How Short Are Patent Titles and Abstracts? Evidence from European Patent Applications

Description

The present dataset encompasses both derived data and reproducibility materials, providing a comprehensive foundation for a large-scale analysis of the length of English-language European patent titles and abstracts. The patent-title data were obtained from the European Patent Office (EPO) PATSTAT database through the Technology Intelligence Platform (TIP), while the abstract data were obtained from the EPO EPAB database through TIP. The two corpora were analyzed independently. The primary title dataset contains 4,686,155 valid English-language EP titles after deduplication at application level. The mean title length is 8.06 words, the median is 7 words, and approximately 75.0% of titles contain no more than 10 words. The primary abstract dataset contains 2,155,267 valid English-language abstracts, which were selected deterministically to ensure a single abstract per EP application. The mean abstract length is 125.20 words, and the median is 122 words. A survey of abstracts reveals that approximately 72.36% of them contain no more than 150 words. The modal abstract length is precisely 150 words (24,232 observations), succeeded by 14,768 abstracts at 151 words, corresponding to a 39.06% relative decrease immediately above the 150-word threshold. The repository contains comprehensive observed word-count frequency distributions, cumulative threshold statistics, annual publication-level descriptive series, application-versus-publication scope checks, language counts, outlier sensitivity analyses, tokenizer validation examples, derived figures, Python code, and supplementary methodological documentation. The raw EPO patent-title and abstract corpus is not redistributed; the deposited files contain aggregate or derived research data only. The extraction encompasses the period from 1978 to 2026. It should be noted that the data for 2026 represent a partial publication year at the time of extraction. The package facilitates verification of the primary findings in the absence of access to the TIP database. Additionally, it provides scripts for users with access to the EPO TIP environment. The OpenAI Codex was utilized as a programming assistant to develop, test, and debug Python workflows for EPO TIP. The research data that has been deposited is derived from EPO data and is not AI-generated.

Steps to reproduce

1. Download the complete dataset and retain the directory structure described in README.md. 2. To verify the main numerical results without EPO TIP access, install Python 3 and pandas and run: python code/reproduce_key_results_from_derived_csv.py from the root directory of the dataset. 3. The script parses the complete application-level title and abstract word-count frequency distributions and reproduces the main findings: mean title length, percentage of titles containing no more than 10 words, mean abstract length, percentage of abstracts containing no more than 150 words, the 150-word mode, and the relative frequency decrease from 150 to 151 words. 4. To ensure the complete reproduction of source data, it is necessary to utilize the EPO Technology Intelligence Platform (TIP). The code "run code/ep_patstat_title_length_frequency_TIP_PROD.py" is intended for the extraction of PATSTAT titles, while the code "run code/epab_abstract_frequency_TIP_PROD_annual_fallback.py" is designed for the extraction of EPAB abstracts. The EPAB workflow first queries exact annual publication-date terms and falls back to monthly terms if required. 5. Both workflows normalize text by decoding HTML entities, removing HTML tags, normalizing Unicode dash characters, collapsing repeated whitespace, and applying the Unicode-aware token regular expression [^\\W_]+(?:[-'][^\\W_]+)*. Expressions that contain hyphens and apostrophes are to be counted as single words. 6. Abstract records are deduplicated by application.number, with the following deterministic priority: (a) a valid English-language abstract; (b) A1/A2 publication kind; (c) earliest publication date; and (d) publication number as the final tie-breaker. 7. Compare regenerated outputs with the CSV files in the data/ directory. Tokenizer validation examples are provided in the validation/subfolder, while extraction quality-control files are stored in qc/ subfolder. 8. The main cross-sectional title and abstract frequency distributions are application-level. The annual descriptive series are publication-level. Data for 2026 are incomplete and should not be interpreted as a complete annual cohort.

Institutions

• Politecnico di Milano

  Lombardy, Milan

Categories

Licence