Skip to content

Latest commit

 

History

History
172 lines (115 loc) · 6.48 KB

README.md

File metadata and controls

172 lines (115 loc) · 6.48 KB
Raw

Repository for Systematically Comparing Multilingual NER Tools

This repository contains the replication materials for the article "Automatically Finding Actors in Texts: A Performance Review of Multilingual Named Entity Recognition Tools." Communication Methods and Measures 2024. doi: https://doi.org/10.1080/19312458.2024.2324789. For the exact version used in the paper, please go to the branch with the corresponding DOI: https://github.com/mrwunderbar666/ner_tool_comparison/tree/10.1080/19312458.2024.2324789

The repository's main branch is intended to be updated with new corpora and NER tools. Contributions are welcome! You can either create an issue with your suggestion, or make a pull request.

NER Tools

  • CoreNLP
  • NLTK
  • ICEWS
  • JRC Names
  • mLUKE (new!)
  • Nametagger
  • OpenNLP
  • spaCy
  • XLM-RoBERTa (via Huggingface)

Data

The datasets for evaluation are the following:

  • CoNLL 2002 (Dutch & Spanish)
  • CoNLL 2003 (English & German*)
  • Europeana (German, French, Dutch)
  • GermEval2014 (German)
  • WNUT Emerging Entities (English)
  • OntoNotes* (English, Chinese, Arabic)
  • WikiANN* (many)
  • CNEC 2.0 (Czech)

Alomst every dataset can be downloaded automatically with the supplied scripts. The datasets marked with an asterisk (*) require user intervention. Please refer to the readme.md files in the corresponding sub-directories for instructions.

Please be aware that some datasets are very large and take a while to download and convert

Installation of Requirements

Make sure to install all required packages (Python & R) before proceeding.

  1. Create a virtual environment
    • python3 -m venv .venv
    • source .venv/bin/activate
  2. Execute the script install_prerequisites.sh

Manual Installation

Install Python Dependencies

python -m pip install -r requirements.txt

Additionally, get a script from huggingface:

curl https://huggingface.co/datasets/conll2003/raw/main/conll2003.py -o utils/conll2003.py

Then, get spaCy models

python -m spacy download zh_core_web_lg
python -m spacy download zh_core_web_trf
python -m spacy download nl_core_news_lg
python -m spacy download en_core_web_lg
python -m spacy download fr_core_news_lg
python -m spacy download de_core_news_lg
python -m spacy download es_core_news_lg
python -m spacy download xx_ent_wiki_sm

Install R Packages

Rscript r_packages.r

Install Tools

python3 tools/corenlp/get_corenlp.py
Rscript tools/icews/get_icews.r
python3 tools/jrcnames/get_jrc.py
python3 tools/nltk/get_dependencies.py
python3 tools/opennlp/get_opennlp.py

Details on Data

Data Conversion Scripts

Collection of scripts that automatically retrieve the datasets (if possible) and then convert them to a common format.

Every script should be run from the root directory: For example, if you want to automatically get the CoNLL2002 dataset run the following python corpora/conll/get_conll2002.py

When you run the scripts that automatically download and convert the corpora, a registry.csv is created that contains meta-information on each corpus. This file is used by the evaluation scripts to automatically find all available datasets and run the tests.

Each corpus is in tokenized long format (one row = one token) and contains the following columns:

  • dataset: name of dataset
  • language: language of dataset / tokens
  • subset: Original name of subset (or split) of dataset. E.g., training, validation, etc.
  • sentence_id: id of sentence (string), typically enumerated from 000001. In some cases the corpus also has document ids, then the sentence_id includes the doc_id as well. E.g, 0001_000001.
  • token_id: id (actually position) of token within the sentence. Always starts at 1.
  • token: actual token in its original form.
  • CoNLL_IOB2: Named entity tag according to Inside-Outside-Beginning scheme as defined by CoNLL. Named entities are limited to Persons, Organizations, Location, and Misc.

Difficult Examples

See the file challenges.json for a set of sentences which pose challenges for NER tools.

Automatically Getting & Installing Tools

Every script should be run from the root directory: For example, if you want to automatically get the CoreNLP run the following python tools/corenlp/get_corenlp.py

Other Tools (not evaluated here)

More Corpora (not included, yet)

English

French

License for Quaero corpus prohibits to train a model with the data and to redistribute the resulting model. Hence corpus only for validation purposes.

Polish

  • NJKP: 1-million-word subcorpus. The manually annotated 1-million word subcorpus of the NJKP, available on CC-BY 4.0.

Russian

Hungarian

Japanese

Italian

Collections of more corpora (other domains)