BiLSTM-CNN-CRF with BERT for Sequence Tagging

This repository is based on BiLSTM-CNN-CRF ELMo implementation.

The model here present is the one presented in Deliverable 2.2 of Embeddia Project.

The dependencies for running the code are present in the environement.yml file. These can be used to create a Anaconda environement.

To run the code, you need to do:

python NER.py -f {input_file} -F {input_type} -l {language} -o {output_file}

The argument -F can support:

plain Plain text, it will tokenize the text based on white spaces
tokens The text is already tokenized and has a format similar to CONLL 2003. Where the first column correspond to the tokens.
IOB The text is already tokenized and has a format similar to CONLL 2003 however, in this case it is already annotated. The second column correspond to the tokens, the third to the labels
XML_tokenized We can support XML files that have the text already split in sentences and tokenized. The format is the one used by http://opus.nlpl.eu/
XML_sentences We can support XML files that have the text already split in sentences. The sentences will be tokenized using white spaces. The format is the one used by http://opus.nlpl.eu/
XML_Large_tokenized Similar to XML_tokenized, however, we store the last sentence processed in order to recover if needed.
XML_Large_sentences Similar to XML_sentences, however, we store the last sentence processed in order to recover if needed.

It has some optional arguments, such as:

--HTMLFormat -H This makes the output to be an HTML file that contains the original text annotated with the predicted labels
--evaluate -e If the document is annotated, it evaluates the document
--lastSentence -s This argument is used when large XML files are parsed and we need to restart the process of it. It takes the id of the last processed sentence.

License

This work has been attributed with an MIT License, with an exception to the code that it was developed by a third-party. Specifically, we have used code from:

• BiLSTM-CNN-CRF ELMo implementation: Apache 2.0 License
• phipleg/keras: MIT License

All the files have their respective license.

How to cite?

@inproceedings{moreno-etal-2019-tlr,
    title = "{TLR} at {BSNLP}2019: A Multilingual Named Entity Recognition System",
    author = "Moreno, Jose G.  and
      Linhares Pontes, Elvys  and
      Coustaty, Mickael  and
      Doucet, Antoine",
    booktitle = "Proceedings of the 7th Workshop on Balto-Slavic Natural Language Processing",
    month = aug,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/W19-3711",
    doi = "10.18653/v1/W19-3711",
    pages = "83--88"
}

Parent project

This work is is result of the European Union H2020 Project Embeddia. Embeddia is a project that creates NLP tools that focuses on European under-represented languages and that has for objective to improve the accessibility of these tools to the general public and to media enterprises. Visit Embeddia's Github to discover more NLP tools and models created within this project.