StAn - Quickly annotate your dataset with Stanford CoreNLP

In natural language processing, algorithms often require additional linguistic features (syntactic and semantic), such as part-of-speech, named entity, and dependency tags; information that is not readily available in most datasets. StAn provides a convenient way to quickly annotate an existing dataset with additional linguistic features computed by Stanford CoreNLP.

Getting Started

Prerequisites

StAn either uses a local CoreNLP installation or an exisiting CoreNLP Server. To use a local installation, download and unpack the latest version from the Stanford CoreNLP website.

Installing

With pip

TBD

From Source

Clone the repository and run:

pip install [--editable] .

Usage

For example, the following command annotates the SemEval 2010 Task 8 relation extraction dataset with POS, NER, and dependency information and saves it in JSONL format.

stan \
    --input-dir $INPUT_PATH/SemEval2010_task8_all_data/ \
    --output-dir $OUTPUT_PATH/ \
    --corenlp $PATH_TO_CORENLP_JAR_OR_SERVER_URL \
    --input-format semeval2010task8 \
    --output-format jsonl \
    --shuffle \
    --validation-size 0.1 \
    --n-jobs 4

Parameters:

• input-dir: the directory containing the dataset or dataset files. StAn expects a specific structure for common datasets (e.g. SemEval 2010 Task 8). The format of the input is specified by input-format.
• output-dir: the directory to store the annotated dataset. The format in which to save the dataset is specified by output-format.
• corenlp: the path to the directory containing the CoreNLP jar file or a url pointing to an exisiting CoreNLP server.
• input-format: the format of the input dataset, can be one of "semeval2010task8", "json" or "jsonl".
• output-format: the format of the output dataset, can be one of "tacred", "json", "jsonl".
• shuffle: whether to shuffle the training dataset before splitting into train and validation (only if validation size > 0).
• validation-size: if > 0, use a validation-size fraction of the training dataset for validation.
• n-jobs: the number of threads to use for concurrent requests to CoreNLP.

Running the tests

Explain how to run the automated tests for this system

Unittests

pytest -v tests/

Typechecker and coding style tests

mypy stan --ignore-missing-imports

Built With

• Stanford CoreNLP
• stanford-corenlp - Python wrapper for Stanford CoreNLP

Authors

• Christoph Alt

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the LICENSE file for details