--- license: apache-2.0 task_categories: - token-classification - feature-extraction language: - en tags: - markuplm - metadata - extraction - tokens - articles - html - books - conferences pretty_name: MarkupLM Metadata Extraction From Scientific Works --- # Dataset Card for Scientific Works Metadata Extraction ## 📖 Description Dataset for training models (including MarkupLM) on the task of metadata extraction from scientific papers in HTML format. Contains BIO tagging for 27 metadata types. **Task**: Token Classification for entity extraction: ```python ['title', 'author', 'date', 'doi', 'issn', 'eissn', 'journal', 'publisher', 'pages', 'first_page', 'last_page', 'language', 'volume', 'issue', 'abstract', 'affiliation', 'keyword', 'doc_type', 'isbn', 'eisbn', 'editor', 'orcid', 'book_version', 'subtitle', 'conference_title', 'book_series', 'book_title'] ``` ### 🎯 Key Features * Preprocessed splits: Train/Validation/Test * MarkupLM-compatible format * BIO tagging for tokens * Multi-domain data (16 sources) * Rich metadata structure: ```python { "tokens": ["", "", ...], "xpaths": ["/html[1]/body[1]", ...], "metadata": { "title": {"text": [...], "xpaths": [...]}, ... }, "node_labels": [0, 1, 2, ...] # BIO labels } ``` ### 📊 Statistics **Distribution by Source** | Resource | Articles | Books | Book Chapters | Conferences | |---------------|----------|-------|---------------|-------------| | ACS | 10000 | 81 | 1000 | - | | AIP | - | 117 | 1756 | - | | BMJ | 10000 | - | - | - | | CAIRN | - | 71 | - | - | | Duke | 9973 | 1000 | 1000 | - | | Emerald | 10000 | 4116 | 5000 | - | | IEEE | 10067 | 7229 | 10000 | 10000 | | IOP | 10000 | 933 | 1426 | - | | Karger | 10000 | 5931 | - | - | | Oxford | 10000 | - | - | - | | RSC | 9178 | 2187 | 10000 | - | | SAE | 10710 | 936 | - | - | | Sage | 1641 | 5716 | - | - | | ScienceDirect | 10000 | 10000 | - | - | | Cambridge | 13172 | 3160 | 9743 | - | | ACM | 5038 | 1502 | - | 10000 | ### 📜 Data Structure ```python Features({ "id": Value("string"), "source_file": Value("string"), # Source file path "resource": Value("string"), # Publisher (ACS, IEEE, etc.) "doc_type": Value("string"), # Document type "tokens": Sequence(Value("string")), "xpaths": Sequence(Value("string")), # XPath for each token "metadata": { # Target metadata fields field: { "text": Sequence(Value("string")), "xpaths": Sequence(Value("string")) } for field in expected_fields }, "node_labels": Sequence(Value("int64")), # BIO labels "processing_time": Value("string") }) ``` ### 📜 Label Map ```json { "id2label": { "0": "O", "1": "B-TITLE", "2": "I-TITLE", "3": "B-AUTHOR", "4": "I-AUTHOR", "5": "B-DATE", "6": "I-DATE", "7": "B-DOI", "8": "I-DOI", "9": "B-ISSN", "10": "I-ISSN", "11": "B-EISSN", "12": "I-EISSN", "13": "B-JOURNAL", "14": "I-JOURNAL", "15": "B-PUBLISHER", "16": "I-PUBLISHER", "17": "B-PAGES", "18": "I-PAGES", "19": "B-FIRST_PAGE", "20": "I-FIRST_PAGE", "21": "B-LAST_PAGE", "22": "I-LAST_PAGE", "23": "B-LANGUAGE", "24": "I-LANGUAGE", "25": "B-VOLUME", "26": "I-VOLUME", "27": "B-ISSUE", "28": "I-ISSUE", "29": "B-ABSTRACT", "30": "I-ABSTRACT", "31": "B-AFFILIATION", "32": "I-AFFILIATION", "33": "B-KEYWORD", "34": "I-KEYWORD", "35": "B-DOC_TYPE", "36": "I-DOC_TYPE", "37": "B-ISBN", "38": "I-ISBN", "39": "B-EISBN", "40": "I-EISBN", "41": "B-EDITOR", "42": "I-EDITOR", "43": "B-ORCID", "44": "I-ORCID", "45": "B-BOOK_VERSION", "46": "I-BOOK_VERSION", "47": "B-SUBTITLE", "48": "I-SUBTITLE", "49": "B-CONFERENCE_TITLE", "50": "I-CONFERENCE_TITLE", "51": "B-BOOK_SERIES", "52": "I-BOOK_SERIES", "53": "B-BOOK_TITLE", "54": "I-BOOK_TITLE" } } ``` ### 🏗️ Dataset Creation Dataset generated using a [custom tool](https://github.com/endthesame/markuplm-dataset-creator).