Armenian Corpus of Printed Periodicals (ACoPPer)

The Armenian Corpus of Printed Periodicals (ACoPPer) is an expanding digital archive of written texts.

By compiling digitized periodicals from diverse domains spanning the late 19th and 20th centuries, ACoPPer serves as a powerful tool for quantitative humanities research, while enabling users to explore cultural phenomena and linguistic trends from the era.

Overview

ACoPPer is designed to support advanced quantitative research in digital humanities. It enables researchers to:

• Track linguistic change over time
• Analyze cultural and social developments reflected in period media
• Study historical trends through printed sources
• Build and evaluate language technologies for Armenian

Data Description

The dataset consists of digitized texts extracted from historical Armenian newspapers and periodicals. It includes content from multiple genres and topics, reflecting real-world usage of Armenian in print media across different historical periods.

Dataset Structure

Each row corresponds to one scanned newspaper page and contains:

Column	Type	Description
page_id	string	Unique page identifier
source	string	Newspaper issue the page comes from
split	string	Dataset split (pilot)
image	Image	Full-page scan
image_width	int	Original image width in pixels
image_height	int	Original image height in pixels
annotations	list	Ground-truth regions and metadata

Each annotation contains:

Field	Type	Description
id	string	Region identifier
label	string	Region type
transcription	string	Ground-truth text ("" for non-text regions)
reading_order	int	Reading order within the parent region
parent_id	string	Parent region identifier
bbox	float[4]	Bounding box [x1, y1, x2, y2] in pixels
rotation	float	Rotation angle in degrees

Annotation Schema

The dataset provides detailed document-layout annotations for historical Armenian newspapers, including:

• Article structure (Kicker, Headline, Deck, Lead, Text, Subhead)
• Newspaper metadata (Nameplate, Issueline, Folioline, Editorline, Masthead, Page number)
• Visual content (Photo, Front picture, Graphics, Cartoon)
• Advertisements and announcements (Ads, Obituaries)
• Editorial content (Editorial, Column, Letters to the editor)
• Auxiliary elements (Caption, Byline, Jump line, Rule, Table, and others)

The annotation scheme contains 38 region categories and preserves the logical hierarchy of newspaper pages through explicit parent–child relationships.

Full Label Taxonomy

Label	Annotation guideline	Formal definition
Ads	Generally, announcements which are both classic advertisements and small general announcements by individuals or businesses offering a service or product, including house sales or rentals, tailoring or shoemaking services, announcements of sporting events, etc. Excludes obituaries (→ Obituaries).	Message paid for by an advertiser to inform readers about a business, product or service. Incl. classified advertisements: short ads placed by individuals, grouped into categories.
Banner	Large headline — necessarily combines several headlines: column, article, short messages. Differs from Kicker, which has almost the same logic but relates to only one headline.	Large headline appearing immediately below the heading and running across multiple columns.
Byline	Data about the author of primarily authored articles and materials — name, surname, including position and any other information related to the author.	Name of the writer.
Caption	Descriptions of photographs, any graphic images, cartoons, tables, formulas, etc., or their accompanying explanatory text pieces, which may also be included in the image.	Short explanatory text accompanying a photograph, image, or illustration.
Cartoon	Used exclusively for caricatures.	Humorous or satirical drawing; usually accompanied by a caption and comments on a news event.
Column	Columns, other than editorial and leading, that are periodic in nature in terms of thematic content or are determined by the columnist's authority and influence.	Regularly published article that presents the comments of one author (reporter or personality) on a chosen subject.
Deck	Subtitle — text following the title, completing the title, and graphically different from the title.	Short block of text under the headline that completes it. (ArticleSubTitle)
Editorial	Editorial.	In-depth article that reflects the collective viewpoint of a newspaper's editorial board.
Editorline	Information about the editor of the periodical or the editor responsible for the issue (e.g. "Editor: A. Poghosyan"), which in Soviet publications is separated from the issue and editorial data. Must have the Issueline ID in the Optional Parent box ID.	—
Folioline	Periodical-related metadata on the second and subsequent pages (not the first), at the very top of the page, sometimes at the bottom — includes the page number, the title of the periodical or section, etc. Necessarily attached to Issueline as the first, second, third, and subsequent pages. An empty Folioline is used for the first page. Materials on a given page are attached to that page's Folioline.	Upper portion of the page; usually features the nameplate or section (sport, culture etc.), the volume number and the date.
Footnote	Ordinary footnotes and references, very rarely found in reviews and on the very last page of the periodical, below editorial addresses — typographical codes (e.g. "ВФ12358") or serial numbers (e.g. "01568"), printing house addresses, etc. Always attached to Masthead.	—
Formula	Any formula that can be recognized by characters.	—
Formula number	Numbering in formula descriptions that will allow citing formula numbers rather than formulas in the text. Rarely found in periodicals unless in special supplements. Not used for tables (use Caption instead).	—
Front picture	Only a photo of the lead article or a photo on the front page that acts as a lead article. Other photos may not be used with this tag under any circumstances.	—
Front story	A leading article, which may be a photograph or graphic image and which reflects the main event of the current or previous day, appearing on the front page of the newspaper.	Splash story.
Graphics	Any graphic image that is not a photograph or cartoon. May also include text pieces.	Any graphic visual representation, incl. infographics, logos, etc.
Heading	The headline section of the periodical, which includes the name of the periodical, logo, slogans, in Soviet times also medals awarded to the periodical, and brief information about the issue: year, date, issue, price, etc.	Upper portion of the front page; usually features the nameplate, the volume number and the date.
Headline	Title of an article or material, but not a heading.	Word or group of words in large print that introduces an article.
Issueline	Release data necessary as metadata for corpus creation — year, date, issue number, serial number, date of first issue, price, etc. Part of the header, either as a line or as a rectangular box.	A part of the heading, as a box or line on the front page; usually features the volume number and the date.
Jump line	The line in an article or story indicating that it continues on the next page of the periodical. Used for "continued on page ..." and "beginning on page ...". For the second part of a story, always use the Optional Parent box ID pointing to the ID of the first part.	Where the article is cut and the continuation can be read on another page.
Kicker	"Proletarians of all countries, unite!" and all other slogans or similar pieces that appear above the headlines and provide clear context for them, e.g. "Sports", under which there is only one article.	Short text appearing above the headline that puts the article in context or highlights certain key points.
Lead	A small introductory or summarizing section of an article or material, a preface, which differs in pagination: it is italicized or bold, can be several columns wide, etc.	Short text at the beginning of an article that introduces it or summarizes its contents.
Letters to the editor	Letters to the editorial board, to the editor, and their responses.	Part of the newspaper where readers' opinions on topics of general interest are published.
Masthead	Editorial addresses, structure, phone numbers, which cannot be included in Issueline. In Soviet periodicals, as a rule, found at the bottom of the last page, along the entire length of the page. If on a different page than Issueline, add Issueline ID in the Optional Parent box ID.	Space that usually contains information about the newspaper such as its address, main contributors and subscription information.
Nameplate	The name of the periodical only: Banvor, Avangard, etc.	Title of the newspaper presented in a specific graphic style.
Obituaries	Exclusively death notices, which are given by individuals, collectives, etc.; can also take the form of articles or reviews. Excludes announcements of birth, marriage, acquaintance, services, sales, rentals, and other similar announcements (→ Ads).	Listing of death notices and anniversaries of deaths, cards of thanks and remembrances.
Page number	Page numbers, usually included in Folioline. Not typically used as a standalone label.	—
Photo	Only photographs, except for the lead photo or group of photos, which is labeled Front picture.	—
Photo credit line	Indicates the photographer or source of the photo(s).	Mandatory mention of the individual holding the rights to the photograph used to illustrate an article or a publication.
Rule	Open solid lines, with varying designs, that separate articles or materials in a periodical. Excludes rectangles or squares that delineate the entire article.	Line of varying thickness used to separate columns, articles and different graphic elements.
SealFigure	Seals placed when registering periodicals, or any other marking that is stamped.	—
Shorts	As a subtype of headlines, which combines small telegraphic messages under headings like "Short News", "The World Today", "The World in Orbit" and similar. Also a small message headline that simply solves a pagination problem and the following text is so short that the headline itself does not make sense to consider a Headline.	Titles of very short, informative texts, not headlines.
Sourceline	Information about the source of brief telegraphic messages or articles reprinted from news agencies — the name of the news agency or source: Armenpress, TASS, APA, etc. Sometimes attached to a Banner if the source of the Shorts attached to it is the same.	Name of the news agencies.
Subhead	A subtitle that is within the text and divides the text, not immediately following the title (that would be Deck). Text segments that follow the subtitle are attached directly to the subtitle.	Secondary title that separates and introduces the various parts of an article.
Supplement	A supplement that sharply differs from the main content of the given daily newspaper and is clearly perceived as a supplement, cultural, etc.	Separate publication that is inserted into a newspaper regularly or from time to time; it is printed in color and often on glossy paper.
Table	Tables not annotated as graphs, or pieces with a clear tabular structure that are possible or meaningful to identify (e.g. quotes in economic newspapers or economic indicators). Also includes crossword puzzles (chessboards are Graphics).	—
Television program schedule	Information about television or radio programs. Also includes theater and concert announcements. Excludes horoscopes, weather forecasts, announcements of sports matches, and entrance exam announcements (→ Ads).	—
Text	The article itself, the material, the text that follows the headline or title or subtitle, or a stand-alone untitled piece of text.	Body of the article; stand-alone text forming a whole; usually presents information, explanation or commentary.

Label hierarchy

Some labels have mandatory parent–child relationships:

Relationship	Description
Issueline → Folioline (at least ×2)	Each issue has one Issueline; at least two Folioline entries (one per page, empty for page 1) are attached to it. All article regions on a page are attached to that page's Folioline.
Issueline → Editorline	Editor-in-chief information is attached to the Issueline.
Issueline → Masthead	Masthead (editorial contact block) is attached to the Issueline, even when it appears on a different page.
Masthead → Footnote	Publishing notes at the bottom of the last page are attached to the Masthead if present.
Jump line → Jump line	The following part of a split article uses the Optional Parent box ID pointing to the ID of the first part.

Full Source Table

Source ID	Newspaper	Year
Araqs_1968_22	Արաքս	1968
AraqsM_1965_41	Արաքս	1965
AraqsM_1969_16	Արաքս	1969
AraqsM_1974_79	Արաքս	1974
AraqsM_1979_21	Արաքս	1979
AraqsM_1981_122	Արաքս	1981
Ashxatanqayin_uxi_1964_3	Աշխատանքային ուղի	1964
Ashxatanqayin_uxi_1966_17	Աշխատանքային ուղի	1966
Ashxatanqayin_uxi_1966_28	Աշխատանքային ուղի	1966
Avangard_1962_11	Ավանգարդ	1962
Avangard_1974_14	Ավանգարդ	1974
Avangard_1978_25	Ավանգարդ	1978
Avangard_1980_141	Ավանգարդ	1980
Avangard_1981_21	Ավանգարդ	1981
BanvorLenin_1966_202	Բանվոր	1966
BanvorLenin_1971_18	Բանվոր	1971
BanvorLenin_1976_29	Բանվոր	1976
BanvorLenin_1981_300	Բանվոր	1981
BanvorLenin_1982_302	Բանվոր	1982
ErekoyanErevan_1964_156	Երեկոյան Երևան	1964
ErekoyanErevan_1966_158	Երեկոյան Երևան	1966
ErekoyanErevan_1977_48	Երեկոյան Երևան	1977
ErekoyanErevan_1978_134	Երեկոյան Երևան	1978
ErekoyanErevan_1980_282	Երեկոյան Երևան	1980
ErekoyanErevan_1982_136	Երեկոյան Երևան	1982
Film_1968_14	Ֆիլմ	1968
Film_1976_23	Ֆիլմ	1976
Film_1976_34	Ֆիլմ	1976
Film_1984_12	Ֆիլմ	1984
Film_1984_13	Ֆիլմ	1984
Grakantert_1983_43	Գրական թերթ	1983
HayastaniFizkulturnik_1962_29	Հայաստանի Ֆիզկուլտուրնիկ	1962
HayastaniFizkulturnik_1974_21	Հայաստանի Ֆիզկուլտուրնիկ	1974
HayastaniFizkulturnik_1977_120	Հայաստանի Ֆիզկուլտուրնիկ	1977
HayastaniFizkulturnik_1980_106	Հայաստանի Ֆիզկուլտուրնիկ	1980
HayastaniFizkulturnik_1985_90	Հայաստանի Ֆիզկուլտուրնիկ	1985
Lori_1978_125	Լոռի	1978
Paros_1962_34	Փարոս	1962
Pionerkanch_1962_98	Պիոներ կանչ	1962
Pionerkanch_1967_99	Պիոներ կանչ	1967
Pionerkanch_1972_67	Պիոներ կանչ	1972
Qajaran_1978_5	Քաջարան	1978
Qajaran_1978_21	Քաջարան	1978
Qajaran_1978_24	Քաջարան	1978
Qajaran_1978_32	Քաջարան	1978
Qajaran_1979_10	Քաջարան	1979
Qajaran_1979_40	Քաջարան	1979
Qajaran_1979_45	Քաջարան	1979
SovetHayastan_1968_204	Սովետական Հայաստան	1968
SovetHayastan_1969_268	Սովետական Հայաստան	1969
SovetHayastan_1972_218	Սովետական Հայաստան	1972
Vardenis_1974_63	Վարդենիս	1974
Vardenis_1974_109	Վարդենիս	1974
Vardenis_1979_134	Վարդենիս	1979
Vardenis_1980_100	Վարդենիս	1980
Vardenis_1985_46	Վարդենիս	1985

Intended Use

This corpus is intended for:

• Computational linguistics research
• Historical and cultural analysis
• Language modeling and NLP tasks for Armenian
• OCR post-processing and text normalization research

OCR Evaluation Results

This section tracks the Character Error Rate (CER) across various baseline architectures and configuration filters on the ACoPPer test split. Note that for CER, lower values indicate better performance.

Document structure-independent models

Document structure-independent models output individual word boxes, which are spatially grouped into lines before scoring. These models return 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	17.78%	17.56%	17.38%	17.31%
tesseract	21.63%	21.40%	21.20%	21.13%
ocarm	5.84%	5.77%	5.65%	5.62%
abbyy	41.54%	41.37%	41.22%	41.17%
google-vision	7.55%	7.49%	7.35%	7.33%
portmind	9.76%	6.96%	9.61%	6.82%

Document structure-aware models

Document structure-aware models output one complete text line per box, scored directly against document structure-aware ground truth. These models return 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 word-level 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	35.45%	35.24%	35.03%	34.96%
tesseract	38.60%	38.38%	38.12%	38.06%
ocarm	5.81%	5.72%	5.58%	5.55%
abbyy	43.15%	42.98%	42.85%	42.80%
google-vision	43.44%	43.47%	43.97%	43.98%
portmind	27.69%	24.88%	27.61%	24.81%

OCArm detector + external recognizers

These variants use OCArm detection model's bounding boxes and recognize only the text cropped from within those specific regions.

Model	w/o filters	filter non-Armenian	filter graphics & headers	filter all
chandra	83.68%	83.77%	83.73%	83.76%
deepseek	96.59%	96.58%	96.16%	96.17%
qwen	82.19%	82.28%	81.85%	81.87%

Acknowledgements

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