README.md

---
language:
- hy
task_categories:
- image-to-text
tags:
- ocr
- Armenian
- mixed
- benchmark
- documents
license: apache-2.0
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: page_id
    dtype: string
  - name: source
    dtype: string
  - name: split
    dtype: string
  - name: image
    dtype: image
  - name: image_width
    dtype: int32
  - name: image_height
    dtype: int32
  - name: annotations
    sequence:
    - name: id
      dtype: string
    - name: label
      dtype: string
    - name: transcription
      dtype: string
    - name: bbox
      sequence: float32
      length: 4
  splits:
  - name: test
    num_bytes: 132031004.0
    num_examples: 200
  download_size: 130758246
  dataset_size: 132031004.0
---

# Armenian Printed Documents OCR Benchmark

A benchmark dataset for evaluating OCR and document-layout systems on Armenian printed
documents. This benchmark covers **20 diverse document types** — legal acts, government 
decisions, court rulings, dictionaries, textbooks, literary works, religious texts, contracts, 
and reports — to test OCR robustness across varied layouts, fonts, and print quality.

The benchmark contains **200 pages** across 20 source documents (10 pages per source),
combining both **born-digital** PDFs (real text layer) and **scanned** PDFs (image-only,
manually transcribed).

## Overview

This corpus is designed to stress-test Armenian OCR and layout-analysis systems beyond
the newspaper domain. It enables:

- Evaluating OCR accuracy across legal, administrative, literary, and religious registers
- Testing layout-analysis robustness on both born-digital and scanned documents
- Benchmarking word-level and line-level recognition systems on the same ground truth
- Studying model behavior under different content filters (non-Armenian text, graphics/headers)

## Data Description

Each source document was rendered to page images and annotated at the paragraph level with
bounding boxes, reading order, region type, and (for a ground-truth subset) full text
transcription. Documents span both **born-digital** PDFs, where paragraph boxes were derived
from the real PDF text layer before being converted to pixel coordinates, and **scanned**
PDFs, where boxes were estimated manually from the rendered page image. In the published
dataset both are stored uniformly as pixel-space `bbox` values (see Dataset Structure below).

## Dataset Structure

Each row corresponds to one document page and contains:

| Column | Type | Description |
|---|---|---|
| `page_id` | string | Unique page identifier |
| `source` | string | Source document the page comes from (see Sources below) |
| `split` | string | Dataset split (`test`) |
| `image` | Image | Full-page render |
| `image_width` | int | Image width in pixels |
| `image_height` | int | Image height in pixels |
| `annotations` | list | Ground-truth regions (see below) |

Each entry in `annotations has:

| Field | Type | Description |
|---|---|---|
| `id` | string | Region identifier, unique within the page |
| `label` | string | Region type (see label taxonomy below) |
| `transcription` | string | Ground-truth text for the region (`""` where transcription was not collected) |
| `bbox` | float[4] | Bounding box `[x1, y1, x2, y2]` in pixels |

Regardless of whether the source page was born-digital or scanned (see the `Text layer`
column in Sources below), `bbox` coordinates are always in pixel space matching
`image_width`/`image_height`.

## Annotation Schema

### Label taxonomy

| Label | Description |
|---|---|
| `Text` | Body text paragraph |
| `Title` | Document or section title |
| `Heading` | Section/chapter heading, below title level |
| `Header` | Running header repeated at the top of pages |
| `Footer` | Running footer / page number at the bottom of pages |
| `List_Item` | Numbered or bulleted list entry |
| `Dialogue` | Dialogue or quoted-speech line (literary texts) |
| `Verse` | Poetic verse line (poetry, prayers) |
| `Signature` | Signature block or signatory name |
| `Photo` | Photograph, illustration, or other non-text graphic |
| `Divider` | Visual separator (rule, line) between regions |
| `Annotation` | Marginal note or annotation outside the main text flow |

## Sources

| Source ID | Document | Pages | Text layer |
|---|---|---|---|
| `Arlis` | Constitution of the Republic of Armenia, via the ARLIS legal database | 10 | born-digital |
| `Assembly` | Three draft laws before the National Assembly of RA: amendments to the Criminal Procedure Code (videoconferencing for detained defendants), to the Law on Prevention of Family and Domestic Violence, and to the Law on Holidays and Days of Remembrance | 10 | born-digital |
| `Bank_Contract` | Two bank service-contract templates: an Ameriabank paid-services contract and an ArmSwissBank custody/safekeeping-services contract (approved by Board Decision N84-L, 28.08.2014) | 10 | born-digital |
| `Chess_Textbook` | "Շախմատ" (Chess), grade-2 school textbook by L. Martirosyan, S. Misakyan, N. Movsisyan (2023) | 10 | born-digital |
| `Code_Of_Laws` | Criminal Code of the Republic of Armenia (adopted 18 April 2003) | 10 | born-digital |
| `Court_Decision` | Constitutional Court of RA decision on the conformity of the EAEU–Indonesia free trade agreement (signed St. Petersburg, 21 Dec 2025) with the Constitution | 10 | born-digital |
| `Dictionary` | "Ամենագործածական տնտեսագիտական տերմինների բացատրական բառարան" — explanatory dictionary of common economic terms | 10 | scanned |
| `Dissertation` | "Էլեկտրոնային մամուլի լեզուն (բառային կազմ, քերականական համակարգ)" — dissertation by A. A. Khachatryan, Institute of Language after Hr. Acharyan, NAS RA | 10 | born-digital |
| `Fables` | Collection of fable-poems ("Առակներ"); individual pieces dated 1910–1945, written in Venice/Paris/Yerevan | 10 | born-digital |
| `Gospel_Images` | Illustrated gospel book (image-heavy pages) | 10 | scanned |
| `Gov_Decision` | Three RA Government decisions/drafts: N1763-L (12 Oct 2023, social assistance for persons displaced from Nagorno-Karabakh), N1820-L (19 Oct 2023, amending decision N922-L), and a 2017 draft resolution approving the Cybersecurity Strategy | 10 | mixed (scanned + born-digital) |
| `Gov_Program` | RA Government Programme 2021–2026 (Annex to Government Decision N1363-A, 18 Aug 2021) | 10 | born-digital |
| `History_Textbook` | "Հայոց պատմություն" (History of Armenia), grade-7 school textbook by Smbat Hovhannisyan (Yerevan, 2023) | 10 | born-digital |
| `Human_Comedy` | Armenian translation of William Saroyan's novel "The Human Comedy," translated by Grigor Keshishyan | 10 | scanned |
| `Jonathan_Livingston` | Armenian translation of Richard Bach's "Jonathan Livingston Seagull," translated by Naira Harutyunyan | 10 | born-digital |
| `Order` | Official Armenian translation of Order No. 249 of the General Administration of Customs of the People's Republic of China | 10 | born-digital |
| `OSCE_Report` | OSCE/ODIHR report on the Republic of Armenia Extraordinary Parliamentary Elections, 20 June 2021 | 10 | born-digital |
| `Poetry` | Poetry collection of Hovhannes Shiraz, compiled by Anush Sarumyan ("Dizak Plus" publishing, Stepanakert, 2016) | 10 | born-digital |
| `Police_Order` | Police orders/decrees (legacy font encoding prevents clean text extraction from the source PDFs) | 10 | scanned |
| `Prayer` | "Ամենօրյա աղոթագիրք" (Daily Prayer Book), published by the Mother See of Holy Etchmiadzin, 2011 | 10 | scanned |

Each source contributes 10 pages (200 pages total).

> **Note on rights:** All source documents were collected from openly available, open-sourced
> sources.

## Intended Use

This corpus is intended for:

- Benchmarking Armenian OCR systems across document types beyond periodicals
- Evaluating document-layout analysis on mixed born-digital/scanned inputs
- Comparing word-level vs. line-level OCR architectures on identical ground truth
- Research on OCR post-processing and text normalization for Armenian

# Evaluation

Predictions against this dataset are scored with the same evaluation kit
used for [RLALT/ACoPPer](https://huggingface.co/datasets/RLALT/ACoPPer):
see [`evaluation_kit/`](https://huggingface.co/datasets/RLALT/ACoPPer/tree/main/evaluation_kit)
(in particular [`evaluate_from_hf.py`](https://huggingface.co/datasets/RLALT/ACoPPer/blob/main/evaluation_kit/evaluate_from_hf.py),
which also accepts `--dataset RLALT/ACoPDoc` directly).

# OCR Evaluation Results

CER (Character Error Rate) summary for each report filter variant.


## Document structure-aware models

These models output one bounding box + recognized text per already-complete text line. Each line becomes its own predicted row as-is — no grouping step is applied, so the line breaks scored are exactly what the model itself produced. This means document structure-aware CER reflects both character recognition accuracy *and* the model's own line-segmentation quality (over- or under-splitting a line shows up directly in the score), unlike document structure-independent scoring where line boundaries come from a separate, model-independent grouping heuristic.

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 4.66% | 4.33% | 4.66% | 4.33% |
| tesseract | 5.91% | 5.58% | 5.91% | 5.58% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **2.06%** | **1.85%** | **2.06%** | **1.85%** |
| portmind | 8.49% | 8.31% | 8.49% | 8.31% |
| abbyy | 3.93% | 3.63% | 3.93% | 3.63% |
| google-vision | 4.34% | 4.33% | 4.34% | 4.33% |

## Document structure-independent models

These models output one bounding box + recognized text per individual word. Before scoring script spatially groups words into lines using vertical-overlap and horizontal-gap heuristics: words that share a line are concatenated in reading order into a single predicted row, which is then matched against whichever ground-truth box(es) it falls inside and compared via character edit distance. Because line breaks are reconstructed rather than given, document structure-independent scores also depend on how well that grouping heuristic reassembles lines — a good recognizer can still lose points here if word boxes are noisy enough to throw off the grouping.

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 3.92% | 3.65% | 3.92% | 3.65% |
| tesseract | 5.28% | 5.01% | 5.28% | 5.01% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 2.06% | 1.85% | 2.05% | 1.85% |
| **portmind** | **1.70%** | **1.50%** | **1.70%** | **1.50%** |
| abbyy | 3.78% | 3.51% | 3.78% | 3.51% |
| google-vision | 3.96% | 3.95% | 3.96% | 3.95% |

**Additional models — [DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR), [Qwen-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), and [Chandra](https://github.com/datalab-to/chandra) — were also evaluated on this dataset but produced substantially higher error rates (50%+ CER) and are not included in the tables above.**


## Per-document breakdown

CER by document/category group, for every model with per-document evaluation
output available, split into document structure-aware and -independent tables per document (see
above for what distinguishes the two).

### 1-Arlis

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.42% | 0.42% | 0.42% | 0.42% |
| tesseract | 0.83% | 0.83% | 0.83% | 0.83% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.23% | 0.23% | 0.23% | 0.23% |
| **portmind** | **0.20%** | **0.20%** | **0.20%** | **0.20%** |
| abbyy | 0.97% | 0.97% | 0.97% | 0.97% |
| google-vision | 1.78% | 1.78% | 1.78% | 1.78% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.54% | 0.54% | 0.54% | 0.54% |
| tesseract | 0.95% | 0.95% | 0.95% | 0.95% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.21% | 0.21% | 0.21% | 0.21% |
| **portmind** | **0.20%** | **0.20%** | **0.20%** | **0.20%** |
| abbyy | 0.97% | 0.97% | 0.97% | 0.97% |
| google-vision | 1.78% | 1.78% | 1.78% | 1.78% |

*10 page(s) in this group.*

### 2-Assembly

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 6.24% | 1.72% | 6.24% | 1.72% |
| tesseract | 7.68% | 3.27% | 7.68% | 3.27% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 3.68% | **0.87%** | 3.68% | **0.87%** |
| portmind | 18.70% | 16.79% | 18.70% | 16.79% |
| abbyy | 5.66% | 1.68% | 5.66% | 1.68% |
| google-vision | **3.01%** | 3.08% | **3.01%** | 3.08% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.49% | 1.82% | 5.49% | 1.82% |
| tesseract | 6.84% | 3.43% | 6.84% | 3.43% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 3.68% | **0.87%** | 3.68% | **0.87%** |
| portmind | 4.21% | 1.56% | 4.21% | 1.56% |
| abbyy | 5.05% | 1.61% | 5.05% | 1.61% |
| google-vision | **3.01%** | 3.08% | **3.01%** | 3.08% |

*10 page(s) in this group.*

### 3-Bank_Contract

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 4.11% | 3.85% | 4.11% | 3.85% |
| tesseract | 4.49% | 4.23% | 4.49% | 4.23% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **2.55%** | **2.56%** | **2.55%** | **2.56%** |
| portmind | 5.93% | 5.92% | 5.93% | 5.92% |
| abbyy | 4.21% | 3.93% | 4.21% | 3.93% |
| google-vision | 4.52% | 4.51% | 4.52% | 4.51% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 3.30% | 3.04% | 3.30% | 3.04% |
| tesseract | 3.72% | 3.45% | 3.72% | 3.45% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **2.55%** | **2.56%** | **2.55%** | **2.56%** |
| portmind | 2.64% | 2.62% | 2.64% | 2.62% |
| abbyy | 4.11% | 3.84% | 4.11% | 3.84% |
| google-vision | 4.67% | 4.66% | 4.67% | 4.66% |

*10 page(s) in this group.*

### 4-Chess_Textbook

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 20.45% | 20.45% | 20.45% | 20.45% |
| tesseract | 22.08% | 22.08% | 22.08% | 22.08% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **2.97%** | **2.97%** | **2.97%** | **2.97%** |
| portmind | 9.33% | 9.33% | 9.33% | 9.33% |
| abbyy | 5.43% | 5.43% | 5.43% | 5.43% |
| google-vision | 6.81% | 6.81% | 6.81% | 6.81% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.28% | 5.28% | 5.28% | 5.28% |
| tesseract | 6.28% | 6.28% | 6.28% | 6.28% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **2.64%** | **2.64%** | **2.64%** | **2.64%** |
| portmind | 3.69% | 3.69% | 3.69% | 3.69% |
| abbyy | 5.05% | 5.05% | 5.05% | 5.05% |
| google-vision | 6.59% | 6.59% | 6.59% | 6.59% |

*10 page(s) in this group.*

### 5-Code_Of_Laws

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 4.33% | 4.33% | 4.33% | 4.33% |
| tesseract | 3.91% | 3.91% | 3.91% | 3.91% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.70% | 0.70% | 0.70% | 0.70% |
| **portmind** | **0.42%** | **0.42%** | **0.42%** | **0.42%** |
| abbyy | 1.62% | 1.62% | 1.62% | 1.62% |
| google-vision | 2.45% | 2.45% | 2.45% | 2.45% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.19% | 5.19% | 5.19% | 5.19% |
| tesseract | 5.15% | 5.15% | 5.15% | 5.15% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.88% | 0.88% | 0.88% | 0.88% |
| **portmind** | **0.42%** | **0.42%** | **0.42%** | **0.42%** |
| abbyy | 1.62% | 1.62% | 1.62% | 1.62% |
| google-vision | 2.45% | 2.45% | 2.45% | 2.45% |

*10 page(s) in this group.*

### 6-Court_Decision

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.93% | 0.93% | 0.93% | 0.93% |
| tesseract | 1.56% | 1.56% | 1.56% | 1.56% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **0.25%** | **0.25%** | **0.25%** | **0.25%** |
| portmind | 1.49% | 1.49% | 1.49% | 1.49% |
| abbyy | 0.52% | 0.52% | 0.52% | 0.52% |
| google-vision | 4.08% | 4.08% | 4.08% | 4.08% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.77% | 0.77% | 0.77% | 0.77% |
| tesseract | 1.40% | 1.40% | 1.40% | 1.40% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.44% | 0.44% | 0.44% | 0.44% |
| portmind | 1.02% | 1.02% | 1.02% | 1.02% |
| **abbyy** | **0.38%** | **0.38%** | **0.38%** | **0.38%** |
| google-vision | 4.08% | 4.08% | 4.08% | 4.08% |

*10 page(s) in this group.*

### 7-Dictionary

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 23.10% | 23.10% | 23.10% | 23.10% |
| tesseract | 23.19% | 23.19% | 23.19% | 23.19% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 2.20% | 2.20% | 2.20% | 2.20% |
| **portmind** | **0.84%** | **0.84%** | **0.84%** | **0.84%** |
| abbyy | 15.18% | 15.18% | 15.18% | 15.18% |
| google-vision | 15.24% | 15.24% | 15.24% | 15.24% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 19.44% | 19.44% | 19.44% | 19.44% |
| tesseract | 19.73% | 19.73% | 19.73% | 19.73% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 2.27% | 2.27% | 2.27% | 2.27% |
| **portmind** | **0.84%** | **0.84%** | **0.84%** | **0.84%** |
| abbyy | 15.18% | 15.18% | 15.18% | 15.18% |
| google-vision | 7.51% | 7.51% | 7.51% | 7.51% |

*10 page(s) in this group.*

### 8-Dissertation

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 3.79% | 3.66% | 3.79% | 3.66% |
| tesseract | 4.47% | 4.32% | 4.47% | 4.32% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **1.47%** | **1.37%** | **1.47%** | **1.37%** |
| portmind | 2.72% | 2.59% | 2.72% | 2.59% |
| abbyy | 4.74% | 4.60% | 4.74% | 4.60% |
| google-vision | 4.94% | 4.80% | 4.94% | 4.80% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 3.64% | 3.50% | 3.64% | 3.50% |
| tesseract | 4.53% | 4.34% | 4.53% | 4.34% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **1.47%** | **1.37%** | **1.47%** | **1.37%** |
| portmind | 2.72% | 2.59% | 2.72% | 2.59% |
| abbyy | 4.48% | 4.34% | 4.48% | 4.34% |
| google-vision | 4.88% | 4.74% | 4.88% | 4.74% |

*10 page(s) in this group.*

### 9-Fables

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| **hye_tesseract** | **6.10%** | **6.10%** | **6.10%** | **6.10%** |
| tesseract | 34.83% | 34.83% | 34.83% | 34.83% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 8.40% | 8.40% | 8.40% | 8.40% |
| portmind | 31.53% | 31.53% | 31.53% | 31.53% |
| abbyy | 35.13% | 35.13% | 35.13% | 35.13% |
| google-vision | 8.93% | 8.93% | 8.93% | 8.93% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.78% | 5.78% | 5.78% | 5.78% |
| tesseract | 34.81% | 34.81% | 34.81% | 34.81% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 8.48% | 8.48% | 8.48% | 8.48% |
| **portmind** | **4.33%** | **4.33%** | **4.33%** | **4.33%** |
| abbyy | 35.08% | 35.08% | 35.08% | 35.08% |
| google-vision | 8.88% | 8.88% | 8.88% | 8.88% |

*10 page(s) in this group.*

### 10-Gospel_Images

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.49% | 5.49% | 5.49% | 5.49% |
| tesseract | 8.01% | 8.01% | 8.01% | 8.01% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **1.86%** | **1.86%** | **1.86%** | **1.86%** |
| portmind | 11.92% | 11.92% | 11.92% | 11.92% |
| abbyy | 5.40% | 5.40% | 5.40% | 5.40% |
| google-vision | 4.74% | 4.74% | 4.74% | 4.74% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.85% | 5.85% | 5.85% | 5.85% |
| tesseract | 8.42% | 8.42% | 8.42% | 8.42% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **1.86%** | **1.86%** | **1.86%** | **1.86%** |
| portmind | 2.32% | 2.32% | 2.32% | 2.32% |
| abbyy | 5.24% | 5.24% | 5.24% | 5.24% |
| google-vision | 5.00% | 5.00% | 5.00% | 5.00% |

*10 page(s) in this group.*

### 11-Gov_Decision

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 3.01% | 2.56% | 3.01% | 2.56% |
| tesseract | 3.72% | 3.14% | 3.72% | 3.14% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **1.65%** | **1.05%** | **1.65%** | **1.05%** |
| portmind | 14.36% | 13.88% | 14.36% | 13.88% |
| abbyy | 2.41% | 1.88% | 2.41% | 1.88% |
| google-vision | 3.13% | 3.16% | 3.13% | 3.16% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | **1.64%** | 1.29% | **1.64%** | 1.29% |
| tesseract | 2.39% | 1.87% | 2.39% | 1.87% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 1.65% | **1.05%** | 1.65% | **1.05%** |
| portmind | 1.74% | 1.17% | 1.74% | 1.17% |
| abbyy | 2.32% | 1.79% | 2.32% | 1.79% |
| google-vision | 2.96% | 2.99% | 2.96% | 2.99% |

*10 page(s) in this group.*

### 12-Gov_Program

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.30% | 0.30% | 0.30% | 0.30% |
| tesseract | 1.35% | 1.35% | 1.35% | 1.35% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **0.17%** | **0.17%** | **0.17%** | **0.17%** |
| portmind | 0.98% | 0.98% | 0.98% | 0.98% |
| abbyy | 0.33% | 0.33% | 0.33% | 0.33% |
| google-vision | 2.30% | 2.30% | 2.30% | 2.30% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.30% | 0.30% | 0.30% | 0.30% |
| tesseract | 1.35% | 1.35% | 1.35% | 1.35% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.17% | 0.17% | 0.17% | 0.17% |
| **portmind** | **0.16%** | **0.16%** | **0.16%** | **0.16%** |
| abbyy | 0.33% | 0.33% | 0.33% | 0.33% |
| google-vision | 2.26% | 2.26% | 2.26% | 2.26% |

*10 page(s) in this group.*

### 13-History_Textbook

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 9.11% | 9.11% | 9.11% | 9.11% |
| tesseract | 9.34% | 9.34% | 9.34% | 9.34% |
| [**ocarm***](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | **2.02%** | **2.02%** | **2.02%** | **2.02%** |
| portmind | 11.85% | 11.85% | 11.85% | 11.85% |
| abbyy | 4.70% | 4.70% | 4.70% | 4.70% |
| google-vision | 3.67% | 3.67% | 3.67% | 3.67% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 5.27% | 5.27% | 5.27% | 5.27% |
| tesseract | 5.93% | 5.93% | 5.93% | 5.93% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 2.01% | 2.01% | 2.01% | 2.01% |
| **portmind** | **1.89%** | **1.89%** | **1.89%** | **1.89%** |
| abbyy | 3.81% | 3.81% | 3.81% | 3.81% |
| google-vision | 3.35% | 3.35% | 3.35% | 3.35% |

*10 page(s) in this group.*

### 14-Human_Comedy

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| **hye_tesseract** | **2.46%** | **2.46%** | **2.46%** | **2.46%** |
| tesseract | 10.91% | 10.91% | 10.91% | 10.91% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 14.15% | 14.15% | 14.15% | 14.15% |
| portmind | 7.55% | 7.55% | 7.55% | 7.55% |
| abbyy | 7.30% | 7.30% | 7.30% | 7.30% |
| google-vision | 7.03% | 7.03% | 7.03% | 7.03% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| **hye_tesseract** | **2.48%** | **2.48%** | **2.48%** | **2.48%** |
| tesseract | 10.95% | 10.95% | 10.95% | 10.95% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 14.12% | 14.12% | 14.12% | 14.12% |
| portmind | 3.29% | 3.29% | 3.29% | 3.29% |
| abbyy | 7.28% | 7.28% | 7.28% | 7.28% |
| google-vision | 6.90% | 6.90% | 6.90% | 6.90% |

*10 page(s) in this group.*

### 15-Jonathan_Livingston

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.65% | 0.65% | 0.65% | 0.65% |
| **tesseract** | **0.33%** | **0.33%** | **0.33%** | **0.33%** |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.96% | 0.96% | 0.96% | 0.96% |
| portmind | 21.27% | 21.27% | 21.27% | 21.27% |
| abbyy | 1.00% | 1.00% | 1.00% | 1.00% |
| google-vision | 3.27% | 3.27% | 3.27% | 3.27% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.94% | 0.94% | 0.94% | 0.94% |
| tesseract | 0.62% | 0.62% | 0.62% | 0.62% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.96% | 0.96% | 0.96% | 0.96% |
| **portmind** | **0.44%** | **0.44%** | **0.44%** | **0.44%** |
| abbyy | 1.00% | 1.00% | 1.00% | 1.00% |
| google-vision | 3.28% | 3.28% | 3.28% | 3.28% |

*10 page(s) in this group.*

### 16-OSCE_Report

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| **hye_tesseract** | **0.93%** | **0.89%** | **0.93%** | **0.89%** |
| tesseract | 1.87% | 1.83% | 1.87% | 1.83% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.96% | 0.96% | 0.96% | 0.96% |
| portmind | 17.38% | 17.39% | 17.38% | 17.39% |
| abbyy | 2.11% | 2.07% | 2.11% | 2.07% |
| google-vision | 2.90% | 2.88% | 2.90% | 2.88% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.70% | 0.68% | 0.70% | 0.68% |
| tesseract | 1.74% | 1.71% | 1.74% | 1.71% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.95% | 0.95% | 0.95% | 0.95% |
| **portmind** | **0.60%** | **0.60%** | **0.60%** | **0.60%** |
| abbyy | 1.88% | 1.88% | 1.88% | 1.88% |
| google-vision | 2.85% | 2.84% | 2.85% | 2.84% |

*10 page(s) in this group.*

### 17-Order

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 0.58% | 0.58% | 0.58% | 0.58% |
| tesseract | 1.08% | 1.08% | 1.08% | 1.08% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.60% | 0.60% | 0.60% | 0.60% |
| portmind | 18.92% | 18.92% | 18.92% | 18.92% |
| **abbyy** | **0.16%** | **0.16%** | **0.16%** | **0.16%** |
| google-vision | 2.05% | 2.05% | 2.05% | 2.05% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 1.29% | 1.29% | 1.29% | 1.29% |
| tesseract | 2.30% | 2.30% | 2.30% | 2.30% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 0.60% | 0.60% | 0.60% | 0.60% |
| portmind | 0.77% | 0.77% | 0.77% | 0.77% |
| **abbyy** | **0.16%** | **0.16%** | **0.16%** | **0.16%** |
| google-vision | 2.03% | 2.03% | 2.03% | 2.03% |

*10 page(s) in this group.*

### 18-Poetry

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 1.47% | 1.47% | 1.47% | 1.47% |
| **tesseract** | **0.99%** | **0.99%** | **0.99%** | **0.99%** |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 1.68% | 1.68% | 1.68% | 1.68% |
| portmind | 4.17% | 4.17% | 4.17% | 4.17% |
| abbyy | 1.05% | 1.05% | 1.05% | 1.05% |
| google-vision | 4.68% | 4.68% | 4.68% | 4.68% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 1.47% | 1.47% | 1.47% | 1.47% |
| tesseract | 0.99% | 0.99% | 0.99% | 0.99% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 1.70% | 1.70% | 1.70% | 1.70% |
| **portmind** | **0.84%** | **0.84%** | **0.84%** | **0.84%** |
| abbyy | 1.05% | 1.05% | 1.05% | 1.05% |
| google-vision | 4.68% | 4.68% | 4.68% | 4.68% |

*10 page(s) in this group.*

### 19-Police_Order

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 1.87% | 1.87% | 1.82% | 1.82% |
| tesseract | 1.71% | 1.71% | 1.66% | 1.66% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 1.09% | 1.09% | 1.06% | 1.06% |
| portmind | 14.40% | 14.40% | 14.36% | 14.36% |
| **abbyy** | **0.96%** | **0.96%** | **0.91%** | **0.91%** |
| google-vision | 4.65% | 4.65% | 4.62% | 4.62% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 1.94% | 1.94% | 1.89% | 1.89% |
| tesseract | 1.77% | 1.77% | 1.72% | 1.72% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 1.01% | 1.01% | 0.99% | 0.99% |
| portmind | 2.42% | 2.42% | 2.37% | 2.37% |
| **abbyy** | **0.96%** | **0.96%** | **0.91%** | **0.91%** |
| google-vision | 4.65% | 4.65% | 4.62% | 4.62% |

*10 page(s) in this group.*

### 20-Prayer

**Document structure-aware models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 21.12% | 21.12% | 21.12% | 21.12% |
| tesseract | 27.99% | 27.99% | 27.99% | 27.99% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 8.44% | 8.44% | 8.44% | 8.44% |
| **portmind** | **7.08%** | **7.08%** | **7.08%** | **7.08%** |
| abbyy | 10.73% | 10.73% | 10.73% | 10.73% |
| google-vision | 8.51% | 8.51% | 8.51% | 8.51% |

**Document structure-independent models**

| Model | w/o filters | filter non-Armenian | filter graphics & headers | filter all |
|---|---|---|---|---|
| hye_tesseract | 20.99% | 20.99% | 20.99% | 20.99% |
| tesseract | 27.86% | 27.86% | 27.86% | 27.86% |
| [ocarm*](https://huggingface.co/datasets/RLALT/ACoPDoc/blob/main/README.md#ocarm) | 7.80% | 7.80% | 7.80% | 7.80% |
| **portmind** | **4.07%** | **4.07%** | **4.07%** | **4.07%** |
| abbyy | 10.73% | 10.73% | 10.73% | 10.73% |
| google-vision | 8.82% | 8.82% | 8.82% | 8.82% |

*10 page(s) in this group.*

### OCArm*

OCArm is an OCR pipeline purpose-built for historical Armenian printed periodicals, combining a text detection stage and a text recognition stage to transcribe scanned newspaper pages. OCArm's text detection stage uses CRAFT (Character Region Awareness for Text Detection, [arXiv:1904.01941](https://arxiv.org/abs/1904.01941)), which localizes individual characters through region and affinity heatmaps and then groups them into word-level bounding boxes, making it robust to curved and irregularly spaced text. For text recognition, OCArm uses a deep-text-recognition-benchmark-style model ([arXiv:1904.01906](https://arxiv.org/abs/1904.01906)), built from a feature-extraction backbone, a sequence-modeling stage, and a prediction head (CTC or attention-based decoding). Detection is run first to produce word crops, which are then passed to the recognition model to transcribe the text in each box.

Since real annotated historical newspaper data is limited, OCArm was trained using two complementary data generation approaches: fully synthetic generation (procedurally created backgrounds, layouts, and rendered Armenian text with noise) and real image-based generation (using actual scanned newspapers with original text removed and new text rendered onto authentic paper textures via layout detection with PP-DocLayoutV2, an RT-DETR-based model ([arXiv:2510.14528](https://arxiv.org/abs/2510.14528))). The recognition model itself was trained specifically on cropped word boxes cut out from these generated images.

## Acknowledgements

The research was supported by the Higher Education and Science Committee of MESCS RA (Research project № 27TARGET-6B173).