vuamc.py

""" English metaphor-annotated corpus. """

import os

import datasets
import logging
import re

import xml.etree.ElementTree as ET
from typing import List, Tuple, Dict


_CITATION = """\
@book{steen2010method,
  title={A method for linguistic metaphor identification: From MIP to MIPVU},
  author={Steen, Gerard and Dorst, Lettie and Herrmann, J. and Kaal, Anna and Krennmayr, Tina and Pasma, Trijntje},
  volume={14},
  year={2010},
  publisher={John Benjamins Publishing}
}
"""

_DESCRIPTION = """\
The resource contains a selection of excerpts from BNC-Baby files that have been annotated for metaphor. 
There are four registers, each comprising about 50,000 words: academic texts, news texts, fiction, and conversations. 
Words have been separately labelled as participating in multi-word expressions (about 1.5%) or as discarded for 
metaphor analysis (0.02%). Main categories include words that are related to metaphor (MRW), words that signal 
metaphor (MFlag), and words that are not related to metaphor. For metaphor-related words, subdivisions have been made 
between clear cases of metaphor versus borderline cases (WIDLII, When In Doubt, Leave It In). Another parameter of 
metaphor-related words makes a distinction between direct metaphor, indirect metaphor, and implicit metaphor.
"""

_HOMEPAGE = "https://hdl.handle.net/20.500.12024/2541"

_LICENSE = "Available for non-commercial use on condition that the terms of the BNC Licence are observed and that " \
           "this header is included in its entirety with any copy distributed."

_URLS = {
    "vuamc": "https://ota.bodleian.ox.ac.uk/repository/xmlui/bitstream/handle/20.500.12024/2541/VUAMC.xml"
}


XML_NAMESPACE = "{http://www.w3.org/XML/1998/namespace}"
VICI_NAMESPACE = "{http://www.tei-c.org/ns/VICI}"
NA_STR = "N/A"


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


def resolve_recursively(el, ns):
    words, pos_tags, met_type, meta_tags = [], [], [], []

    if el.tag.endswith("w"):
        # A <w>ord may be
        #   (1) just text,
        #   (2) a metaphor (text fully enclosed in another seg)
        #   (3) a partial metaphor (optionally some text, followed by a seg, optionally followed by more text)
        idx_word = 0
        _w_text = el.text.strip() if el.text is not None else ""
        if len(_w_text) > 0:
            words.append(_w_text)
            pos_tags.append(el.attrib["type"])
            meta_tags.append(NA_STR)
            idx_word += 1

        met_els = el.findall(f"{ns}seg")
        for met_el in met_els:
            parse_tail = True
            if met_el.text is None:
                # Handle encoding inconsistency where the metaphor is encoded without a closing tag (I hate this format)
                # <w lemma="to" type="PRP"><seg function="mrw" type="met" vici:morph="n"/>to </w>
                parse_tail = False
                _w_text = met_el.tail.strip()
            else:
                _w_text = met_el.text.strip()

            curr_met_type = met_el.attrib[f"function"]

            # Let the user decide how they want to aggregate metaphors
            if "type" in met_el.attrib:
                curr_met_type = f"{curr_met_type}/{met_el.attrib['type']}"

            if "subtype" in met_el.attrib:
                curr_met_type = f"{curr_met_type}/{met_el.attrib['subtype']}"

            words.append(_w_text)
            pos_tags.append(el.attrib["type"])
            meta_tags.append(NA_STR)
            met_type.append({"type": curr_met_type, "word_indices": [idx_word]})
            idx_word += 1

            if not parse_tail:
                continue

            _w_text = met_el.tail.strip() if met_el.tail is not None else ""
            if len(_w_text) > 0:
                words.append(_w_text)
                pos_tags.append(el.attrib["type"])
                meta_tags.append(NA_STR)
                idx_word += 1

    elif el.tag.endswith("vocal"):
        desc_el = el.find(f"{ns}desc")
        description = desc_el.text.strip() if desc_el is not None else "unknown"

        words.append("")
        pos_tags.append(NA_STR)
        meta_tags.append(f"vocal/{description}")  # vocal/<desc>

    elif el.tag.endswith("gap"):
        words.append("")
        pos_tags.append(NA_STR)
        meta_tags.append(f"gap/{el.attrib.get('reason', 'unclear')}")  # gap/<reason>

    elif el.tag.endswith("incident"):
        desc_el = el.find(f"{ns}desc")
        description = desc_el.text.strip() if desc_el is not None else "unknown"

        words.append("")
        pos_tags.append(NA_STR)
        meta_tags.append(f"incident/{description}")

    elif el.tag.endswith("shift"):
        # TODO: this is not exposed
        new_state = el.attrib.get("new", "normal")
        children = list(iter(el))
        # NOTE: Intentionally skip shifts like this, without children:
        # <u who="#PS05E"> <shift new="crying"/> </u>
        if len(children) > 0:
            for w_el in el:
                _words, _pos, _mets, _metas = resolve_recursively(w_el, ns=ns)
                words.extend(_words)
                pos_tags.extend(_pos)
                meta_tags.extend(_metas)

    elif el.tag.endswith("seg"):
        # Direct <seg> descendant of a sentence indicates truncated text
        word_el = el.find(f"{ns}w")

        words.append(word_el.text.strip())
        pos_tags.append(word_el.attrib["type"])
        meta_tags.append(NA_STR)

    elif el.tag.endswith("pause"):
        words.append("")
        pos_tags.append(NA_STR)
        meta_tags.append(f"pause")

    elif el.tag.endswith("sic"):
        for w_el in el:
            _words, _pos, _mets, _metas = resolve_recursively(w_el, ns=ns)
            words.extend(_words)
            pos_tags.extend(_pos)
            meta_tags.extend(_metas)

    elif el.tag.endswith("c"):
        words.append(el.text.strip())
        pos_tags.append(el.attrib["type"])
        meta_tags.append(NA_STR)

    elif el.tag.endswith("pb"):
        words.append("")
        pos_tags.append(NA_STR)
        meta_tags.append(NA_STR)

    elif el.tag.endswith("hi"):
        # TODO: this is not exposed
        rendition = el.attrib.get("rend", "normal")

        for child_el in el:
            _words, _pos, _mets, _metas = resolve_recursively(child_el, ns=ns)
            words.extend(_words)
            pos_tags.extend(_pos)
            meta_tags.extend(_metas)

    elif el.tag.endswith("choice"):
        sic_el = el.find(f"{ns}sic")
        _words, _pos, _mets, _metas = resolve_recursively(sic_el, ns=ns)
        words.extend(_words)
        pos_tags.extend(_pos)
        met_type.extend(_mets)
        meta_tags.extend(_metas)

    elif el.tag.endswith(("ptr", "corr")):
        # Intentionally skipping these:
        # - no idea what <ptr> is
        # - <sic> is being parsed instead of <corr>
        pass

    else:
        logging.warning(f"Unrecognized child element: {el.tag}.\n"
                        f"If you are seeing this message, please open an issue on HF datasets.")

    return words, pos_tags, met_type, meta_tags


def parse_sent(sent_el, ns) -> Tuple[List[str], List[str], List[str], List[Dict], List[str]]:
    all_words, all_pos_tags, all_met_types, all_metas = [], [], [], []
    for child_el in sent_el:
        word, pos, mtype, meta = resolve_recursively(child_el, ns=ns)
        # Need to remap local (index inside the word group) `word_indices` to global (index inside the sentence)
        if len(mtype) > 0:
            base = len(all_words)
            mtype = list(map(lambda met_info: {
                "type": met_info["type"],
                "word_indices": list(map(lambda _i: base + _i, met_info["word_indices"]))
            }, mtype))

        all_words.extend(word)
        all_pos_tags.extend(pos)
        all_met_types.extend(mtype)
        all_metas.extend(meta)

    return all_words, all_pos_tags, all_met_types, all_metas


def parse_text_body(body_el, ns):
    all_words: List[List] = []
    all_pos: List[List] = []
    all_met_type: List[List] = []
    all_meta: List[List] = []

    # Edge case#1: <s>entence
    if body_el.tag.endswith("s"):
        words, pos_tags, met_types, meta_tags = parse_sent(body_el, ns=ns)
        all_words.append(words)
        all_pos.append(pos_tags)
        all_met_type.append(met_types)
        all_meta.append(meta_tags)

    # Edge case#2: <u>tterance either contains a sentence of metadata or contains multiple sentences as children
    elif body_el.tag.endswith("u"):
        children = list(filter(lambda _child: not _child.tag.endswith("ptr"), list(iter(body_el))))
        is_utterance_sent = all(map(lambda _child: not _child.tag.endswith("s"), children))
        if is_utterance_sent:
            # <u> contains elements as children that are not a <s>entence, so it is itself considered a sentence
            words, pos_tags, met_types, meta_tags = parse_sent(body_el, ns=ns)
            all_words.append(words)
            all_pos.append(pos_tags)
            all_met_type.append(met_types)
            all_meta.append(meta_tags)
        else:
            # <u> contains one or more of <s>entence children
            for _child in children:
                words, pos_tags, met_types, meta_tags = parse_sent(_child, ns=ns)
                all_words.append(words)
                all_pos.append(pos_tags)
                all_met_type.append(met_types)
                all_meta.append(meta_tags)

    # Recursively go deeper through all the <p>aragraphs, <div>s, etc. until we reach the sentences
    else:
        for _child in body_el:
            _c_word, _c_pos, _c_met, _c_meta = parse_text_body(_child, ns=ns)

            all_words.extend(_c_word)
            all_pos.extend(_c_pos)
            all_met_type.extend(_c_met)
            all_meta.extend(_c_meta)

    return all_words, all_pos, all_met_type, all_meta


class VUAMC(datasets.GeneratorBasedBuilder):
    """English metaphor-annotated corpus. """

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        features = datasets.Features(
            {
                "document_name": datasets.Value("string"),
                "words": datasets.Sequence(datasets.Value("string")),
                "pos_tags": datasets.Sequence(datasets.Value("string")),
                "met_type": [{
                    "type": datasets.Value("string"),
                    "word_indices": datasets.Sequence(datasets.Value("uint32"))
                }],
                "meta": datasets.Sequence(datasets.Value("string"))
            }
        )

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION
        )

    def _split_generators(self, dl_manager):
        urls = _URLS["vuamc"]
        data_path = dl_manager.download_and_extract(urls)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"file_path": os.path.join(data_path)}
            )
        ]

    def _generate_examples(self, file_path):
        curr_doc = ET.parse(file_path)
        root = curr_doc.getroot()
        NAMESPACE = namespace(root)
        root = root.find(f"{NAMESPACE}text")

        idx_instance = 0
        for idx_doc, doc in enumerate(root.iterfind(f".//{NAMESPACE}text")):
            document_name = doc.attrib[f"{XML_NAMESPACE}id"]
            body = doc.find(f"{NAMESPACE}body")
            body_data = parse_text_body(body, ns=NAMESPACE)

            for sent_words, sent_pos, sent_met_type, sent_meta in zip(*body_data):
                # TODO: Due to some simplifications (not parsing certain metadata), some sentences may be empty
                if len(sent_words) == 0:
                    continue

                yield idx_instance, {
                    "document_name": document_name,
                    "words": sent_words,
                    "pos_tags": sent_pos,
                    "met_type": sent_met_type,
                    "meta": sent_meta
                }
                idx_instance += 1