Ibnuck/poster-schedule-information-extraction

Multimodal Visual-Text Dataset for Poster Schedule Information Extraction

A ready-to-train Indonesian document AI dataset combining pixels, OCR tokens, spatial layout, and BIO entity labels.

---

Overview

What it is	127 Indonesian seminar and religious-study event posters with multimodal token-level annotations
Primary task	Schedule information extraction as token classification
Modalities	Image + text + 2D spatial layout
Coordinates	Standard LayoutLM boxes normalized independently on both axes to the 0–1000 range
Labels	24 BIO token labels, including O
Splits	Rare-label-aware train / validation / test split with seed 42

From poster pixels to structured event information:
original poster → OCR tokens and source boxes → standard LayoutLM boxes (0–1000) → BIO labels

Dataset Snapshot

Split	Documents	Share
train	101	79.5%
validation	12	9.4%
test	14	11.0%
Total	127	100%

Quick Start

from datasets import load_dataset

dataset = load_dataset(
    "Ibnuck/poster-schedule-information-extraction"
)

print(dataset)
sample = dataset["train"][0]

print(sample["document_id"])
print(sample["original_size"])
print(sample["tokens"][:10])
print(sample["bboxes"][:10])
print(sample["ner_tags"][:10])
sample["image"].show()

Expected split structure:

DatasetDict({
    train:      101 documents
    validation:  12 documents
    test:        14 documents
})

Dataset Schema

Feature	Type	Description
id	string	Identifier retained from dataset preparation
document_id	string	Human-readable unique document identifier
image	Image	Embedded original poster image
tokens	list[string]	Word-level OCR tokens in reading order
bboxes	list[list[int]]	Standard LayoutLM boxes in [x0, y0, x1, y1]; x and y are normalized independently to 0–1000
ner_tags	list[ClassLabel]	BIO token-classification labels aligned with tokens
source_category	ClassLabel	Poster category: seminar or kajian
original_size	list[int]	Original image size as [width, height]

For every example:

len(tokens) == len(bboxes) == len(ner_tags)

Complete 24-label vocabulary

Group	Labels
Outside	O
Event content	B-ACARA, I-ACARA, B-DESKRIPSI, I-DESKRIPSI
Calendar	B-HARI, B-TANGGAL, B-BULAN, B-TAHUN, B-TANGGALLENGKAP
Time	B-WAKTUMULAI, B-WAKTUSELESAI, B-WAKTULENGKAP, B-ZONAWAKTU, I-ZONAWAKTU, B-PENUNJUKWAKTU
People and place	B-PEMBICARA, I-PEMBICARA, B-LOKASI, I-LOKASI
Online access	B-LINK, I-LINK, B-KONFERENSI, I-KONFERENSI

Why Multimodal?

Event posters rarely follow a fixed reading order. Important information can be separated by icons, typography, columns, portraits, or decorative elements. Text alone misses these visual cues.

This dataset lets a model learn from three complementary signals:

Signal	What the model receives
Textual	OCR-extracted word tokens
Spatial	Standard LayoutLM coordinates normalized to 0–1000 on both axes
Visual	The poster image itself

That combination makes it a natural fit for LayoutLMv3 and related document-understanding architectures.

OCR Compatibility

The annotations are aligned at the word level: each item in tokens represents one OCR word and has one corresponding bboxes entry and one ner_tags entry.

The source annotations were created from Google Cloud Vision API output. For the closest consistency between training and inference, Google Cloud Vision API with word-level text and bounding boxes is therefore recommended. This reduces differences in word segmentation, reading order, and detected text regions between the OCR used to build the dataset and the OCR used during deployment.

Other OCR engines can also be used, provided that their output is prepared as follows:

• split detected text into individual words;
• provide one bounding box for every word;
• preserve a consistent reading order;
• normalize x and y coordinates independently to the LayoutLM range 0–1000; and
• ensure that len(tokens) == len(bboxes) before sending the inputs to the model.

Different OCR engines may segment words or detect their boundaries differently, so end-to-end performance can differ even when the same trained model is used.

Collection and Annotation

1. Poster images were collected into two source categories: seminar and kajian.
2. Before OCR, each image was resized without changing its aspect ratio so that its longest side targeted 1000 pixels.
3. Google Cloud Vision API produced tokens and pixel coordinates on that resized canvas.
4. OCR tokens were manually assigned schedule-related entity labels.
5. The published bboxes were derived by normalizing x coordinates against the OCR-canvas width and y coordinates against its height, each to the 0–1000 range.
6. Documents were divided using a rare-label-aware strategy with a fixed random seed of 42.

The split strategy attempts to retain examples containing infrequent labels across the evaluation splits where possible.

Only the standard bboxes field is published. The temporary OCR canvas is part of source-data validation and conversion, not part of the public training schema.

Data Integrity Checks

The published artifacts were validated before upload:

• 127/127 images decoded successfully
• All 127 source annotation files were converted using their corresponding OCR canvas dimensions
• Token, bboxes, and label sequences are aligned for every document
• Every source box fits the OCR canvas used during conversion
• Every published LayoutLM box contains four ordered coordinates within the inclusive range 0–1000
• All three remote splits load successfully with Hugging Face Datasets
• Validation errors found: 0

Intended Uses

• LayoutLMv3 token classification
• Multimodal document understanding
• Indonesian event-information extraction
• OCR and spatial-layout experimentation
• Low-resource document AI research
• Benchmarking preprocessing and annotation pipelines

Limitations and Responsible Use

• The dataset is small and focuses on Indonesian seminar and religious-study event posters.
• Layout, typography, OCR quality, and entity frequency vary considerably.
• OCR output and human annotations may contain errors.
• Source OCR processing used integer dimensions after aspect-ratio-preserving scaling; small rounding differences may occur if the conversion is independently reimplemented.
• Results may not generalize to other languages, regions, document types, or current events.
• Names, portraits, contact details, logos, and organizational information may appear in source posters.
• Do not treat automatically extracted dates, locations, links, or speaker names as verified facts without checking the original document.

License and Image Rights

The author-created annotations, schema, split definitions, and documentation are covered only to the extent described in the repository's custom license.

Poster images, OCR-reproduced poster text, logos, photographs, trademarks, and other third-party materials are excluded from that grant and may remain protected by their respective rights holders. Permission to use the annotations does not automatically grant permission to reuse or redistribute the underlying poster material.

Please review LICENSE before using this dataset.

Suggested Citation

@misc{ahraza2026poster_schedule_dataset,
  author       = {Ibnu Taufick Ahraza},
  title        = {Multimodal Visual-Text Dataset for Poster Schedule Information Extraction},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Ibnuck/poster-schedule-information-extraction}}
}

Author

Ibnu Taufick Ahraza
Information-extraction research project associated with Universitas Pembangunan Nasional “Veteran” Yogyakarta.

---

Built for Indonesian document AI research.
Questions, feedback, and collaboration ideas are welcome in the repository's Community tab.