AmyIvan/mosaic-acl2026

MOSAIC

Dataset Summary

MOSAIC is a course-centric multimodal dataset released with an ACL 2026 paper. The dataset centers on mosaic.jsonl, a JSONL file that stores course-level metadata together with nested video-level summaries, subtitles, captions, and auxiliary references.

The dataset also includes:

• data/graph_p_results/: course-level knowledge graph JSON files keyed by kg
• data/all.csv: URL-to-filename mapping for slide references
• data/pdfs/shard_xx/: sharded reference slide PDFs

Supported Tasks

• multimodal educational data understanding
• subtitle and caption analysis
• document-aware summarization
• course knowledge graph grounding
• retrieval over linked videos, graphs, and slides

Languages

The dataset is primarily in Chinese, with a smaller amount of English content in slide titles, references, and course materials.

Dataset Structure

.
├── README.md
└── data/
    ├── mosaic.jsonl
    ├── all.csv
    ├── graph_p_results/
    │   ├── BIT-1001604004.json
    │   └── ...
    └── pdfs/
        ├── shard_00/
        ├── shard_01/
        └── ...

Data Instances

Main file: data/mosaic.jsonl

Each line is one course record with the following top-level fields:

• url
• course_title
• contents
• kg
• caption_anno
• overview
• objectives
• prerequisites
• references

Each video entry inside contents[*].courses[*] contains:

• video_url
• srt_url
• summary
• subtitle
• caption
• video_title
• ref

The ref object includes:

• cate: reference category
• doc: list of reference document URLs

Knowledge graphs: data/graph_p_results/*.json

Each knowledge graph file contains a top-level object with:

• code
• message
• sampled
• traceId
• result

The main graph payload is stored in:

• result.mocKgNodeDtoList

Slide mapping: data/all.csv

Columns:

• doc_url: document URL referenced in mosaic.jsonl
• filename: corresponding PDF filename

PDFs: data/pdfs/shard_xx/

Reference slide PDFs are sharded into directories of up to 500 files each for more reliable upload and browsing.

Dataset Creation

MOSAIC is constructed from public courses on iCourse163, a major Chinese MOOC platform. The source data follows a four-level hierarchy of course, chapter, video, and topic. Each course provides course-level metadata such as objectives and prerequisite information; chapters group related videos and associated slide decks; videos include timestamped ASR transcripts, instructor-provided knowledge-point outlines, and summaries; and topics correspond to the predefined knowledge points used for alignment. Because the platform does not provide high-quality alignment between transcripts, topic inventories, and slides, the dataset constructs these links from scratch. MOSAIC is released in two subsets: MOSAIC-G, a fully human-annotated gold benchmark built from 6 diverse courses with utterance-level topic labels and utterance-to-slide alignment, and MOSAIC-S, a large silver subset for the remaining courses produced with DORA, a two-stage pipeline that first refines noisy topic inventories and then performs joint segmentation and topic assignment. For slide linkage in MOSAIC-S, the paper describes an automatic pipeline combining title matching, rule-based filtering, and LLM verification.

Statistics

Metric	Value
Courses	179
Videos	14,942
Knowledge graph JSON files	167
PDF files	10,566
Slide mapping rows	10,566
Raw size	~12.17 GB (11.34 GiB)

Licensing Information

This dataset is released under CC BY-NC-SA 4.0.

Citation Information

@inproceedings{ai-etal-2026-mosaic,
  title = {MOSAIC: A Large-Scale Multimodal Open-Course Segmentation and Alignment Corpus in Chinese},
  author = {Ai, Yuming and Fan, Shuai and Xu, Hua and Kong, Fang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics},
  year = {2026}
}