qanom.py

# coding=utf-8
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""A Dataset loading script for the QANom dataset (klein et. al., COLING 2000)."""


from dataclasses import dataclass
from typing import Optional, Tuple, Union, Iterable, Set
import datasets
from pathlib import Path
import pandas as pd
import gzip
import json
import itertools

_CITATION = """\
@inproceedings{klein2020qanom,
  title={QANom: Question-Answer driven SRL for Nominalizations},
  author={Klein, Ayal and Mamou, Jonathan and Pyatkin, Valentina and Stepanov, Daniela and He, Hangfeng and Roth, Dan and Zettlemoyer, Luke and Dagan, Ido},
  booktitle={Proceedings of the 28th International Conference on Computational Linguistics},
  pages={3069--3083},
  year={2020}
}
"""


_DESCRIPTION = """\
The dataset contains question-answer pairs to model predicate-argument structure of deverbal nominalizations. 
The questions start with wh-words (Who, What, Where, What, etc.) and contain a the verbal form of a nominalization from the sentence; 
the answers are phrases in the sentence. 
See the paper for details: QANom: Question-Answer driven SRL for Nominalizations (Klein et. al., COLING 2020)
For previewing the QANom data along with the verbal annotations of QASRL, check out "https://browse.qasrl.org/".  
This dataset was annotated by selected workers from Amazon Mechanical Turk.
"""

_HOMEPAGE = "https://github.com/kleinay/QANom"

_LICENSE = """MIT License

Copyright (c) 2020 Ayal Klein (kleinay)

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE."""


_URLs = {
    "qanom_csv":  "https://github.com/kleinay/QANom/raw/master/qanom_dataset.zip",
    "qanom_jsonl":  "https://github.com/julianmichael/qasrl.org/raw/refs/heads/master/data/qanom.tar"
}

SpanFeatureType = datasets.Sequence(datasets.Value("int32"), length=2)

SUPPOERTED_DOMAINS = {"wikinews", "wikipedia"}

@dataclass
class QANomBuilderConfig(datasets.BuilderConfig):
    """ Allow the loader to re-distribute the original dev and test splits between train, dev and test. """
    redistribute_dev: Tuple[float, float, float] = (0., 1., 0.)
    redistribute_test: Tuple[float, float, float] = (0., 0., 1.)
    load_from: str = "jsonl" # "csv" or "jsonl"
    domains: Union[str, Iterable[str]] = "all" # can provide also a subset of acceptable domains.
    # Acceptable domains are {"wikipedia", "wikinews"} for dev and test (qasrl-2020)
    # and {"wikipedia", "wikinews", "TQA"} for train (qasrl-2018)    

# TODO: Name of the dataset usually match the script name with CamelCase instead of snake_case
class Qanom(datasets.GeneratorBasedBuilder):
    """QANom: Question-Answer driven SRL for Nominalizations corpus.
    Notice: This datasets genrally follows the format of `qa_srl` and `kleinay\qa_srl2018` datasets. 
    However, it extends Features to include "is_verbal" and "verb_form" fields (required for nominalizations).
    In addition, and most critically, unlike these verbal qasrl datasets, in the qanom datset some examples
    are for canidate nominalization which are judged to be non-predicates ("is_verbal"==False) or predicates with no QAs.
    In these cases, the qa fields (question, answers, answer_ranges) would be empty lists. """

    VERSION = datasets.Version("1.2.0")
    
    BUILDER_CONFIG_CLASS = QANomBuilderConfig

    BUILDER_CONFIGS = [
        QANomBuilderConfig(
            name="default", version=VERSION, description="This provides the QANom dataset"#, redistribute_dev=(0,1,0)
        ),
    ]

    DEFAULT_CONFIG_NAME = (
        "default"  # It's not mandatory to have a default configuration. Just use one if it make sense.
    )

    def _info(self):
        features = datasets.Features(
            {
                "sentence": datasets.Value("string"),
                "sent_id": datasets.Value("string"),
                "predicate_idx": datasets.Value("int32"),
                "predicate": datasets.Value("string"),
                "is_verbal": datasets.Value("bool"),
                "verb_form": datasets.Value("string"),                
                "question": datasets.Sequence(datasets.Value("string")),
                "answers": datasets.Sequence(datasets.Value("string")),
                "answer_ranges": datasets.Sequence(SpanFeatureType)
            }
        )
        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,
            # specify them here. They'll be used if as_supervised=True in
            # builder.as_dataset.
            supervised_keys=None,
            # Homepage of the dataset for documentation
            homepage=_HOMEPAGE,
            # License for the dataset if available
            license=_LICENSE,
            # Citation for the dataset
            citation=_CITATION,
        )

    def _prepare_wiktionary_verb_inflections(self, dl_manager):
        wiktionary_url = "https://raw.githubusercontent.com/nafitzgerald/nrl-qasrl/master/data/wiktionary/en_verb_inflections.txt"
        wiktionary_path = dl_manager.download(wiktionary_url)
        verb_map = {}
        with open(wiktionary_path, 'r', encoding="utf-8") as f:
            for l in f.readlines():
                inflections = l.strip().split('\t')
                stem, presentsingular3rd, presentparticiple, past, pastparticiple = inflections
                for inf in inflections:
                    verb_map[inf] = {"Stem" : stem, "PresentSingular3rd" : presentsingular3rd, "PresentParticiple":presentparticiple, "Past":past, "PastParticiple":pastparticiple}
        self.verb_inflections = verb_map 
                
    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""            
         
        assert self.config.load_from in ("csv", "jsonl")

        # Handle domain selection
        domains: Set[str] = [] 
        if self.config.domains == "all":
            domains = SUPPOERTED_DOMAINS
        elif isinstance(self.config.domains, str):
            if self.config.domains in SUPPOERTED_DOMAINS:
                domains = {self.config.domains}
            else:
                raise ValueError(f"Unrecognized domain '{self.config.domains}'; only {SUPPOERTED_DOMAINS} are supported")
        else:
            domains = set(self.config.domains) & SUPPOERTED_DOMAINS
            if len(domains) == 0:
                raise ValueError(f"Unrecognized domains '{self.config.domains}'; only {SUPPOERTED_DOMAINS} are supported")
        self.config.domains = domains
                
        self.corpus_base_path = Path(dl_manager.download_and_extract(_URLs[f"qanom_{self.config.load_from}"]))        
        if self.config.load_from == "csv":
            # prepare wiktionary for verb inflections inside 'self.verb_inflections'
            self._prepare_wiktionary_verb_inflections(dl_manager)
            self.dataset_files = [
                self.corpus_base_path / "annot.train.csv",
                self.corpus_base_path / "annot.dev.csv",
                self.corpus_base_path / "annot.test.csv"
            ]
        elif self.config.load_from == "jsonl":
            self.dataset_files = [
                self.corpus_base_path / "qanom" / "train.jsonl.gz",
                self.corpus_base_path / "qanom" / "dev.jsonl.gz",
                self.corpus_base_path / "qanom" / "test.jsonl.gz"
            ]
             
        
        # proportional segment (start,end) to take from every original split to returned SplitGenerator 
        orig_dev_segments = ((0, self.config.redistribute_dev[0]),
                             (self.config.redistribute_dev[0], sum(self.config.redistribute_dev[:2])),
                             (sum(self.config.redistribute_dev[:2]), 1))
        orig_tst_segments = ((0, self.config.redistribute_test[0]),
                             (self.config.redistribute_test[0], sum(self.config.redistribute_test[:2])),
                             (sum(self.config.redistribute_test[:2]), 1))
        train_proportion = ((0,1),                  # from train
                            orig_dev_segments[0],   # from dev
                            orig_tst_segments[0])   # from test
        dev_proportion = ((0,0),                    # from train
                          orig_dev_segments[1],     # from dev
                          orig_tst_segments[1])     # from test
        test_proportion = ((0,0),                   # from train
                           orig_dev_segments[2],    # from dev
                           orig_tst_segments[2])    # from test
        
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "split_proportion": train_proportion
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "split_proportion": dev_proportion
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "split_proportion": test_proportion
                },
            ),
        ]

    def _generate_examples(self, split_proportion):
        if self.config.load_from == "csv":
            return self._generate_examples_from_csv(split_proportion=split_proportion)
        elif self.config.load_from == "jsonl": 
            return self._generate_examples_from_jsonl(split_proportion=split_proportion)
    
    def _generate_examples_from_jsonl(self, split_proportion):
        """ Yields examples from a jsonl.gz file, in same format as qasrl-v2."""
        empty_to_underscore = lambda s: "_" if s=="" else s
        def read_lines(filepath):
            with gzip.open(filepath, "rt") as f:
                return [line.strip() for line in f]
            
        
        orig_splits_jsons = [read_lines(filepath)
                           for filepath in self.dataset_files]  # train, dev, test
        # Each json-line stands for a sentence with several predicates and QAs; we will redistribute  
        #  the new proportions of the splits on the sentence level for convenience 
        lines_from_orig_splits = [jsonlines[int(len(jsonlines)*start) : int(len(jsonlines)*end)] 
                                  for jsonlines, (start,end) in zip(orig_splits_jsons, split_proportion)] 
        this_split_lines = list(itertools.chain(*lines_from_orig_splits))
        qa_counter = 0
        for line in this_split_lines:
            sent_obj = json.loads(line.strip())
            tokens = sent_obj['sentenceTokens']
            sentence = ' '.join(tokens)
            sent_id = sent_obj['sentenceId']
            # consider only selected domains
            sent_domain = sent_id.split(":")[1]
            if sent_domain not in self.config.domains:
                continue
            for predicate_idx, verb_obj in sent_obj['verbEntries'].items():
                verb_forms = verb_obj['verbInflectedForms']
                predicate = tokens[int(predicate_idx)]
                for question_obj in verb_obj['questionLabels'].values():
                    question_slots = question_obj['questionSlots']
                    verb_form = question_slots['verb'] 
                    verb_surface = verb_forms[verb_form.split(" ")[-1]] # if verb_form in verb_forms else verb_forms['stem']
                    question_slots_in_order = [
                        question_slots["wh"],
                        question_slots["aux"],
                        question_slots["subj"],                            
                        verb_surface,
                        question_slots["obj"],
                        empty_to_underscore(question_slots["prep"]), # fix bug in data
                        question_slots["obj2"],
                        '?'
                    ] 
                    # retrieve answers
                    answer_spans = []
                    for ans in question_obj['answerJudgments']:
                        if ans['isValid']: 
                            answer_spans.extend(ans['spans']) 
                    answer_spans = list(set(tuple(a) for a in answer_spans))
                    # answer_spans = list(set(answer_spans))
                    answer_strs = [' '.join([tokens[i] for i in range(*span)])
                                for span in answer_spans]
                    
                    yield qa_counter, {
                        "sentence": sentence,
                        "sent_id": sent_id,
                        "predicate_idx": predicate_idx,
                        "predicate": predicate,
                        "is_verbal": True,
                        "verb_form": verb_forms['stem'],
                        "question": question_slots_in_order,
                        "answers": answer_strs,
                        "answer_ranges": answer_spans
                    }
                    qa_counter += 1
            # also return non-predicates with empty data
            for non_predicate_idx, non_predicate in sent_obj["nonPredicates"].items():
                yield qa_counter, {
                    "sentence": sentence,
                    "sent_id": sent_obj['sentenceId'],
                    "predicate_idx": int(non_predicate_idx),
                    "predicate": non_predicate,
                    "is_verbal": False,
                    "verb_form": "",
                    "question": [],
                    "answers": [],
                    "answer_ranges": []
                }
                qa_counter += 1
                
        
    @classmethod
    def span_from_str(cls, s:str):
        start, end = s.split(":")
        return [int(start), int(end)]
    
    def _generate_examples_from_csv(self, split_proportion):

        """ Yields examples from a 'annot.?.csv' file in QANom's format."""

        # construct concatenated DataFrame from different source splits
        orig_splits_dfs = [pd.read_csv(filepath)
                           for filepath in self.dataset_files]  # train, dev, test
        segment_df_from_orig_splits = [df.iloc[int(len(df)*start) : int(len(df)*end)] 
                                       for df, (start,end) in zip(orig_splits_dfs, split_proportion)] 
            
        df = pd.concat(segment_df_from_orig_splits, ignore_index=True)
        for counter, row in df.iterrows():
            # Each record (row) in csv is a QA or is stating a predicate/non-predicate with no QAs
            
            # consider only selected domains
            sent_domain = row.qasrl_id.split(":")[1]
            if sent_domain not in self.config.domains:
                continue
            
            # Prepare question (slots)
            na_to_underscore = lambda s: "_" if pd.isna(s) else str(s)
            question = [] if pd.isna(row.question) else list(map(na_to_underscore, [
                    row.wh, row.aux, row.subj, row.verb_slot_inflection, row.obj, row.prep, row.obj2 
                ])) + ['?']
            # fix verb slot - replace with actual verb inflection, and prepend verb_prefix 
            if question:
                if row.verb_form in self.verb_inflections and not pd.isna(row.verb_slot_inflection):
                    verb_surface = self.verb_inflections[row.verb_form][row.verb_slot_inflection] 
                else:
                    verb_surface = row.verb_form
                if not pd.isna(row.verb_prefix):
                    verb_surface = row.verb_prefix.replace("~!~", " ") + " " + verb_surface
                question[3] = verb_surface
            answers = [] if pd.isna(row.answer) else row.answer.split("~!~")
            answer_ranges = [] if pd.isna(row.answer_range) else [Qanom.span_from_str(s) for s in row.answer_range.split("~!~")]
            
            yield counter, {
                "sentence": row.sentence,
                "sent_id": row.qasrl_id,
                "predicate_idx": row.target_idx,
                "predicate": row.noun,
                "is_verbal": row.is_verbal,
                "verb_form": row.verb_form,
                "question": question,
                "answers": answers,
                "answer_ranges": answer_ranges
            }