GENIA-Term-Corpus.py

import random
import re
import xml.etree.ElementTree as ET
from typing import Tuple, List, Set
from tqdm import tqdm

import csv
import json
import os

import datasets

# TODO: Add BibTeX citation
# Find for instance the citation on arxiv or on the dataset repo/website
_CITATION = """\
@InProceedings{huggingface:dataset,
title = {A great new dataset},
author={huggingface, Inc.
},
year={2020}
}
"""

# TODO: Add description of the dataset here
# You can copy an official description
_DESCRIPTION = """\
GENIA Term corpus
"""

# TODO: Add a link to an official homepage for the dataset here
_HOMEPAGE = "http://www.geniaproject.org/genia-corpus/term-corpus"

# TODO: Add the licence for the dataset here if you can find it
_LICENSE = ""

# TODO: Add link to the official dataset URLs here
# The HuggingFace Datasets library doesn't host the datasets but only points to the original files.
# This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
_URLS = "http://www.nactem.ac.uk/GENIA/current/GENIA-corpus/Term/GENIAcorpus3.02.tgz"

def _split_files(data_dir):
    root = ET.parse(os.path.join(data_dir, "GENIA_term_3.02", "GENIAcorpus3.02.xml")).getroot()
    articles = root.findall(".//article")

    train_root = ET.Element("set")
    dev_root = ET.Element("set")
    test_root = ET.Element("set")

    for a in articles:
        root.remove(a)

    random.shuffle(articles)

    for a in articles[:1600]:
        train_root.append(a)

    for a in articles[1600:1800]:
        dev_root.append(a)

    for a in articles[1800:]:
        test_root.append(a)

    ET.ElementTree(train_root).write(os.path.join(data_dir, "train.xml"))
    ET.ElementTree(dev_root).write(os.path.join(data_dir, "dev.xml"))
    ET.ElementTree(test_root).write(os.path.join(data_dir, "test.xml"))

class GENIATermCorpus(datasets.GeneratorBasedBuilder):

    VERSION = datasets.Version("0.9.0")

    pattern = re.compile(r"[,\.;:\[\]\(\)]")

    def _info(self):

        features = datasets.Features(
            {
                "tokens": datasets.Sequence(datasets.Value("string")),
                "folded_tokens": datasets.Sequence(datasets.Value("string")),
                "labels": datasets.Sequence(datasets.Value("string"))
                    # datasets.features.ClassLabel(
                    #     names=["O", ]
                    # )
                # )
            }
        )

        return datasets.DatasetInfo(
            # This is the description that will appear on the datasets page.
            description=_DESCRIPTION,
            # This defines the different columns of the dataset and their types
            features=features,  # Here we define them above because they are different between the two configurations
            # If there's a common (input, target) tuple from the features, uncomment supervised_keys line below and
            # specify them. They'll be used if as_supervised=True in builder.as_dataset.
            # supervised_keys=("sentence", "label"),
            # Homepage of the dataset for documentation
            homepage=_HOMEPAGE,
            # License for the dataset if available
            license=_LICENSE,
            # Citation for the dataset
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):

        data_dir = dl_manager.download_and_extract(_URLS)
        # Split the dataset files in train/dev/test
        _split_files(data_dir)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "filepath": os.path.join(data_dir, "train.xml"),
                    "split": "train",
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "filepath": os.path.join(data_dir, "test.xml"),
                    "split": "test"
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "filepath": os.path.join(data_dir, "dev.xml"),
                    "split": "dev",
                },
            ),
        ]


    def _generate_examples(self, filepath:str, split):
        root = ET.parse(filepath)
        articles = root.findall(".//article")
        for idx, article in enumerate(articles):
            article_id, data= self.parse_article(article)
            for sen_ix, (tokens, entities) in enumerate(data):
                yield f"{split}_{idx}_{sen_ix}", {
                    "tokens": tokens,
                    "folded_tokens": [t.lower() for t in tokens],
                    "labels": entities
                }

    def parse_article(self, article:ET):
        # Get the id of the article
        article_id = article.find("./articleinfo/bibliomisc").text
        # Select all sentences in the article object
        sentences = article.findall(".//sentence")
        data = list()
        for sentence in sentences:
            data.append(self. build_bio_tags(*self.flatten_tree(sentence)))

        return article_id, data

    def build_bio_tags(self, text_segments:List[str], entities:List[str]) -> Tuple[List[str], List[str]]:

        # Hacky tokenizer
        tokens, tags = list(), list()
        for seg, entity in zip(text_segments, entities):
            # Insert whitespaces
            seg = self.pattern.sub(r" \g<0> ", seg).strip() # Remove trailing whitespaces
            t = seg.split()
            tokens.extend(t)
            tags.extend( [f"B-{entity}"] + [f"I-{entity}"] * (len(t) - 1) if entity != "O" else ["O"] * len(t))
        return tokens, tags


    def flatten_tree(self, elem:ET) -> Tuple[List[str], List[str]]:
        # Just keep the simple (not the nested) annotations
        text_segments, entities = list(), list()
        if  elem.text:
            text_segments.append(elem.text)
            if elem.tag == "cons" and "sem" in elem.attrib:
                tag = elem.attrib['sem'].replace("G#", "")
            else:
                tag = "O"
            entities.append(tag)
        for child in elem:
            c_segments, c_entities = self.flatten_tree(child)
            text_segments.extend(c_segments)
            entities.extend(c_entities)
        if elem.tail and elem.tail != '\n':
            text_segments.append(elem.tail)
            entities.append("O")


        return text_segments, entities