atheer2104/swedish-patent-cpc-group

Patent CPC Group Multi-Label Classification Dataset

This dataset contains historical Swedish patent documents (1885-1972) classified according to the Cooperative Patent Classification (CPC) system at the group level. Each document can have multiple labels, making this a challenging multi-label classification task with significant class imbalance and data sparsity characteristics.

Dataset Description

The dataset includes patent claims text extracted from digitally recreated versions of historical Swedish patents, generated using Optical Character Recognition (OCR) from original paper documents. The text quality varies due to OCR limitations, but all CPC labels were manually assigned by patent engineers at PRV (Swedish Patent and Registration Office), ensuring high reliability for machine learning applications.

Background and Motivation

Historical Swedish patents from 1885-1972 were originally classified using the German Patent Classification (DPK) system. When Sweden adopted the International Patent Classification (IPC) system in 1972, approximately 20% of documents were manually reclassified. Later, when the Cooperative Patent Classification (CPC) system was introduced, most IPC labels were automatically converted to CPC, with manual classification when automatic conversion wasn't possible.

This dataset addresses the challenge of making historical patents more accessible by providing fine-grained CPC group classifications that enable precise patent search and retrieval systems to effectively index and retrieve older Swedish patents.

Group Classification

• Examples: 27,339 documents (after filtering)
• Unique labels: 1,781 CPC group labels
• Single-label documents: 16,110 (58.9%)
• Multi-label documents: 11,229 (41.1%)
• Average labels per document: 1.65
• Max labels per document: 15
• Average documents per label: 25.37
• Task: Multi-label classification of patent claims text to CPC group level

Data Characteristics

The dataset exhibits extreme class imbalance and data sparsity typical of fine-grained patent classification:

• The average number of documents per label (25.37) is approximately 4 times lower than at the subclass level
• More documents have multiple labels compared to subclass classification (41.1% vs 34.7%)
• Long-tailed distribution with many labels having only the minimum threshold of 10 documents
• Higher classification complexity due to finer granularity of CPC group categories

Label Distribution Breakdown:

• 1 label: 16,111 documents (58.9%)
• 2 labels: 7,083 documents (25.9%)
• 3 labels: 2,658 documents (9.7%)
• 4 labels: 935 documents (3.4%)
• 5+ labels: 552 documents (2.0%, decreasing exponentially to max of 15)

Classification Hierarchy Context

CPC group classification represents a finer level of granularity than subclass classification:

• Section → Class → Subclass → Main Group → Subgroup
• This dataset focuses on the group level (combining main groups and subgroups)
• Higher precision requirements but increased difficulty due to:
  • More subtle distinctions between categories
  • Fewer training examples per label
  • Greater label vocabulary size (1,781 vs 497 at subclass level)

Data Fields

• text: Patent claims text (Swedish) extracted from OCR-processed historical documents
• labels: List of CPC group classification codes (multi-label format)
• dokID: Unique document identifier
• dokNr: Patent document number
• dokKod: Document code classification
• full_cpc_codes: Complete list of all CPC codes assigned to the document

Data Processing

Text Extraction

Patent claims were automatically extracted using spaCy with the "sv_core_news_sm" pipeline for Swedish text processing. Claims were identified by:

• Start markers: "patentanspråk" or "patentkrav"
• End markers: "härtill ritning" or "kongl boktryckeriet"

Label Preprocessing

• CPC labels related to index codes and section Y were removed (providing only supplementary information)
• Labels with fewer than 10 associated documents were filtered out to ensure adequate training data
• CPC group level labels include both main groups and subgroups for comprehensive coverage

Quality Considerations

• Original documents: 269,052 total patent documents
• Used documents: 27,339 (after filtering for quality and label requirements)
• Text quality varies due to OCR processing of historical documents
• All CPC labels manually verified by patent engineering professionals

Usage

from datasets import load_dataset

# Load group classification dataset
dataset = load_dataset("atheer2104/swedish-patent-cpc-group")

# Example of accessing multi-label data
example = dataset['train'][0]
text = example['text']
labels = example['labels']  # List of CPC group codes

Multi-Label Classification Challenges

This dataset presents significant challenges that make it particularly demanding for machine learning models:

1. Extreme Class Imbalance: Highly skewed label distribution with some labels having hundreds of examples while many have only 10-20 examples
2. Data Sparsity: Average of only 25.37 documents per label makes learning difficult for rare categories
3. Fine-grained Classification: Subtle distinctions between closely related patent groups require deep domain understanding
4. Multi-label Complexity: 41.1% of documents have multiple labels with complex co-occurrence patterns
5. Historical Text: OCR artifacts and historical Swedish language variations impact text quality
6. Large Label Vocabulary: 1,781 possible labels create a high-dimensional prediction space

Evaluation Metrics

For this challenging multi-label classification task, comprehensive evaluation is essential:

Primary Metrics:

• Precision@k (P@k): Precision of top-k predictions (k = 1, 3, 5)
• Recall@k (R@k): Recall of top-k predictions (k = 1, 3, 5)
• F1@k: Harmonic mean of P@k and R@k (k = 1, 3, 5)
• nDCG@k: Normalized Discounted Cumulative Gain at rank k (k = 1, 3, 5)

Aggregate Metrics:

• Micro-F1: Overall F1 score across all labels (emphasizes frequent labels)
• Macro-F1: Average F1 score across all labels (treats all labels equally)

Recommended Focus: Given the extreme class imbalance, pay particular attention to macro-averaged metrics and performance on rare labels.

Dataset Splits

Use stratified sampling to maintain similar label distributions across splits:

• Train: 90%
• Test: 10%
• Validation: Use nested 90/10 split of training data for hyperparameter tuning

Important: Due to label sparsity, ensure that rare labels appear in both training and test sets.

Baseline Performance Context

Research from the source thesis indicates that group-level classification is significantly more challenging than subclass classification:

• Lower overall performance across all metrics
• Greater benefit from incorporating additional information sources
• Higher variance in performance across different label categories

Related Research

This dataset was created as part of research investigating whether incorporating legacy classification information (DPK) improves fine-grained CPC classification performance. The research demonstrated that:

• Group-level classification is more challenging than subclass classification
• Data sparsity and class imbalance significantly impact model performance
• Combining textual content with legacy classification data shows promise for improving results

Limitations

• Severe class imbalance: Many labels have minimal training examples
• OCR quality: Text quality varies significantly across documents
• Historical scope: Limited to Swedish patents from 1885-1972
• Language specificity: All text in Swedish, limiting broader applicability
• Filtering effects: Minimum threshold of 10 documents per label excludes very rare categories
• Fine-grained complexity: Subtle distinctions between groups may be inconsistently labeled

Recommended Approaches

Given the dataset characteristics, consider:

• Class balancing techniques: Oversampling, undersampling, or focal loss
• Hierarchical methods: Leverage CPC hierarchy for better predictions
• Few-shot learning: Handle rare labels with limited examples
• Ensemble methods: Combine multiple models for robust predictions
• Transfer learning: Leverage pre-trained Swedish language models

Citation

@mastersthesis{Salim2025PatentClassification,
   author = {Salim, Atheer},
   institution = {KTH, School of Electrical Engineering and Computer Science (EECS)},
   pages = {70},
   school = {KTH, School of Electrical Engineering and Computer Science (EECS)},
   title = {Machine Learning for Classifying Historical Swedish Patents: A Comparison of Textual and Combined Data Approaches},
   series = {TRITA-EECS-EX},
   number = {2025:571},
   keywords = {Multi-label Text Classification, Machine Learning, Patent Classification, Deep Learning, Natural Language Processing, Textklassificering med flera Klasser, Maskininlärning, Patentklassificering, Djupinlärning, Språkteknologi},
   abstract = {Patents are essential for protecting intellectual property and advancing innovation, but the accessibility of historical patents is often limited by outdated classification systems. The thesis investigates whether incorporating DPK information alongside patent claims improves the performance of machine learning models in classifying historical Swedish patents into the CPC system. While group-level predictions were more challenging due to label imbalance and fewer samples per label, the research provides valuable insights into fine-grained patent classification challenges.},
   year = {2025},
   url = "https://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-368254"
}