• 2022-07: MultiNERD, our fine-grained WikiNEuRal extension, has been published and the produced dataset is available here!
• 2022-01: A Multilingual Named Entity Recognition model trained on WikiNEuRal is now available on HuggingFace Models!
• 2021-12: Our WikiNEuRal dataset is available on HuggingFace Datasets!

Data and evaluation code for the paper WikiNEuRal: Combined Neural and Knowledge-based Silver Data Creation for Multilingual NER.

This repository is mainly built upon Pytorch and Pytorch-Lightning.

@inproceedings{tedeschi-etal-2021-wikineural-combined,
    title = "{W}iki{NE}u{R}al: {C}ombined Neural and Knowledge-based Silver Data Creation for Multilingual {NER}",
    author = "Tedeschi, Simone  and
      Maiorca, Valentino  and
      Campolungo, Niccol{\`o}  and
      Cecconi, Francesco  and
      Navigli, Roberto",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-emnlp.215",
    pages = "2521--2533",
    abstract = "Multilingual Named Entity Recognition (NER) is a key intermediate task which is needed in many areas of NLP. In this paper, we address the well-known issue of data scarcity in NER, especially relevant when moving to a multilingual scenario, and go beyond current approaches to the creation of multilingual silver data for the task. We exploit the texts of Wikipedia and introduce a new methodology based on the effective combination of knowledge-based approaches and neural models, together with a novel domain adaptation technique, to produce high-quality training corpora for NER. We evaluate our datasets extensively on standard benchmarks for NER, yielding substantial improvements up to 6 span-based F1-score points over previous state-of-the-art systems for data creation.",
}

Please consider citing our work if you use data and/or code from this repository.

In a nutshell, WikiNEuRal consists in a novel technique which builds upon a multilingual lexical knowledge base (i.e., BabelNet) and transformer-based architectures (i.e., BERT) to produce high-quality annotations for multilingual NER. It shows consistent improvements of up to 6 span-based F1-score points against state-of-the-art alternative data production methods on common benchmarks for NER:

Moreover, in our paper we also present a new approach for creating interpretable word embeddings together with a Domain Adaptation algorithm, which enable WikiNEuRal to create domain-specific training corpora.

Further datasets, such as the combination of WikiNEuRal with gold-standard training data (i.e., CoNLL), can be obtained by simply concatenating the two train.conllu files together (e.g., data/conll/en/train.conllu and data/wikineural/en/train.conllu give CoNLL+WikiNEuRal).

We trained a Multilingual Named Entity Recognition model on WikiNEuRal and made it available on 🤗 Models (see the Tutorial Notebook). Specifically, we fine-tuned Multilingual BERT (mBERT) for 3 epochs on our dataset (on all 9 languages jointly). Therefore, the system supports the 9 languages covered by WikiNEuRal (de, en, es, fr, it, nl, pl, pt, ru). For a stronger system (mBERT + Bi-LSTM + CRF) look at Reproduce Paper Results Section.

1. To train 10 models on WikiNEuRal English, run:

   python run.py -m +train.seed_idx=0,1,2,3,4,5,6,7,8,9 data.datamodule.source=wikineural data.datamodule.language=en
   

   note 1: we show how to train 10 models on the same training set because in our paper we used exactly 10 different seeds to compute the mean and standard deviation, which were then used to measure the statistical significance w.r.t. other training sets.

   note 2: for the EN, ES, NL and DE versions of WikiNEuRal, you can combine the train, val and test sets of WikiNEuRal to obtain a larger training set (as we did in the paper), and use the CoNLL splits as validation and testing material (e.g., copy data/conll/en/val.conllu into data/wikineural/en/). Similarly, for RU and PL you can use the BSNLP splits for validation and testing. For the other languages instead, since no gold-standard data are available, you can use the standard WikiNEuRal splits to validate your model. Finally, for the domain-adapted datasets you can use both the val and test sets of the datasets on which they have been adapted (if available) and the val and test sets of WikiNEuRal for the same language.

2. To produce results for the 10 trained models, run:

   bash test.sh
   

   test.sh also contains more complex bash for loops that can produce results on multiple datasets / models at once.

WikiNEuRal is licensed under the CC BY-SA-NC 4.0 license. The text of the license can be found here.

We underline that the source from which the raw sentences have been extracted is Wikipedia (wikipedia.org) and the NER annotations have been produced by Babelscape.

We gratefully acknowledge the support of the ERC Consolidator Grant MOUSSE No. 726487 under the European Union’s Horizon2020 research and innovation programme (http://mousse-project.org/).

This work was also supported by the PerLIR project (Personal Linguistic resources in Information Retrieval) funded by the MIUR Progetti di ricerca di Rilevante Interesse Nazionale programme (PRIN2017).

The code in this repository is built on top of .