In this paper, prerequisite graphs from prior educational-resource work — LectureBank, MOOC-style concept-prerequisite graphs (Pan et al. 2017), and subsequent prerequisite-learning methods (Liang et al. 2015; Roy et al. 2019; Jia et al. 2021) — are used as **fixed retrieval substrates**. The paper does **not** infer new prerequisite edges, learn a prerequisite-relation classifier, or otherwise extend the graph. Instead, the graphs are taken as given inputs to graph-aware retrieval, and the paper's contribution is restricted to how a retriever **traverses** them. This scoping decision separates the paper's claims from the prerequisite-learning line it cites and is what makes the strict-parity headline comparisons attributable to graph **traversal** rather than graph **construction**.

Google

**LectureBank** is a publicly released corpus of English lecture slide files collected from university courses in NLP, ML, AI, deep learning, and information retrieval. The original release contains $$1{,}352$$ lecture files and $$208$$ manually annotated prerequisite-relation topic pairs. Each topic is classified against a fixed taxonomy, and each ordered pair of topics is labeled as a prerequisite link or not. The dataset is the canonical resource for prerequisite-chain learning in NLP education.

LectureBank Dataset

**Prerequisite Relation Learning for Concepts in MOOCs** is the canonical method paper by Pan, Li, Li, and Tang (ACL 2017) that formulates and supervises concept-level prerequisite-relation prediction over Massive Open Online Course materials. Course concepts are first extracted from Coursera lecture captions (in Data Structures & Algorithms and Machine Learning), and each ordered pair of concepts is represented by **contextual, structural, and semantic features**, including two link-style cues introduced by the paper: a **video reference distance** and a **sentence reference distance** that exploit the order in which concepts are first introduced across videos and sentences. A binary classifier (with logistic regression and SVM variants) is trained on manually labeled concept pairs to predict whether concept $$A$$ is a prerequisite of concept $$B$$. The work establishes the supervised feature-based baseline for MOOC concept-prerequisite learning and releases the **MOOC-CS** dataset that the same paper introduces and that later work treats as the standard MOOC prerequisite benchmark.

Pan et al. (2017) Prerequisite Relation Learning for Concepts in MOOCs

Chen Liang and collaborators contributed two foundational works to the concept-prerequisite-learning line cited in the paper. The earlier work, **Measuring Prerequisite Relations Among Concepts** (Liang, Wu, Huang, Giles, EMNLP 2015), introduces a single link-based metric called **Reference Distance (RefD)** that measures how asymmetrically two concepts cite each other in a reference network, and shows that this unsupervised metric outperforms then-current supervised classifiers across seven domains on two evaluation datasets. The follow-up work, **Recovering Concept Prerequisite Relations from University Course Dependencies** (Liang, Ye, Wu, Pursel, Giles, AAAI 2017), formalizes the problem of recovering **concept-level** prerequisite relations from observed **course-level** prerequisite dependencies, proposes an optimization-based framework that jointly models concept relations and course-concept memberships, and releases the first real dataset built from university computer-science course listings. Together these papers establish a core supply of prerequisite-relation methods and benchmarks that subsequent prerequisite-learning work (e.g., LectureBank, MOOC-CS, hetGNN-style approaches) builds on or compares against, and they are part of the substrate the paper reuses rather than re-derives.

Liang et al. (2015) Recovering Concept Prerequisite Relations

**PREREQ** is a supervised method for inferring prerequisite relations between educational concepts, introduced by Roy, Madhyastha, Lawrence, and Rajan at IAAI-19 (AAAI 2019). Concepts are represented by latent topic vectors learned with **Pairwise Latent Dirichlet Allocation (Pairwise-LDA)** over concept-aligned online educational text (e.g., course materials and Wikipedia), and a **Siamese neural network** then scores an ordered pair of concept representations to predict whether the first concept is a prerequisite of the second. The method targets the asymmetric, directed nature of prerequisite links and is evaluated on online-course and university-course concept datasets, where it outperforms prior unsupervised and supervised baselines. The work is one of the standard references in the subsequent prerequisite-learning literature alongside Liang et al. (2015) and the LectureBank line.

Roy et al. (2019) PREREQ Concept Prerequisite Inference

**CPRL** (Concept Prerequisite Relation Learning) is a method introduced by Jia, Shen, Tang, Sun, and Lu at NAACL 2021 for predicting prerequisite relations between educational concepts. It builds a **heterogeneous graph** over concepts together with auxiliary node types such as learning objects, applies a heterogeneous graph neural network to obtain concept representations, and feeds those representations together with **concept-pairwise features** to a classifier that predicts whether one concept is a prerequisite of another. The authors further extend CPRL to **weakly supervised** settings, leveraging learning-object dependencies and **data programming** to generate training labels when explicit concept-pair annotations are scarce. CPRL reports state-of-the-art results across four prerequisite-relation benchmarks and is a standard recent reference in the prerequisite-learning literature alongside Liang et al. (2015) and Roy et al. (2019).

Jia et al. (2021) Heterogeneous GNN for Concept Prerequisite Relation Learning

This organizer groups the paper content from Related Work > Prerequisite graphs and educational structure..

Prerequisite Graphs And Educational Structure (Related Work) in Auditable Strict-Parity Evaluation of Prerequisite-Graph Retrieval for RAG under Leakage Controls

Root URL: https://1iman.com/
Canonical paper URL: https://1iman.com/assets/working-papers/pdf/emnlp-auditable-strict-parity-evaluation-of-prerequisite-graph-retrieval-for-rag-under-leakage-controls.pdf
Portfolio PDF URL: https://1iman.com/assets/working-papers/pdf/emnlp-auditable-strict-parity-evaluation-of-prerequisite-graph-retrieval-for-rag-under-leakage-controls.pdf
Paper title: Auditable Strict-Parity Evaluation of Prerequisite-Graph Retrieval for RAG under Leakage Controls
Use this Reference node when citing paper-derived 1Cademy proposals from the Researcher agent.

Reference: Auditable Strict-Parity Evaluation of Prerequisite-Graph Retrieval for RAG under Leakage Controls

**LectureBank-Full** is the expanded configuration of the LectureBank corpus used as a prerequisite-QA / retrieval benchmark. It contains $$1{,}717$$ English NLP-related lecture slide files and manual concept-pair prerequisite annotations over $$322$$ topics. The task is to retrieve, for a given target concept (query), the set of gold prerequisite concepts from a fixed candidate pool. Systems are scored with ranking metrics such as Recall@$$k$$ (R@$$k$$); R@$$10$$ measures the fraction of gold prerequisites recovered in the top-$$10$$ retrieved candidates. Standard splits hold out concept pairs for evaluation, and headline comparisons fix the encoder, candidate pool, cutoff, matching rule, and split policy.

LectureBank-Full Prerequisite QA Benchmark

Prerequisite Graphs Consumed as Retrieval Substrates, Not Inferred

In this paper, **MOOC-CS** is consumed as a prerequisite graph with **$$225$$ concepts**, **$$516$$ prerequisite edges**, and **$$1{,}016$$ QA pairs** derived from the **MOOCCubeX computer-science curriculum**, with prerequisite-relation lineage from Pan et al. (2017). Graph descriptors computed from the released graph and concept texts are: mean depth $$7.2$$, fraction of nodes with depth $$\geq 1$$ equal to $$77.3\%$$, mean degree $$4.6$$, and mean concept-text length $$6.1$$ tokens. Compared to LectureBank-Full, MOOC-CS is deeper but sparser in degree and uses much shorter concept texts.

MOOC-CS Configuration Used in Hierarchical Prerequisite RAG (225 Concepts, 516 Edges, 1,016 QA)

The paper states a **pragmatic retrieval objective**: improve **recall of gold prerequisite chains** while keeping retrieved contexts **interpretable and reproducible**. Recall of gold chains is the optimization target (so missing a true prerequisite is the failure mode that matters), and the **interpretability + reproducibility** constraint motivates deterministic, auditable graph policies and fixed-interface comparisons rather than opaque end-to-end retrievers. This objective is what justifies the paper's later choice of deterministic diffusion and strict-parity headline reporting.

Learn Before

Related