--- language: - sv task_categories: - token-classification dataset_info: features: - name: text dtype: string - name: entity_types sequence: string - name: entity_start_chars sequence: int64 - name: entity_end_chars sequence: int64 - name: id dtype: int64 - name: word_start_chars sequence: int64 - name: word_end_chars sequence: int64 splits: - name: train num_bytes: 10734126 num_examples: 677 - name: eval num_bytes: 2264347 num_examples: 170 download_size: 5163846 dataset_size: 12998473 configs: - config_name: default data_files: - split: train path: data/train-* - split: eval path: data/eval-* --- # Swener-1800 Dataset for nested named entity recognition in historical Swedish. ## Dataset Details ### Dataset Description This is a unique dataset for nested named entity recognition in historical Swedish. The texts and entity types were selected with the aid of a group of historians and researches within the humanities, and the annotation was done by a group of domain experts. The selection of texts range from historical newspapers, court-records, letters, fiction and police reports, ensuring a diverse range of styles and subjects. The year of origin for the selected texts range from 1730-1900, making it ideal for evaluating NER-systems, and underlying language models, on their ability to represent historical Swedish. - **Curated by:** Språkbanken, Uppsala University and The Swedish National Archives - **Funded by:** Språkbanken - **Language(s) (NLP):** Swedish (historical) - **License:** [More Information Needed] ### Dataset Sources [optional] - **Repository:** [More Information Needed] - **Paper:** https://sweclarin.se/sites/sweclarin.se/files/SCR-01-2024.pdf ## Uses ### Direct Use The dataset is intended for training specialized nested NER systems on historical Swedish texts, but as these systems are usually based on pre-trained LLM:s, an equally important intended use for this dataset is to evaluate pre-trained LLM:s on their ability to represent historical Swedish, which has important downstream applications for language technology on historical text. ### Out-of-Scope Use The dataset is not intended for training NER-systems on modern Swedish, or other languages than Swedish ## Dataset Structure The dataset contains the following fields: - `text`: The original sentence or text passage. - `entity_types`: A list of entity types corresponding to the named entities found in the text. - `entity_start_chars`: A list of start character positions for each entity in the `text`. - `entity_end_chars`: A list of end character positions for each entity in the `text`. - `id`: A unique identifier for each sample. Each entry in the dataset captures a portion of text along with the annotated entities and their positions. ### Entity Types Below is a description of the entity types found in this dataset: - **EVN**: The event category refers to specific events, gatherings, or incidents. This includes a wide range of activities and events, such as historical events, sporting events, concerts, conferences, festivals, natural disasters, political elections, and ceremonies. - **LOC**: The location category includes geographical locations of all kinds, real or fictional, large or small: continents, countries, regions, cities, villages, areas, parks, streets, mountains, rivers, and so on. - **MSR-AREA**: The area category contains measurements used to describe the size of a surface or area, such as square kilometers, hectares, square meters, or other area units. - **MSR-DIST**: The distance category includes measurements describing the distance between two points, such as kilometers, meters, miles, or other units used to express distance. - **MSR-LEN**: The length category contains measurements used to describe the size of something in a single dimension, such as the length of an object or the height of a building, often given in meters, centimeters, feet, inches, etc. - **MSR-MON**: The money category includes all things related to money and monetary values, including various currencies such as kronor, öre, or dollars, as well as monetary expressions like interest rates, prices, and valuations. - **MSR-OTH**: This category includes other types of measurements that do not fall under specific measurement categories. - **MSR-VOL**: The volume category refers to measurements that describe how much space an object or substance occupies, such as liquid volume in liters or milliliters, or space volume in cubic meters or cubic feet. - **MSR-WEI**: The weight category includes measurements used to determine how heavy something is. Weight can be given in kilograms, grams, pounds, ounces, and similar units, used to describe the weight of objects, from small items to larger structures. - **OCC**: The occupation category includes specific names of occupations or professions, encompassing all types of job roles and services individuals can hold, such as 'doctor', 'teacher', or 'carpenter'. - **ORG-COMP**: The company category includes all types of commercial enterprises, from small sole proprietorships to large multinational companies. - **ORG-INST**: The institution category includes various types of formal organizations or established societal institutions. It includes, but is not limited to, educational institutions (such as schools and universities), public authorities, religious organizations, research institutes, and non-governmental organizations. - **ORG-OTH**: Other organizations that do not fall under specific categories. - **PER**: The persons category includes real and fictional individuals, gods, and mythological figures, either as individuals or groups, but not animals or other creatures. - **TME-DATE**: The date category encompasses specific calendar dates, including days, months, and years. It can refer to both historical and future dates. - **TME-INTRV**: The interval category refers to a period of time with a defined start and end point. It can include short periods like 'from 9 to 17' or longer periods like 'January to March' or '2020 to 2024'. It also includes more undefined periods like 'during the 19th century' or 'during World War II'. - **TME-TIME**: The time expression category includes broader and more general time-related expressions that do not necessarily refer to specific dates or intervals. It can include expressions like 'the day before yesterday', 'next week', 'in the morning', 'in an hour', 'during the summer', or specific times of day. - **WRK**: The work of art category includes names or titles referring to works of art, such as books, plays, brand names of commercial products, newspapers, and magazines. - **SYMP**: The symptom category includes specific medical or health-related conditions experienced by an individual. It includes physical, emotional, or cognitive signs that indicate illness, injury, or other health conditions. --- ### Usage Users can load the dataset using the `datasets` library as follows: ```python from datasets import load_dataset dataset = load_dataset("Riksarkivet/swener_1800") ``` ## Dataset Creation ### Curation Rationale AI for modern languages is well researched, and off-the-shelf-models for standard NLP-tasks are readily available. For historical language the situation is a bit different, but the potential for NLP to aid in historical research and research within the humanities is huge. This dataset will serve, both as a training dataset for training NER-systems for historical Swedish, an important task on it's own, but also, and perhaps more importantly as a benchmark dataset for evaluating large language models' ability to represent historical Swedish, which will further research within, and implementation of, NLP for historical lanuguage. ### Source Data 19th century Swedish newspapers 19th century Finish court-records (in Swedish) 18th century letters to Swedish authorities (suppliker) 19th century Swedish short-stories 19th century Sweish police-reports #### Data Collection and Processing The texts an annotated entity categories were selected in discussion with an expert group of historians and researchers within the humanities #### Annotation process The annotation process was iterative where annotation guidelines step by step materialized in the form of a complete annotation manual. Systematic double-annotations were done, and inter-annotator agreement were calculated at several steps in the annotation process. For more information on the annotation process, see the paper: [SWENER-1800: A Corpus for Named Entity Recognition in 19th Century Swedish](https://sweclarin.se/sites/sweclarin.se/files/SCR-01-2024.pdf) #### Who are the annotators? An expert group of annotators funded by Språkbanken ## Bias, Risks, and Limitations The entities are not balanced, some entities are very common, while others are very rare. ## Citation [SWENER-1800: A Corpus for Named Entity Recognition in 19th Century Swedish](https://sweclarin.se/sites/sweclarin.se/files/SCR-01-2024.pdf) **BibTeX:** [More Information Needed] ## Dataset Card Authors [optional] Erik Lenas ## Dataset Card Contact erik.lenas@riksarkivet.se