gkomet.py

"""Metaphor corpus G-KOMET 1.0"""
import logging
import os
import re
import xml.etree.ElementTree as ET
from typing import List, Tuple

import datasets

_CITATION = """\
@InProceedings{antloga2022gkomet,
title = {Korpusni pristopi za identifikacijo metafore in metonimije: primer metonimije v korpusu gKOMET},
author={Antloga, \v{S}pela},
booktitle={Proceedings of the Conference on Language Technologies and Digital Humanities (Student papers)},
year={2022},
pages={271-277}
}
"""


_DESCRIPTION = """\
G-KOMET 1.0 (a corpus of metaphorical expressions in spoken Slovene language) is a corpus of speech transcriptions and 
conversations that covers 50,000 lexical units. The corpus contains samples from the Gos corpus of spoken Slovene 
and includes a balanced set of transcriptions of informative, educational, entertaining, private, and public discourse.

The annotation scheme was based on the MIPVU metaphor identification process. 
This protocol was modified and adapted to the specifics of the Slovene language and the specifics of the spoken 
language. Corpus was annotated for the following relations to metaphor: indirect metaphor, direct metaphor, borderline 
cases and metaphor signals. In addition, the corpus introduces a new ‘frame’ tag, which gives information about the 
concept to which it refers. 
"""

_HOMEPAGE = "http://hdl.handle.net/11356/1490"

_LICENSE = "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)"

_URLS = {
    "gkomet": "https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1490/G-Komet.zip"
}


XML_NAMESPACE = "{http://www.w3.org/XML/1998/namespace}"
EL_LEAF, EL_TYPE, EL_FRAME = range(3)


def namespace(element):
    # https://stackoverflow.com/a/12946675
    m = re.match(r'\{.*\}', element.tag)
    return m.group(0) if m else ''


def word_info(sent_el):
    def _resolve_recursively(element) -> List:
        """ Knowingly ignored tags: name (anonymized, without IDs), gap, vocal, pause, del,
        linkGrp (handled separately in linkgroup_info()) """
        # Leaf node: word or punctuation character
        if element.tag.endswith(("w", "pc")):
            id_curr = element.attrib[f"{XML_NAMESPACE}id"]
            return [(id_curr, element.text)]

        # Annotated word or word group - not interested in the annotations in this function
        elif element.tag.endswith("seg"):
            parsed_data = []
            for child in element:
                if child.tag.endswith(("c", "vocal", "pause")):  # empty space betw. words or "special" word
                    continue

                res = _resolve_recursively(child)
                if isinstance(res, list):
                    parsed_data.extend(res)
                else:
                    parsed_data.append(res)

            return parsed_data

    id_words, words = [], []
    for child_el in sent_el:
        curr_annotations = _resolve_recursively(child_el)
        if curr_annotations is not None:  # None = unrecognized ("unimportant") element
            for ann in curr_annotations:
                id_words.append(ann[0])
                words.append(ann[1])

    return id_words, words


def seg_info(sent_el):
    def _resolve_recursively(element) -> Tuple:
        """ Returns (type[, subtype], deeper_elements, latest_element)"""
        # Leaf node: word or punctuation character
        if element.tag.endswith(("w", "pc")):
            id_curr = element.attrib[f"{XML_NAMESPACE}id"]
            return EL_LEAF, [], [id_curr]

        # Annotated word or word group
        elif element.tag.endswith("seg"):
            subtype = element.attrib["subtype"]
            if element.attrib["type"] == "frame":
                ann_type = EL_FRAME
            elif element.attrib["type"] == "metaphor":
                ann_type = EL_TYPE
            elif element.attrib["type"] == "idiom":
                ann_type = EL_TYPE
            else:
                raise ValueError(f"Unrecognized seg type: {element.attrib['type']}")

            deeper_elements = []
            latest_element = []
            for child in element:
                if child.tag.endswith(("c", "vocal", "pause")):  # empty space betw. words or "special" word
                    continue

                res = _resolve_recursively(child)
                if res[0] == EL_LEAF:
                    latest_element.extend(res[2])
                else:
                    deeper_elements.extend(res[2])
                    deeper_elements.append((res[0], res[1], res[3]))
                    latest_element.extend(res[3])

            return ann_type, subtype, deeper_elements, latest_element

    annotations = []
    for child_el in sent_el:
        if not child_el.tag.endswith("seg"):
            continue

        ann_type, subtype, deeper_elements, latest_element = _resolve_recursively(child_el)
        annotations.extend(deeper_elements)
        annotations.append((ann_type, subtype, latest_element))

    return annotations


def linkgroup_info(sent_el):
    annotations = []
    for child_el in sent_el:
        if not child_el.tag.endswith("linkGrp"):
            continue

        for curr_link in child_el:
            ann_type = EL_TYPE
            if child_el.attrib["type"] not in {"metonymy", "frame", "metaphor", "idiom"}:
                logging.warning(f"Uncovered linkGrp element type, skipping: {child_el.attrib['type']}")
                continue

            if child_el.attrib["type"] == "metonymy":
                subtype = curr_link.attrib["ana"]
            elif child_el.attrib["type"] in {"frame", "metaphor"}:
                ann_type = EL_TYPE if child_el.attrib["type"] == "metaphor" else EL_FRAME
                subtype = curr_link.attrib["ana"].split(":")[-1]
            else:
                subtype = "idiom"

            tokens_involved = list(map(lambda _tok_id: _tok_id[1:] if _tok_id.startswith("#") else _tok_id,
                                       curr_link.attrib["target"].split(" ")))
            annotations.append((ann_type, subtype, tokens_involved))

    return annotations


class GKomet(datasets.GeneratorBasedBuilder):
    """G-KOMET 1.0 is a corpus of metaphorical expressions in spoken Slovene language. """

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        features = datasets.Features(
            {
                "document_name": datasets.Value("string"),
                "idx": datasets.Value("uint32"),  # index inside current document
                "idx_paragraph": datasets.Value("uint32"),
                "idx_sentence": datasets.Value("uint32"),  # index inside current paragraph
                "sentence_words": datasets.Sequence(datasets.Value("string")),
                "met_type": [{
                    "type": datasets.Value("string"),
                    "word_indices": datasets.Sequence(datasets.Value("uint32"))
                }],
                "met_frame": [{
                    "type": datasets.Value("string"),
                    "word_indices": datasets.Sequence(datasets.Value("uint32"))
                }]
            }
        )
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        data_dir = dl_manager.download_and_extract(_URLS["gkomet"])
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"data_dir": os.path.join(data_dir, "G-Komet")},
            )
        ]

    def _generate_examples(self, data_dir):
        data_files = []
        for fname in os.listdir(data_dir):
            curr_path = os.path.join(data_dir, fname)
            if os.path.isfile(curr_path) and fname.endswith(".xml") and fname != "G-Komet.xml":  # G-Komet.xml = meta-file
                data_files.append(fname)
        data_files = sorted(data_files)

        idx_example = 0
        for fname in data_files:
            fpath = os.path.join(data_dir, fname)
            curr_doc = ET.parse(fpath)
            root = curr_doc.getroot()
            NAMESPACE = namespace(root)

            idx_sent_glob = 0
            for idx_par, curr_par in enumerate(root.iterfind(f".//{NAMESPACE}p")):
                id2position = {}  # {<idx_sent> -> {<id_word>: <position> foreach word} foreach sent}
                all_words = []

                # Pass#1: extract word information
                for idx_sent, curr_sent in enumerate(curr_par.iterfind(f"{NAMESPACE}s")):
                    id_words, words = word_info(curr_sent)

                    id2position[idx_sent] = dict(zip(id_words, range(len(words))))
                    all_words.append(words)

                all_types, all_frames = [], []

                # Pass#2: extract annotations from <seg>ments
                for idx_sent, curr_sent in enumerate(curr_par.iterfind(f"{NAMESPACE}s")):
                    annotated_segs = seg_info(curr_sent)
                    all_types.append([])
                    all_frames.append([])

                    for curr_ann in annotated_segs:
                        ann_type, ann_subtype, words_involved = curr_ann
                        if ann_type == EL_TYPE:
                            all_types[idx_sent].append({
                                "type": ann_subtype,
                                "word_indices": [id2position[idx_sent][_id_word] for _id_word in words_involved
                                                 if _id_word in id2position[idx_sent]]
                            })
                        elif ann_type == EL_FRAME:
                            all_frames[idx_sent].append({
                                "type": ann_subtype,
                                "word_indices": [id2position[idx_sent][_id_word] for _id_word in words_involved
                                                 if _id_word in id2position[idx_sent]]
                            })

                # Pass#3: extract annotations from <linkGrp>s
                for idx_sent, curr_sent in enumerate(curr_par.iterfind(f"{NAMESPACE}s")):
                    annotated_linkgroups = linkgroup_info(curr_sent)

                    for curr_ann in annotated_linkgroups:
                        ann_type, ann_subtype, words_involved = curr_ann

                        if ann_type == EL_TYPE:
                            all_types[idx_sent].append({
                                "type": ann_subtype,
                                "word_indices": [id2position[idx_sent][_id_word] for _id_word in words_involved
                                                 if _id_word in id2position[idx_sent]]
                            })
                        elif ann_type == EL_FRAME:
                            all_frames[idx_sent].append({
                                "type": ann_subtype,
                                "word_indices": [id2position[idx_sent][_id_word] for _id_word in words_involved
                                                 if _id_word in id2position[idx_sent]]
                            })

                idx_sent = 0
                for curr_words, curr_types, curr_frames in zip(all_words, all_types, all_frames):
                    if len(curr_words) == 0:
                        continue

                    yield idx_example, {
                        "document_name": fname,
                        "idx": idx_sent_glob,
                        "idx_paragraph": idx_par,
                        "idx_sentence": idx_sent,
                        "sentence_words": curr_words,
                        "met_type": curr_types,
                        "met_frame": curr_frames
                    }
                    idx_example += 1
                    idx_sent += 1
                    idx_sent_glob += 1