Intelligent Grading: Technical Exploration and Practice in AI‑Powered Education

Speaker: Li Chao, Senior Researcher at Tencent (DataFun Summit).

Background

The talk introduces the evolution of education from traditional to online and now to intelligent education, highlighting how AI can improve learning efficiency for students and reduce teachers' workload.

Typical Educational Scenarios

Classroom : Use class performance data to tailor lesson pace and select targeted exercises.

Homework : Automatic grading with video explanations and AI virtual teachers to pinpoint student errors, following Polya’s problem‑solving steps.

Review : Personalized learning paths based on each student’s knowledge gaps.

Exams : Automated test generation and grading.

Intelligent Education Architecture

Resource layer: expert‑built knowledge graphs, question banks, exam analysis, difficulty prediction, and multimedia resources.

Algorithm layer: NLP tasks such as semantic mining, planning, text classification, and multimodal processing of images and audio.

Engine layer: automatic grading and adaptive learning engines applied to test generation, homework, diagnostics, tutoring, and exams.

Typical Problems

Construction of a knowledge‑point taxonomy aligned with curriculum standards.

Mapping questions to knowledge points and difficulty levels.

Parsing entire test papers, handling varied formats, and aligning questions with answers.

Identifying knowledge points, difficulty, and video content for each question.

Tracking student error steps to distinguish between unknown concepts and partial misunderstandings.

Adaptive Learning

After grading, the system builds a student profile indicating mastery probabilities for each knowledge point, plans personalized learning paths based on weak points and prerequisite knowledge, and recommends similar or repeated questions.

Technical Solutions

The grading system handles both subjective (essay) and objective questions.

1. Essay Grading

Word‑level analysis: idioms, poetry, conjunctions, verbs, onomatopoeia, and quoted sayings.

Sentence‑level analysis: rhetorical devices (metaphor, parallelism, personification, exaggeration), expression types (narration, argumentation, lyricism, description).

Paragraph‑level analysis: structure detection (intro‑body‑conclusion, total‑part‑total, etc.) and coherence.

Overall feedback: scoring consistency, teacher‑machine agreement, class‑level analysis, and writing quality.

2. Spell‑checking

Based on the ACL Findings 2021 paper “Read, Listen, and See: Leveraging Multimodal Information Helps Chinese Spell Checking”, the model combines semantic (Transformer), phonetic (GRU‑encoded pinyin), and graphic (ResNet‑encoded character images) encoders, concatenates their outputs, and feeds them to a Transformer for error detection.

3. Grammatical Error Detection

English GEC follows a sequence‑labeling approach similar to Grammarly, using token‑level tags for keep, delete, replace, and tense changes, with model ensembles and dependency‑based masking to improve long‑range dependencies.

4. Content Understanding

Multitask classification identifies rhetorical devices, metaphor components, parallelism, expression styles, descriptive types, narrative elements, argument strength ranking, and genre classification (narrative, argumentative, expository).

5. Other Writing Tasks

Imitation writing: matching sentence patterns.

Continuation writing: assessing coherence and completeness.

Picture‑based composition: extracting scene, objects, and actions from images to evaluate content coverage.

6. Application‑Question Grading

Handles textual answers with semantic similarity models and numeric answers by parsing equations and verifying consistency with the problem statement.

7. Fill‑in‑the‑Blank Grading

Combines OCR results with textual matching, multimodal cues, and error‑tolerant strategies to address missing strokes or ambiguous characters.

8. Other Grading Scenarios

Open‑ended reading comprehension requiring deep semantic understanding.

Handwriting quality assessment.

Geometry drawing verification.

Related Resources

References include papers on dependency‑masked BERT, multimodal Chinese spell‑checking, essay scoring regression models, argumentative structure detection, multimodal surveys, and OCR correction.

Q&A session covered methods for extracting argument points, data generation for spell‑checking, and details of Chinese grammatical error correction.

Thank you for attending.