🤖 AI

T-TExTS (Teaching Text Expansion for Teacher Scaffolding): Enhancing Text Selection in High School Literature through Knowledge Graph-Based Recommendation

This paper introduces T-TExTS, a knowledge graph-based recommendation system that leverages pedagogical ontologies and graph embedding strategies to assist high school English teachers in selecting diverse, thematically aligned literature texts, demonstrating that hybrid models combining structural and pedagogical signals offer an optimal balance between interpretability and ranking performance.

Original authors: Nirmal Gelal, Chloe Snow, Ambyr Rios, Kathleen M. Jagodnik, Hande Küçük McGinty

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nirmal Gelal, Chloe Snow, Ambyr Rios, Kathleen M. Jagodnik, Hande Küçük McGinty

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Overwhelmed Librarian"

Imagine a high school English teacher as a librarian who needs to build a special bookshelf for their class. They don't just want any books; they need a specific set of stories that all talk about the same big idea (like "rebellion" or "dystopia") but come from different authors, cultures, and time periods.

However, the teacher is busy. They don't have hours to read every book in the world to find the perfect matches. Usually, they end up picking the same old classics they know well, which means the bookshelf lacks diversity. They need a smart assistant to help them find new, diverse books that still fit the lesson plan perfectly.

The Solution: T-TExTS (The "Smart Map")

The researchers built a tool called T-TExTS. Think of this not as a simple search engine that looks for keywords (like "dystopia"), but as a giant, interactive map of the literary world.

The Map (Knowledge Graph): Instead of just a list of books, the system builds a map where every book is a city. The roads connecting these cities aren't just "similar words"; they are deep, educational connections. One road might say, "These two books both deal with censorship," while another says, "These two have similar reading difficulty."
The Blueprint (Ontology): Before drawing the map, the researchers worked with expert teachers to create a strict blueprint (called an ontology). This blueprint ensures the map only uses "pedagogical" roads—connections that actually matter for teaching—rather than just surface-level similarities.

How the System "Walks" the Map

To find the best books, the system sends out little "explorers" (algorithms) to walk along the roads of this map. The paper tested four different ways these explorers could walk:

DeepWalk (The Random Tourist): This explorer walks randomly, taking every road with equal chance. It gets a broad, general feel of the neighborhood.
Biased Random Walk (The Guided Tour): This explorer is led by the expert teachers. If the teachers say, "Genre is the most important connection," this explorer is forced to take those specific roads more often. It's like a tour guide who only shows you the "best" spots based on a specific rule.
Node2Vec (The Strategic Hiker): This explorer is smart. It can choose to stay close to home (looking at immediate neighbors) or hike far away to discover new, distant territories. It balances between exploring the local area and finding hidden gems far away.
Hybrid (The Best of Both Worlds): This explorer carries a backpack containing the maps from both the "Random Tourist" and the "Guided Tour." It combines the broad view with the specific teacher rules.

What They Discovered (The Race Results)

The researchers tested these explorers on maps of different sizes (98 books, 196 books, and 351 books) to see which one found the best matches.

The Surprise: The "Guided Tour" (Biased Random Walk), which relied heavily on the teachers' specific weightings, actually performed worse at finding the overall structure of the map than the "Strategic Hiker" (Node2Vec).
- The Analogy: It turns out the map was drawn so well by the experts in the first place that the "Guided Tour" didn't need to force the explorer down specific paths. The map itself already held the right clues. Forcing the explorer to stick to one type of road actually made them miss some great connections.
The Winner: Node2Vec (the Strategic Hiker) was the best at finding the right books, especially as the library got bigger. It knew when to stay close and when to hike far to find the perfect thematic match.
The Practical Choice: The Hybrid model was the runner-up. It wasn't quite as fast as Node2Vec, but it was almost as good. Crucially, because it included the "Guided Tour" part, teachers could look at the results and say, "Ah, I see why it picked this book; it's because of the genre rule I set." This transparency is vital for teachers who need to trust the tool.

A Real-World Test: The "1984" Challenge

To prove it worked, they gave the system the book 1984 by George Orwell and asked, "What other books should we teach with this?"

The system correctly identified Fahrenheit 451 as the #1 match (because they are both famous dystopian stories about censorship).
The Node2Vec explorer was the only one to find all five of the expert-approved "perfect matches" (including The Hunger Games and Animal Farm).
The Hybrid model found four out of five, and it also found some "near misses" (like The Giver) that weren't on the official list but were still excellent teaching choices.

The Bottom Line

The paper concludes that you don't need to constantly tweak the rules of the road to get good recommendations. If you build a really high-quality map (the Knowledge Graph) based on expert teaching principles, a smart explorer (Node2Vec) can navigate it better than a rigidly guided one.

However, for a tool used in a classroom, the Hybrid approach is the sweet spot. It gives you the high performance of the smart explorer but keeps the "teacher's voice" visible in the results, ensuring the recommendations are both accurate and explainable.

What the paper does NOT claim:

It does not claim this tool has been tested in real classrooms with real students yet (that is planned for the future).
It does not claim to replace teachers; it claims to help them save time and find diverse books.
It does not claim to work for every subject; it is specifically designed for High School English Literature.

Technical Summary: T-TExTS (Teaching Text Expansion for Teacher Scaffolding)

Problem Statement
High school English Literature teachers face significant barriers in assembling diverse, thematically aligned text sets due to limited planning time and a lack of pedagogical resources. Traditional recommendation systems, such as Collaborative Filtering (CF) and Content-Based Filtering (CBF), are ill-suited for this domain. CF suffers from the cold-start problem with sparse interaction data, while CBF struggles to introduce novel topics beyond a user's existing preferences. Furthermore, existing Knowledge Graph (KG) applications in education have largely focused on linear concept sequencing in STEM or MOOCs, failing to address the non-linear, semantic connections (e.g., shared themes, social justice attributes, genre complexity) required for literacy education. There is a distinct gap in tools that support the pedagogical scaffolding of text sets for secondary English Language Arts (ELA).

Methodology
The authors propose T-TExTS, a KG-based recommendation system that prioritizes pedagogical merit over surface-level metadata. The methodology consists of three primary components:

Domain-Specific Ontology Construction:
Using the Knowledge Acquisition and Representation Methodology (KNARM), the authors constructed a domain-specific ontology grounded in pedagogical criteria. This process involved:
- Data Collection: Aggregating texts from Kansas state studies, educator consultations, and higher-education syllabi (ENGL 545 and ENGL 580).
- Text Analysis: Applying quantitative tools (Lexile, Flesch-Kincaid, ATOS) and qualitative rubrics (Achieve the Core) to evaluate text complexity and instructional merit.
- Ontology Design: Separating the schema into a Terminological Box (TBox) defining classes (Text, Author, Genre, Theme, TextComplexity) and an Assertional Box (ABox) containing specific assertions. The ontology explicitly models relationships such as hasGenre, hasTheme, and hasTextComplexity.
Knowledge Graph Generation:
The ontology was instantiated as a Knowledge Graph using RDFLib and stored in a GraphDB repository. The graph represents texts and authors as individuals connected by pedagogical and structural relations.
Embedding and Recommendation Strategies:
The system projects the symbolic relationships of the KG into a continuous vector space using random-walk-based graph embedding techniques. Four strategies were evaluated:
- DeepWalk: Uniform random walks.
- Biased Random Walk: Expert-guided walks where transition probabilities are weighted by domain experts to prioritize specific pedagogical relations (e.g., genre, theme).
- Node2Vec: Parameterized random walks controlled by return ( $p$ ) and in-out ( $q$ ) parameters to balance Breadth-First Search (BFS) and Depth-First Search (DFS) exploration.
- Hybrid Model: A concatenation of DeepWalk (uniform) and Biased Random Walk (expert-weighted) embeddings to combine structural coverage with pedagogical focus.
Recommendations are generated by computing Cosine Similarity between the embedding of an anchor text and all other texts in the corpus.

Experimental Setup
The system was evaluated across three dataset configurations containing 98, 196, and 351 texts. The graphs were split into training (80%), validation (10%), and test (10%) sets. Hyperparameters were tuned using Optuna with link prediction (AUC) as the optimization objective. Performance was measured using:

Link Prediction: Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC).
Ranking Metrics: Hits@K, Mean Reciprocal Rank (MRR), and normalized Discounted Cumulative Gain (nDCG@10).
Sensitivity Analysis: Comparison of default weighting versus a "genre-emphasized" weighting scheme.

Key Results

Structural Tuning vs. Expert Weighting: Traversal-level expert weighting (Biased Random Walk) alone did not outperform algorithmic structural tuning. The Biased Random Walk achieved the lowest AUC (0.6888–0.7466) across all scales.
Node2Vec Dominance: Node2Vec achieved the highest AUC at every dataset size (0.9642–0.9750) and the strongest ranking metrics (Hits@K, MRR, nDCG) on larger scales (196 and 351 texts). The optimal hyperparameter configuration consistently favored a high in-out parameter ( $q=4.0$ ), indicating a preference for Depth-First Search (DFS) exploration to capture longer chains of thematic association.
Hybrid Model Performance: The Hybrid model (concatenation of DeepWalk and Biased Walk) maintained a high AUC (0.9122–0.9350) across all scales and remained within a few percentage points of Node2Vec on ranking metrics. Crucially, the Hybrid model demonstrated robustness to weight perturbations; while the Biased Walk's performance shifted by ~3% when genre weights were adjusted, the Hybrid model's shift was dampened to ~1.5%.
Qualitative Case Study: For the anchor text 1984, Node2Vec was the only configuration to retrieve all five expert-curated ground-truth recommendations (including structurally distant but thematically aligned texts like Marrow Thieves). The Hybrid model successfully recovered ground-truth texts that the Biased Walk missed (e.g., Animal Farm), demonstrating the value of combining structural breadth with pedagogical bias.

Significance and Claims
The paper claims that T-TExTS provides a novel, pedagogy-first approach to educational recommendation, distinct from user-similarity or popularity-based systems. The central empirical finding is that a high-fidelity, expert-curated ontology inherently encodes sufficient pedagogical signal such that algorithmic structural tuning (specifically Node2Vec) can recover it without the need for additional traversal-level weighting.

However, the authors argue that the Hybrid model represents the most practical deployment choice for teacher-facing tools. It offers a balance of competitive ranking quality and interpretability, allowing educators to trace which expert-assigned weights influenced the recommendations. The study demonstrates that ontology-driven KG embeddings are viable for educational scaffolding even at modest data scales (up to 351 texts) and that the relative performance of embedding methods remains stable as the curriculum expands.

Limitations
The authors acknowledge that the dataset, while expert-curated, is modest in scale compared to general benchmarks. The evaluation relies on expert-curated ground truth rather than live classroom user studies, meaning the system's impact on planning time and actual text-set diversity in real classrooms remains unmeasured. Additionally, standard ranking metrics may penalize "pedagogically defensible near-misses" that do not match the specific ground-truth list but are still valuable for curriculum design. The authors suggest future work should focus on teacher-facing validation and the development of new evaluation metrics better suited for neuro-symbolic, expert-driven recommendation systems.