glossAPI/elocus

E-Locus

Dataset Language:
Greek (primary), with English alternative titles, abstracts and bibliographic terms throughout, and occasional French or German material in older volumes. The content column carries the full Greek text of each thesis; the alt_title and most of the subjects field are English by convention of the source repository.

Dataset Info:
This dataset consists of text-extracted PDFs from E-Locus (elocus.lib.uoc.gr), the Institutional Repository of the Library and Information Center of the University of Crete. E-Locus was launched in 1998 and went public in its current Keystone DLS form on 2 July 2009. Its mission is the unified collection, preservation and dissemination of the scientific work produced at the University of Crete. The material is organised into collections and currently consists of undergraduate theses, postgraduate theses and doctoral dissertations completed at the University of Crete, alongside departmental study guides and technical reports. Each record in this dataset corresponds to one such item: one author, one defense / publication event, one source PDF.

Metadata Info:
This dataset consists of 7,731 records, structured as a flat tabular dataset with thirteen string-typed columns. Each row is uniquely identified by either url (the landing page on E-Locus) or pdf_url (the direct PDF link); both are unique by construction. The collection column carries a hierarchical breadcrumb that encodes both the School / Department / Programme of origin and the type of work, in the form Σχολή/Τμήμα--...--... | Τύπος Εργασίας--.... The dataset preserves bibliographic and physical-description metadata from the source MARC-style record alongside the full extracted text of the attached PDF.

Metadata Field Descriptions:

Field	Type	Description
title	string	The Greek title of the thesis as recorded in E-Locus. Always populated. Length ranges from 8 to 425 characters (median ≈ 99).
content	string	Full text extraction of the attached PDF, preserving paragraph breaks via newline characters. Length ranges from 336 characters (a conference proceedings cover) to ≈ 4.67 million characters; median ≈ 190 K. Tables detected by the PDF parser are emitted as <table>/<tr>/<th> markup with HTML entities (&gt;, &lt;) - this is intentional structured output, not corruption.
abstract	string	The thesis abstract as supplied to the repository by the depositor. May be empty (≈ 36.5 % of records - 2,824 / 7,731 - have an empty abstract at the source). When present, length ranges from a few hundred characters up to ≈ 25 K.
author	string	Name of the thesis author, in the form Surname, Given Name(s) (Greek script). Always populated for 99.6 % of rows; 31 records carry an empty author field at the source.
advisor	string	Name of the thesis supervisor / advisor. Often populated, but 1,418 records (≈ 18.3 %) carry an empty value at the source (older or pre-policy records).
committee_members	string	Semicolon-separated list of the rest of the examination / supervisory committee, e.g. Surname1, Given1; Surname2, Given2. Often empty at the source (4,610 records ≈ 59.6 %).
subjects	string	Subject / keyword string in the form term1; term2; …, typically bilingual (English and Greek terms interleaved). May be empty (3,006 records ≈ 38.9 %). Maps to Dublin Core dc:subject at the source.
alt_title	string	Alternative title - almost always the English translation of the Greek title. May be empty (3,237 records ≈ 41.9 %).
date	string	Date of the defense / publication, stored as it appears at the source. Most values are full YYYY-MM-DD strings; some are bare YYYY (year-only). Date range spans 1988-07-01 → 2026-03-30. Maps to Dublin Core dc:date.
collection	string	Hierarchical breadcrumb encoding (a) the School / Department / Programme that issued the thesis and (b) the type of work, in the form Σχολή/Τμήμα--<School>--<Department>--<Programme> | Τύπος Εργασίας--<Type>. 103 distinct collection values across the dataset. The most-populous collections are Medical School theses, Department of Psychology programmes, and various Faculty of Science and Engineering postgraduate programmes.
physical_description	string	MARC-style physical description (page count, illustrations, dimensions) - e.g. 49 σ. : χάρτ., πίν., έγχ. εικ. ; 30 εκ.. Often empty (1,207 records).
url	string	Direct HTTPS link to the E-Locus metadata landing page for the record, e.g. https://elocus.lib.uoc.gr/dlib/8/8/7/metadata-dlib-1772878887-83781-18163.tkl. Always populated, always under elocus.lib.uoc.gr, and unique per row.
pdf_url	string	Direct HTTPS link to the attached PDF file, e.g. https://elocus.lib.uoc.gr//dlib/8/8/7/attached-metadata-dlib-1772878887-83781-18163/EYAGGELIA_GARA_2026_METAPTYXIAKH_ERGASIA.pdf. Always populated, always under elocus.lib.uoc.gr, and unique per row. The double slash //dlib after the hostname is the form produced by the source repository and is preserved verbatim.

Dataset Statistics:
Records: 7,731

Words: 279,726,210

Tokens: 568,723,589
(WordPiece tokens from nlpaueb/bert-base-greek-uncased-v1, computed on title + abstract + content combined. The three columns split as: title 162,107 tokens, abstract 2,851,103 tokens, content 565,710,379 tokens - i.e. content carries > 99.4 % of the corpus.)

Size: 1.34 GB (parquet on disk) - ≈ 2.04 GB of plain text once decoded.

Known Data Quality Notes:
The source documents are PDFs deposited at E-Locus over a span of nearly forty years (1988 – present), and the content column carries the text extraction of those PDFs. Some level of extraction noise is therefore expected and inherent to the source material. Born-digital theses produced with modern LaTeX or Word tooling extract cleanly, but older scanned or low-quality PDFs do not. Our policy is to publish the extracted text as faithfully as possible to the underlying PDFs rather than aggressively post-process it: this preserves reproducibility and lets downstream users apply their own filtering, re-OCR or normalization strategy as appropriate to their task.

• Empty bibliographic fields at the source. Several metadata columns carry empty strings (not nulls) where the depositor did not supply that field - most notably committee_members (≈ 59.6 % empty), alt_title (≈ 41.9 %), subjects (≈ 38.9 %), abstract (≈ 36.5 %), advisor (≈ 18.3 %), physical_description (≈ 15.6 %), and author (31 records). These reflect repository data as supplied; we do not synthesise placeholders.
• Mixed date granularity. Most dates are full YYYY-MM-DD; some are bare YYYY. Both forms are preserved verbatim, parseable with pandas.to_datetime(..., errors='coerce').
• Generated HTML markup inside content. The PDF-to-text extractor emits <table>/<tr>/<th> tags with HTML entities (>, <,  ) when it detects a table on the page. This is structured output, not corruption: it appears in 6,133 of 7,731 records and the embedded HTML entities account for the ≈ 1,520 "encoding artifact" hits in the QA scan. Strip or parse it depending on whether your downstream pipeline wants structured tables.
• Trailing whitespace. 710 of 7,731 content values end in trailing whitespace (rogue newlines from the extractor) and 3 collection values carry trailing whitespace. Easy to .strip() if desired, but left as-is in the published file to preserve byte-for-byte fidelity to the extraction step.
• Unicode replacement characters. 59 content rows contain \ufffd (the replacement character), typically clustered in a few hard-to-OCR pages of older scanned PDFs. These are PDF-extraction defects at the source and are not removed.
• Duplicate non-content cells are expected. Bibliographic fields like advisor, committee_members, collection, date, subjects and physical_description legitimately repeat across rows - e.g. one advisor supervised 75 theses, the most-populous collection has 827 rows, and the most common physical description (92 σ. ; 30 εκ.) appears on 31 records. No content, url or pdf_url value is duplicated.
• Mojibake-shaped substrings inside content. The QA scan flags ≈ 1,520 rows for "encoding artifacts"; on inspection essentially all hits are HTML entities inside the extractor's emitted <table> markup, not actual UTF-8/Latin-1 corruption. Treat the figure as an upper bound, not as the count of corrupt records.
• Date coverage is not uniform. The repository is sparsely populated for early years (1 – 2 records / year in 1988 – 1993) and reaches its peak coverage from 2007 onwards (150 – 295 records / year). The biggest coverage gap is 638 days between 1990-01-01 and 1991-10-01. Consider this when using date for temporal slicing.

License:
The dataset compilation, metadata schema, text-extraction pipeline and accompanying tooling are released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

The underlying texts (each individual thesis or dissertation) remain the intellectual property of their original authors, as established under Greek copyright law. E-Locus hosts these works under an open-access mandate of the University of Crete Library, but does not declare a machine-readable per-item licence. The dataset is therefore offered for non-commercial academic and research use only. Downstream users who wish to use any individual thesis commercially, or to redistribute it outside research contexts, must independently obtain permission from the thesis's author and/or the Library and Information Center of the University of Crete.

When using this dataset, please attribute both the original repository and this derivative compilation:

E-Locus Dataset, derived from the open-access institutional repository of the Library and Information Center of the University of Crete (elocus.lib.uoc.gr), licensed under CC BY-NC-SA 4.0.

Source:
E-Locus - Institutional Repository of the Library and Information Center of the University of Crete (Ιδρυματικό Καταθετήριο Βιβλιοθήκης Πανεπιστημίου Κρήτης), elocus.lib.uoc.gr. Registered in OpenDOAR (#1464) and ROAR (#4552).