Processed patent corpus by sector

Dataset overview This dataset provides processed patent corpora for five technological industries: - Automotive- Camera- Mobile phone- Robotics- Semiconductor Each industry folder contains two CSV files: - `df_id_year_document.csv`: the processed patent corpus- `df_company_id_year.csv`: the corresponding company–patent relationships What the dataset contains Processed patent corpus For each patent, `df_id_year_document.csv` contains: - `id`: patent identifier- `year`: publication year- `document`: cleaned English-language text compiled from the patent title, abstract, and selected elements of the claims, prepared for embedding and topic-modelling applications- `topic`: the topic assignment produced by the corpus-wide preprocessing and filtering model The deposited `topic` field represents the preliminary topic assignment used during patent-corpus filtering. It should not be interpreted as a final moving-window or cross-window topic assignment. A value of `-1` denotes a patent treated as noise or unassigned by that preliminary model. Patent-level embeddings are not included in the deposited files. They can be reconstructed from the `document` field using the accompanying code. Similarly, the moving-window topic models, cross-window topic links, transition indicators, and other downstream measures must be generated through the analytical pipeline. Company–patent relationships For each industry, `df_company_id_year.csv` contains: - `company`: normalized company or assignee name- `id`: patent identifier- `year`: publication year Because a patent may be associated with more than one company, the company table can contain multiple records for the same patent. It can be joined to the processed patent corpus using `id` and, where necessary, `year`. What the dataset enables The dataset provides a common processed input for researchers and practitioners who want to: - construct semantic representations of industry-specific patent corpora;- build and inspect fine-grained technological topic structures;- estimate topic models over overlapping historical time windows;- connect related topics across successive windows;- trace the evolution and reorientation of technological activity;- identify transition-driving topics and the patents associated with them;- compare technological trajectories across industries; and- conduct company-level analyses by linking patent topics and transition measures to assignees, including portfolio positioning, proximity to transition-driving patents, incumbent–entrant comparisons, and network-based analyses. The deposited documents and company links provide the empirical inputs for these applications. Embeddings, dynamic topics, transition measures, patent-level predictivity measures, and company-level indicators are derived outputs that must be calculated using the accompanying analytical workflow. Reproducibility materials A companion GitHub repository provides the Jupyter notebooks and utility scripts required to reproduce and extend the analytical pipeline: [https://github.com/mingchennnn/TechSeis](https://github.com/mingchennnn/TechSeis) The repository covers: - patent-data preparation and filtering;- document embedding;- corpus-wide BERTopic modelling;- topic-level relevance evaluation and patent filtering;- moving-window topic modelling;- dynamic topic linking;- technological-transition and predictivity analyses; and- downstream patent- and company-level analyses. Users should generate embeddings and derived analytical outputs from the deposited `document` field rather than treating the preliminary `topic` field as the final topic structure. Data-format notes - The files are provided in CSV format.- Some `year` and `topic` values may be serialized as decimal-form numbers, such as `1988.0` or `-1.0`; these fields can be converted to integers after loading.- Company names are normalized identifiers and may contain underscores.- The patent tables and company tables have a one-to-many relationship because one patent may be linked to multiple companies. Intended use and citation This release is intended as a public benchmark and starting point for reproducing and extending technology-seismograph methods across technological domains. If you use the dataset or accompanying code, please cite: - the [Zenodo dataset record](https://doi.org/10.5281/zenodo.22005545); and- the [TechSeis GitHub repository](https://github.com/mingchennnn/TechSeis).