# CzechHistoricalNewspapers — Dataset Recap

## Overview

`CzechHistoricalNewspapers.json` is a COCO-style annotation dataset of scanned Czech historical newspaper pages. It is the primary data source for `Clusterring.ipynb`, which uses the annotations as text input for clustering experiments and the cluster arrays as ground-truth labels for evaluation.

### Top-level keys

| Key | Description |
|---|---|
| `images` | 30 scanned newspaper page records |
| `categories` | 9 region type definitions |
| `annotations` | 1 328 bounding-box annotations across all pages |
| `clusters` | 30 arrays of ground-truth cluster labels, one per page |

---

## Categories

Each annotation belongs to one category, describing the type of layout region it covers.

| ID | Name | Supercategory | Role in notebook |
|---|---|---|---|
| 1 | SeparatorRegion | — | **Filtered out** (no text) |
| 2 | Heading | Article | Kept — article headline text |
| 3 | Caption | Article | **Filtered out** |
| 4 | Enumeration | Article | Kept — list-style text |
| 5 | TextRegion | Article | Kept — main body text (largest group) |
| 6 | ImageRegion | Article | **Filtered out** (no text) |
| 7 | Article | — | Defined but **zero annotations in dataset** |
| 8 | TableRegion | Article | Kept — tabular text |
| 9 | AdvertRegion | — | Kept — advertisement text |

The notebook filters `category_id in [1, 3, 6, 7]` before building page text. Category 7 has no annotations, so its filter entry is a no-op.

### Annotation counts by category

| ID | Name | Count | % of total | Has text content |
|---|---|---|---|---|
| 1 | SeparatorRegion | 297 | 22.4 % | No |
| 2 | Heading | 132 | 9.9 % | Yes |
| 3 | Caption | 67 | 5.0 % | Yes (but filtered) |
| 4 | Enumeration | 17 | 1.3 % | Yes |
| 5 | TextRegion | 572 | 43.1 % | Yes |
| 6 | ImageRegion | 83 | 6.2 % | No |
| 7 | Article | 0 | 0.0 % | — |
| 8 | TableRegion | 1 | 0.1 % | Yes |
| 9 | AdvertRegion | 159 | 12.0 % | Yes |
| — | **Total** | **1 328** | **100 %** | — |

Of the 1 328 annotations, **940 carry text content** (`metadata.content`) and **388 do not** (mainly SeparatorRegions and ImageRegions). The notebook wraps the content access in `try/except` to silently skip these.

---

## Images

30 pages from two newspaper series:

| Series | Pages |
|---|---|
| Brnenske | 12 |
| Lidove | 18 |
| **Total** | **30** |

Image IDs run 1–30, matching the order of the `pictures` dict in the notebook.

---

## Clusters (ground-truth labels)

One cluster array per page. Array length equals the number of text-bearing annotations on that page (after category filtering). The notebook uses `max(array) + 1` to set the number of clusters for each algorithm.

| Page | Text annotations | Unique clusters |
|---|---|---|
| Brnenske_48_D.jpg | 17 | 8 |
| Brnenske_48_C.jpg | 49 | 3 |
| Brnenske_48_B.jpg | 40 | 8 |
| Brnenske_48_A.jpg | 36 | 5 |
| Brnenske_44_D.jpg | 12 | 7 |
| Brnenske_44_C.jpg | 35 | 11 |
| Brnenske_44_B.jpg | 36 | 6 |
| Brnenske_44_A.jpg | 33 | 5 |
| Brnenske_09_D.jpg | 13 | 8 |
| Brnenske_09_C.jpg | 39 | 9 |
| Brnenske_09_B.jpg | 33 | 1 |
| Brnenske_09_A.jpg | 17 | 6 |
| Lidove_07_B.jpg | 43 | 9 |
| Lidove_07_A.jpg | 35 | 7 |
| Lidove_06_L.jpg | 44 | 14 |
| Lidove_06_K.jpg | 12 | 12 |
| Lidove_06_J.jpg | 11 | 11 |
| Lidove_06_I.jpg | 33 | 25 |
| Lidove_06_H.jpg | 43 | 6 |
| Lidove_06_G.jpg | 38 | 3 |
| Lidove_06_F.jpg | 19 | 19 |
| Lidove_06_E.jpg | 31 | 27 |
| Lidove_06_D.jpg | 20 | 4 |
| Lidove_06_C.jpg | 33 | 4 |
| Lidove_06_B.jpg | 39 | 10 |
| Lidove_06_A.jpg | 28 | 4 |
| Lidove_04_D.jpg | 13 | 13 |
| Lidove_04_C.jpg | 23 | 15 |
| Lidove_04_B.jpg | 37 | 3 |
| Lidove_04_A.jpg | 19 | 3 |

Notable: `Brnenske_09_B.jpg` has 33 annotations all assigned to the same cluster (unique_clusters = 1), meaning every text region on that page belongs to one article. Pages like `Lidove_06_E.jpg` (27 clusters from 31 annotations) are highly fragmented, with nearly every annotation being a separate article.

---

## Coherence with Clusterring.ipynb

The JSON is fully aligned with the notebook's expectations:

- All keys the notebook reads (`categories`, `annotations`, `clusters`) are present.
- The 30 cluster keys match the 30 filenames in the notebook's `pictures` dict (the notebook strips the trailing `_` from filenames via `base_name[::-1].replace('_','',1)[::-1]`).
- Cluster label array lengths match filtered annotation counts exactly for every page.
- Czech text in `metadata.content` is valid UTF-8; Python on Colab reads it correctly.

### Known bugs in the notebook (not JSON issues)

| Location | Bug |
|---|---|
| Image-masking cell | Uses `annota` (undefined); should be `annotations = data['annotations']` — causes `NameError` |
| Image+text embedding loop | `counter` is never incremented inside the loop, so `pages[counter]` always returns `pages[0]` for all 30 pages |