Dataset Card for BotanicalNER

Dataset Summary

BotanicalNER is a German-English parallel dataset for Named Entity Recognition (NER) of scientific and vernacular plant names. The resources were created for the master thesis project "Grasping the Nettle" at the University of Zurich in 2019.

The main focus of the project was to identify and disambiguate scientific and vernacular plant names across multiple text genres to provide a valuable tool for extracting and preserving (ethno-)botanical knowledge. The dataset is structured into several sub-corpora from different domains, which are available as separate configurations:

• botlit: Botanical literature
• plantblog: Plant-themed blog posts
• wiki: Wikipedia abstracts
• textberg: The TextBerg corpus of Alpine Club yearbooks
• gold: Gold-standard test sets
• fungi: A specialized test set for fungi names

Supported Tasks and Leaderboards

• Tasks: Named Entity Recognition, Token Classification
• Leaderboards: N/A

Languages

The dataset contains texts in German (de) and English (en).

Dataset Structure

The dataset is composed of multiple configurations, one for each sub-corpus and language (e.g., botlit_de, wiki_en, gold_de).

Data Instances

An example from the gold_de configuration:

{
  "id": "0",
  "tokens": ["Die", "Brennnessel", "(", "Urtica", "dioica", ")", "ist", "eine", "wichtige", "Heilpflanze", "."],
  "pos_tags": ["ART", "NN", "$(", "NE", "NE", "$)", "VAFIN", "ART", "ADJA", "NN", "$."],
  "ner_tags": [0, 3, 0, 1, 2, 0, 0, 0, 0, 0, 0]
}

Data Fields

All configurations share the same data fields:

• id: A unique identifier for the example, a string feature.
• tokens: The list of tokens in the sentence, a Sequence of string features.
• pos_tags: The list of part-of-speech tags, a Sequence of string features.
• ner_tags: The list of NER tags, a Sequence of ClassLabel features. The mapping from ID to tag is as follows:

{
    "0": "O",
    "1": "B-Scientific",
    "2": "I-Scientific",
    "3": "B-Vernacular",
    "4": "I-Vernacular"
}

Data Splits

The data is provided as several distinct corpora, loaded via configurations.

• The botlit, plantblog, textberg, and wiki configurations each contain a train split composed of silver-standard data.
• The gold and fungi configurations each contain a test split composed of gold-standard data.

Dataset Creation

Curation Rationale

The project aimed to create a robust tool for extracting plant names from diverse texts, which is a crucial step for digitizing and preserving botanical and ethno-botanical knowledge.

Source Data

The data was collected from various sources, tokenized, and POS-tagged using TreeTagger.

Annotations

The dataset was annotated using a dictionary-based method (silver standard) and manual correction (gold standard). The annotation scheme is IOB (Inside, Outside, Beginning) for two entity types: Scientific and Vernacular plant names.

Personal and Sensitive Information

The dataset does not contain personal or sensitive information as it focuses on botanical and scientific content.

Considerations for Using the Data

Social Impact of Dataset

This dataset can have a positive social impact by enabling better extraction of botanical knowledge, supporting biodiversity research, and assisting in the preservation of ethnobotanical knowledge.

Discussion of Biases

Potential biases include geographic bias in plant names, source bias from different text genres, and domain bias from the specific sources used. Most training data also uses automatic annotation which may contain errors.

Additional Information

Dataset Curators

Isabel Meraner, Institute of Computational Linguistics, University of Zurich (2019).

Licensing Information

The dataset is licensed under the GNU General Public License v3.0 (GPL-3.0).

Citation Information

@mastersthesis{meraner2019grasping,
    title={Grasping the Nettle: Neural Entity Recognition for Scientific and Vernacular Plant Names},
    author={Meraner, Isabel},
    year={2019},
    school={Institute of Computational Linguistics, University of Zurich}
}