Reading System Learns From Every Click, Pause and Search

A new artificial-intelligence recommendation system could make English reading platforms feel less like digital libraries and more like tutors that notice how people learn. The research combines a mixture-of-experts model, which assigns different parts of the recommendation task to specialized sub-models, with a neural preference-embedding system, or NoPE, that reconstructs a learner’s interests from digital traces such as pauses, word searches, skipped sections and reading time. In the reported experiments, the integrated system reached 81.23% recall for highly relevant content, compared with 56.7% for traditional collaborative filtering, and produced recommendations in an average of 0.9 seconds. These results suggest that a combination of specialized reasoning and richer behavioral signals may help digital reading services overcome a persistent problem: learners rarely provide the explicit ratings that conventional recommendation engines expect.

The motivation comes from a mismatch between how people read and how recommender systems are often built. Collaborative filtering treats explicit ratings and dense user-item interactions as the foundation of personalization, while content-based systems compare articles through keywords, topics or tags. English learners, however, tend to browse much more than they rate. Their useful preferences may be hidden in the duration of a reading session, the frequency of word lookups, the points at which they jump away, or the number of times they bookmark a passage. The new framework treats these weak signals not as isolated counts but as sequences whose timing, order and transitions can reveal changing goals, such as a move from narrative material toward technical or academic texts.

NoPE was designed specifically for that sparse-feedback environment. Its architecture combines a feature input layer, a preference-embedding layer, a sequence-mining layer and an output layer. Behavioral trajectories are converted into dense vectors, with multi-scale pooling used to retain information about recent, average and especially salient interactions. Bidirectional attention then measures relationships between those preference representations and article semantics. A time-aware component further adjusts the sequence representation, allowing the system to respond when interests move rather than treating every user as permanently attached to a fixed set of subjects. In the study’s dynamic example, a learner who shifted from story-based to technology-focused reading had the preference vector updated within 15 days, with recommendation accuracy moving from 68.50% at the novice level to 73.20% at an intermediate stage.

The second half of the system tackles the fact that choosing an appropriate article involves several different questions. Is its vocabulary too difficult for the learner? Does its discourse structure match the material being studied? Is the subject relevant to the user’s interests? And does the article occupy an appropriate place in a developing learning path? Instead of forcing one general model to balance all these factors, the mixture-of-experts architecture divides the work among four specialists: a vocabulary-difficulty expert, a discourse-structure expert, a theme-interest expert and a learning-path expert. A gating network then assigns each expert a weight for the current user and article, combining their outputs into a final recommendation.

That routing system is more than a simple selector. Conventional mixture-of-experts models can suffer from unclear task allocation, redundant specialists and “expert failure,” in which some components are rarely activated and therefore receive little useful training. The researchers addressed this by adding separate attention mechanisms for task features and user features, allowing the gate to consider both the characteristics of an article and the learner’s present state. The resulting allocation matrix is calculated from user, content, explicit-preference, sequence-preference and global-context representations, then normalized to weight the expert outputs. This hierarchical fusion makes the reasoning path more responsive: a user showing frequent word searches may trigger stronger weighting for vocabulary support, while a demonstrated interest in academic writing may raise the contribution of semantic and discourse specialists.

The model also adjusts the final ranking with signals that matter in an educational setting. A timeliness term can favor newer material, while a cost term represents the difficulty burden of a recommendation relative to the learner’s current ability. The study uses a pairwise ranking loss to place a positive article ahead of a negative one, alongside regularization to control parameter growth and a consistency penalty for the reconstructed preference representation. At the data level, the system combines reading behavior, language-proficiency assessments, article semantics and contextual information such as device and usage setting. Feature-ablation experiments indicated that implicit behaviors contributed the largest share of the observed performance gain, at 42%, followed by content semantics at 31%, proficiency information at 18% and scenario context at 9%.

To find relevant articles efficiently, the system separates candidate recall from fine-ranking. Approximate nearest-neighbor search and hierarchical pruning retrieve semantically similar items from high-dimensional vector indexes, after which the expert network and preference model produce a more precise ordering. The experiments used 18,620 valid users and 52,370 English reading items drawn from school, university, workplace and overseas-examination groups. Interaction data were sparse, with fewer than 12 records per user on average. The first 70% of historical behavior was used for training, the next 15% for validation and the most recent 15% as an independent test set, an arrangement intended to resemble the passage of time in real online use. The integrated model was compared with traditional, content-based and popularity-based methods as well as sequence-based, semantic-matching, Transformer and attention-enhanced educational recommenders under the same chronological split and experimental conditions.

Performance varied substantially with learner experience. Expert-level readers achieved 89.70% recommendation accuracy, 94.07% user satisfaction and 95.07% precision, while novice users reported 48.00% satisfaction. Difficulty matching improved with proficiency: the reported rates were 73.94% for novice users, 81.64% for intermediate users and 95.07% for expert users. The authors also reported an overall content-adaptation rate of 99.1%, an average of 78.91 adaptive contents per user and a 67.8% rate of feedback indicating that recommendations closely matched users’ needs. They reported that users read an average of 44.89 recommended articles per day and that 72.1% experienced a significant improvement in reading efficiency. Ablation comparisons indicated that both attention-enhanced routing and NoPE independently improved performance, with their combination producing an additional benefit rather than merely increasing model size.

The speed results are as notable as the accuracy figures. The gating network’s expert activation rate exceeded 95%, compared with a situation in which the authors say roughly 30% of experts in a traditional mixture-of-experts setup could remain inactive for long periods. Candidate retrieval operated at millisecond scale, while the complete recommendation process averaged 0.9 seconds. Multi-source feature fusion was reported to improve recommendation accuracy by 15% to 20% over a single-data model, with a user complaint rate of 3.45%. These findings position the system as a possible engine for real-time reading support in which a recommendation changes not only according to subject preference but also according to ability, discourse demands, recent behavior and the learner’s immediate goal.

There are important limits to the evidence. The experiments are based on one assembled dataset and the reported performance varies with the amount and type of behavioral history available. The strongest results occur for expert users with long, dense interaction sequences, whereas beginners remain less satisfied, suggesting that cold-start learning is not solved by the architecture alone. The paper also notes that its datasets are available from the corresponding author on reasonable request rather than through an openly documented public repository, limiting independent replication. No school or workplace trial, long-term learning outcome or comparison with expert teacher recommendations is described. The work therefore does not establish that faster, more accurate recommendations necessarily produce better language acquisition. It does, however, offer a technically coherent response to a central weakness of educational recommendation: the most revealing signals about what a learner wants are often scattered across ordinary reading behavior, and a model that can interpret those signals while routing different parts of the task to specialists may extract more value from the small traces that explicit ratings leave behind.

Subject of Research: A hybrid artificial-intelligence system for personalized English reading recommendations using expert routing and neural preference embeddings.

Article Title: Research on a personalized English reading recommendation system combining expert mixture model and NoPE

Article References: Wang, J. (2026). Research on a personalized English reading recommendation system combining expert mixture model and NoPE. Discover Artificial Intelligence, 6(1), Article 1200. https://doi.org/10.1007/s44163-026-02037-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02037-x

Keywords: artificial intelligence, English reading, recommendation systems, mixture of experts, neural embeddings, personalized learning, weak feedback, behavioral modeling, attention mechanisms, sparse data, Research, personalized