1 of 14

CFL-KG

A Knowledge Graph for Instructional Support�in Chinese as a Foreign Language

Han Wang, Qianyu Wang, Yunshi Lan, Ye Wang, Yuanyuan Liang, and Anqi Ding

East China Normal University

KDD 2026 Special Day of AI for Education

Presenter: Han Wang

2 of 14

Roadmap

The talk moves from the problem to the graph design,�then to evidence from evaluation.

1

Motivation

Why CFL tasks need explicit relations

2

Representation

How CFL-KG organizes educational evidence

3

Construction and system

How the graph is built and deployed

4

Evaluation

Intrinsic quality and teacher-facing utility

5

Takeaways

When graph structure adds value

2

3 of 14

Motivation: CFL Tasks Need Traceable Relations

3

CFL resources are fragmented, but teacher-facing tasks require connected evidence.

Fragmented resources

Proficiency standards and level descriptors

Characters, words, grammar, and cultural content

Assessment items, rubrics, corpora, and learner errors

Why keyword search is insufficient

Teachers must explain what a question assesses

They must trace errors to knowledge and causes

They need inspectable evidence chains, not isolated results

Core challenge: connect standards, knowledge points, questions, learner errors, and cultural evidence in one reusable structure.

English script:�International Chinese language education relies on many heterogeneous resources. Teachers and researchers need to work across proficiency standards, vocabulary lists, grammar descriptions, assessment materials, learner-error records, corpora, and cultural resources. These resources are usually stored in separate documents or systems. More importantly, many educational tasks are relational. For example, a teacher may need to identify which knowledge point a question assesses, which level it belongs to, and which learner errors are related to it. Keyword search can retrieve objects, but it does not directly provide inspectable relationships.��中文提示：�这一页讲动机。重点不是“资料多”，而是教师需要解释资源之间为什么相连。普通检索能找到材料，但不能直接给出题目—知识点—等级—偏误这样的证据链。

4 of 14

Research Gap and Research Questions

Research gap

Many educational KGs focus on single-course or single-standard settings.

CFL resources coexist across multiple standards and heterogeneous sources.

High-value links such as item–knowledge and error–knowledge still require explicit construction and validation.

RQ1

How can a multi-standard graph unify language knowledge, assessment resources, learner-error evidence, and cultural content?

RQ2

Can such a graph support teacher-facing tasks more effectively and more traceably than document search and keyword-only retrieval?

Main idea: treat educational relations as first-class, inspectable objects.

4

English script:�Existing educational knowledge graphs show the value of structured representations, but many of them focus on a single course, a single curricular standard, or general resource organization. They also often assume that educational relations, such as item-to-knowledge mappings, already exist. This assumption does not hold in international Chinese education. Therefore, our first research question is how to build a multi-standard graph that unifies language knowledge, assessment resources, learner-error evidence, and cultural content. Our second question is whether this graph can support teacher-facing tasks more effectively and more traceably than document search and keyword-only retrieval.��中文提示：�这一页把研究空白和两个研究问题说清楚。关键词是 traceably，即不仅得到答案，还能看到答案背后的关系和证据。

5 of 14

CFL-KG Overview: Data, Construction, and Functions

5

CFL-KG connects heterogeneous CFL resources, a construction pipeline, and teacher-facing functions.

Three parts are connected: data sources → KG construction pipeline → teacher-facing functions.

The graph turns heterogeneous resources into inspectable evidence paths for teaching and assessment.

6 of 14

Ontology: Three-Layer Educational Schema

Key design

Language layer: characters, words, grammar, radicals, and dependencies

Assessment layer: questions, templates, papers, and scoring standards

Cultural layer: cultural points, idioms, and intercultural cases

Level and Error nodes act as cross-layer hubs

The schema makes task-relevant relations explicit: level alignment, item–knowledge links, and error–knowledge links.

6

English script:�This slide shows the schema of CFL-KG. The graph contains three main layers. The language knowledge layer includes characters, words, grammar points, radicals, levels, and learning dependencies. The assessment resource layer contains questions, question templates, sample papers, and scoring standards, where questions are connected to the knowledge units they assess. The cultural extension layer includes cultural points, idioms, national-condition knowledge, and intercultural cases. Level and Error nodes serve as important hubs. The value of this schema is not simply storing different entities, but making educationally meaningful relations explicit, such as item-to-knowledge, error-to-knowledge, and knowledge-to-level relations.��中文提示：�这是核心结构页。不要只报实体名，要强调教育意义：题目—知识点支持评估追溯，偏误—知识点支持干预，等级对齐支持跨标准备课。

7 of 14

Data Construction and Graph Scale

7

Construction combines controlled automation with expert review for high-impact educational relations.

Automated construction

Parsing and cleaning from standards, textbooks, dictionaries, corpora, and assessment resources

Stable identifiers, normalization, deduplication, and provenance preservation

Rule-based and statistical candidate generation for alignments and links

Expert review

Domain experts review item–knowledge, error–knowledge, and cross-standard links

Approved, revised, pending, and rejected candidates are recorded

This protects assessment tracing and intervention planning from unsupported links

Graph scale

20K+

characters

250K+

words

1K+

grammar�points

500+

cultural�nodes

10K+

questions

2K+

reading�materials

High-impact educational relations are reviewed before entering the deployed graph.

English script:�CFL-KG was constructed from multiple educational resources, including standards, examination documents, dictionaries, textbooks, question banks, reading materials, corpora, and learner-error records. The process combines deterministic parsing, rule-based normalization, statistical candidate generation, and expert review. Each entity receives a stable identifier, and source provenance is preserved. Deterministic relations such as level alignment and composition can be generated from structured sources. High-impact relations such as item-to-knowledge and error-to-knowledge links are generated as candidates and then reviewed by domain experts.��中文提示：�这一页回答“数据怎么来的”。重点强调不是全自动生成，而是自动化候选 + 专家审核，尤其是会影响教学判断的高价值关系。

8 of 14

Deployed System: From Graph Data to Teacher Workflows

8

What users can do

Browse by standard and level

Search entities and inspect node details

Explore one-hop and multi-hop relations

Trace questions back to target knowledge points

Combine graph paths with corpus evidence

English script:�CFL-KG is not only a static dataset. We implemented a deployed system based on Neo4j, backend APIs, and a web interface. Users can browse entities under different standard modes, search by node type or keyword, and filter results by level. They can also explore one-hop or multi-hop relations among knowledge points, questions, errors, and cultural nodes. For assessment traceability, a user can start from a question and inspect what knowledge points it assesses, or start from a knowledge point and retrieve related questions. The system also supports direct Cypher queries and natural-language-to-Cypher querying.��中文提示：�这一页说明系统已经部署，不是只有概念图。讲的时候按照用户操作来讲：选标准、检索、看节点、展开关系、追溯题目和偏误证据。

9 of 14

Intrinsic Evaluation: Is the Graph Reliable Enough?

Before testing teacher workflows, we first audit the graph itself.

Evaluation dimensions

Standards coverage

Schema and structural quality

High-value relation correctness

Short-hop reachability

Manual samples

220 item–knowledge links

210 error–knowledge links

200 level-alignment links

Human judgment

Two raters with CFL backgrounds

Endpoint type, relation direction, source evidence, and pedagogical justification

Third adjudicator for disagreements

Inter-rater agreement is substantial, with Cohen’s κ ranging from .81 to .90 for key relation groups.

9

English script:�We evaluated CFL-KG from both intrinsic and extrinsic perspectives. The intrinsic evaluation asks whether the graph itself is sufficiently reliable for downstream educational tasks. We examined standards coverage, schema violations, correctness of key educational relations, and short-hop reachability. For relation correctness, we sampled 220 item-to-knowledge links, 210 error-to-knowledge links, and 200 level-alignment links. Two raters with international Chinese education backgrounds evaluated endpoint types, relation direction, source evidence, and educational justification. Disagreements were resolved by a third adjudicator.��中文提示：�这一页讲内在评估设计。先看图谱本身是否可靠，再看教师使用效果。样本量和 Cohen’s κ 要读出来，体现评估严谨。

10 of 14

Intrinsic Results: Coverage, Structure, and Relation Quality

97.8%

entities aligned to standards

0.6%

schema-violating edge rate

92.0%

item–knowledge precision

89.0%

error–knowledge precision

95.0%

level-alignment precision

91.2%

reachability within ≤3 hops

Interpretation: high-value educational links reach usable precision, and evidence chains remain short enough to inspect.

10

English script:�The intrinsic results show that 97.8 percent of entities were aligned with at least one standard. The schema-violating edge rate was only 0.6 percent. For high-value relations, item-to-knowledge precision reached 92 percent, error-to-knowledge precision reached 89 percent, and level-alignment precision reached 95 percent. In addition, 91.2 percent of relevant entities were reachable within three hops. These results suggest that the graph has reached a usable quality level for standards alignment, assessment tracing, and error-informed planning.��中文提示：�这一页是关键数据结果。讲数字时要解释含义：标准对齐支撑按等级筛选，题目—知识点支撑评估追溯，偏误—知识点支撑教学干预，三跳可达说明证据链不会太长。

11 of 14

Extrinsic Evaluation: Does CFL-KG Improve Teacher Workflows?

We compare graph-based evidence tracing with document search and keyword-only retrieval.

Participants

12 teachers or researchers

International Chinese education backgrounds

Within-subject design

Conditions

Manual document search

Keyword-only retrieval

Full CFL-KG

Task families

T1: standards-aligned preparation

T2: assessment traceability

T3: error-informed planning

Measures: task quality, evidence completeness, and completion time. The setup separates graph relations from simple retrieval.

11

English script:�We then conducted an extrinsic evaluation to examine whether the deployed graph improves teacher-facing workflows. Twelve teachers or researchers with international Chinese education backgrounds participated in a within-subject user study. Each participant completed tasks under three conditions: manual document search, keyword-only retrieval using the same data and interface but without relation tracing, and the full CFL-KG system. The tasks covered standards-aligned preparation, assessment traceability, and error-informed planning. We measured task quality, evidence completeness, and completion time.��中文提示：�这一页强调实验公平性。keyword-only 使用相同数据但关闭关系追踪，因此可以区分提升来自图谱关系，而不只是数字化检索。

12 of 14

Extrinsic Results: Higher Quality, Stronger Evidence, Faster Work

Task quality

0.78 → 0.91

Evidence completeness

0.71 → 0.88

Median time

245s → 109s

Largest gains appear in relation-intensive tasks, especially assessment tracing and error-informed planning.

12

English script:�The full graph condition achieved the best overall results. Average task quality increased from 0.78 for manual search and 0.84 for keyword-only retrieval to 0.91 with CFL-KG. Evidence completeness increased from 0.71 and 0.76 to 0.88. At the same time, median completion time decreased from 245 seconds for manual search and 168 seconds for keyword-only retrieval to 109 seconds with the full graph. The largest improvements appeared in assessment traceability and error-informed planning, because these tasks require users not only to find an object, but also to verify why different resources are connected.��中文提示：�这是最重要的实验结果页。三句话：质量更高、证据更完整、速度更快。最后强调 T2 和 T3 提升最明显，因为它们最需要关系链和证据链。

13 of 14

Discussion: When Does Graph Structure Matter?

Graph structure matters most when teachers need alignment, diagnosis, and justification.

What CFL-KG adds

Relational verification: inspect how items, standards, errors, and examples connect.

Evidence packaging: support intervention planning and explanation.

Efficiency for relation-intensive teacher workflows.

Limitations and next steps

Sample size is modest and tasks are controlled.

Some cultural and learner-error links still need expert judgment.

Future work: confidence metadata, review status, and authentic classroom deployment.

CFL-KG supports teachers and researchers rather than autonomous high-stakes decisions about learners.

13

English script:�Our results suggest that graph structure is especially valuable when a task requires alignment, diagnosis, and justification. For simple lookup tasks, keyword retrieval may already be sufficient. However, when users need to verify item-to-knowledge mappings or combine learner errors, causes, and examples, explicit graph paths provide a clear advantage. It is also important to clarify the role of CFL-KG. The system supports teachers and researchers rather than making autonomous high-stakes decisions about learners. The study also has limitations, including a modest participant sample and controlled tasks. Future work will expose confidence values and review-status metadata more explicitly.��中文提示：�讨论页不要重复数据，而是解释什么时候图谱有价值：简单查找时关键词可能够用；需要解释、诊断和追溯时，图谱关系更有优势。

14 of 14

Takeaways

1

CFL-KG unifies language knowledge, standards, assessment resources, learner errors, and cultural content.

2

Its value comes from inspectable educational relations, not only from data scale.

3

Evaluations show improved task quality, stronger evidence completeness, and faster teacher-facing workflows.

Thank you! Q&A

Han Wang · East China Normal University

14

English script:�To conclude, we presented CFL-KG, a multi-standard knowledge graph for international Chinese language education. It integrates language knowledge, assessment resources, learner errors, and cultural content within an explicit educational schema. We also developed a deployed system that supports cross-standard retrieval, relation exploration, assessment traceability, and error-informed planning. Both intrinsic and extrinsic evaluations indicate that CFL-KG provides a usable and traceable evidence infrastructure for teacher-facing educational tasks. More broadly, we believe that an educational knowledge graph should be evaluated not only by its size, but by whether its relations help teachers justify instructional and assessment decisions. Thank you for listening. I look forward to your questions.��中文提示：�结尾总结三个贡献：多标准知识表示、已部署系统、内外评估证明有效。最后一句是 take-home message：教育知识图谱不只看规模，更要看关系是否能帮助教师解释教学决策。