from typing import List import os import csv import ast import gzip import json import datasets from datasets.utils.logging import get_logger logger = get_logger(__name__) _URL = "https://asappresearch.github.io/slue-toolkit/" _DL_URLS = { "slue-hvb": "data/slue-hvb_blind.zip", "slue-sqa5": "data/slue-sqa5_blind.zip", "slue-vp_nel": "data/slue-vp_nel_blind.zip", "slue-ted": "data/slue-ted", } _LICENSE = """ ======================================================= The license of this script MIT License Copyright (c) 2023 ASAPP Inc. 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. ======================================================= SLUE-HVB dataset contains a subset of the Gridspace-Stanford Harper Valley speech dataset and the copyright of this subset remains the same with the original license, CC-BY-4.0. See also original license notice (https://github.com/cricketclub/gridspace-stanford-harper-valley/blob/master/LICENSE) Additionally, we provide dialog act classification annotation and it is covered with the same license as CC-BY-4.0. ======================================================= SLUE-SQA-5 Dataset SLUE-SQA-5 Dataset contains question texts and answer strings (question_text, normalized_question_text, and answer_spans column in .tsv files) from these datasets, * SQuAD1.1 (for questions whose question_id starts with ‘squad-’) * Natural Questions (for questions whose question_id starts with ‘nq-’) * WebQuestions (for questions whose question_id starts with ‘wq-’) * CuratedTREC (for questions whose question_id starts with ‘trec-’) * TriviaQA (for questions whose question_id starts with ‘triviaqa-’) Additionally, we provide audio recordings (.wav files in “question” directories) of these questions. For questions from TriviaQA (questions whose question_id starts with ‘triviaqa-’), their question texts, answer strings, and audio recordings are licensed with the same Apache License 2.0 as TriviaQA (for more detail, please refer to https://github.com/mandarjoshi90/triviaqa/blob/master/LICENSE). For questions from the other 4 datasets, their question texts, answer strings, and audio recordings are licensed with Creative Commons Attribution-ShareAlike 4.0 International license. SLUE-SQA-5 also contains a subset of Spoken Wikipedia, including the audios placed in “document” directories and their transcripts (document_text and normalized_document_text column in .tsv files). Additionally, we provide the text-to-speech alignments (.txt files in “word2time” directories).These contents are licensed with the same Creative Commons (CC BY-SA 4.0) license as Spoken Wikipedia. ======================================================= SLUE-vp_nel Dataset SLUE-vp_nel includes word-level time stamps for dev and test splits of the SLUE-voxpopuli corpus. For the dev split, the dataset also contains named entity annotations and corresponding time-stamps in a tsv format. ======================================================= SLUE-TED Dataset SLUE-TED Dataset contains TED Talk audios along with the associated abstracts and title, which were concatenated to create reference summaries. This corpus is licensed with the same Creative Commons (CC BY–NC–ND 4.0 International) license as TED talks. For further information, please refer to the details provided below. ======================================================= """ _CITATION = """\ @inproceedings{shon2023slue_phase2, title={SLUE Phase-2: A Benchmark Suite of Diverse Spoken Language Understanding Tasks}, author={Shon, Suwon and Arora, Siddhant and Lin, Chyi-Jiunn and Pasad, Ankita and Wu, Felix and Sharma, Roshan and Wu, Wei-Lun and Lee, Hung-Yi and Livescu, Karen and Watanabe, Shinji}, booktitle={ACL}, year={2023}, } """ _DESCRIPTION = """\ Spoken Language Understanding Evaluation (SLUE) benchmark Phase 2. """ def parse_qa_answer_spans(answer_spans): answer_spans = ast.literal_eval(answer_spans) return [{"answer": a, "start_second": s, "end_second": e} for a, s, e in answer_spans] def load_word2time(word2time_file): word2time = [] with open(word2time_file, "r") as f: for line in f.readlines(): entity = line.strip().split('\t') if len(entity)==1: word = entity[0] normalized_word, start_sec, end_sec = "", -1.0, -1.0 else: word, normalized_word, start_sec, end_sec = entity start_sec, end_sec = float(start_sec), float(end_sec) word2time.append( { "word": word, "normalized_word": normalized_word, "start_second": start_sec, "end_second": end_sec, } ) return word2time def parse_nel_time_spans(nel_timestamps): nel_timestamps = ast.literal_eval(nel_timestamps) if nel_timestamps is None: return [] return [ { "ne_label": ne, "start_char_idx": start, "char_offset": off, "start_sec": t0, "end_sec": t1, } for ne, start, off, t0, t1 in nel_timestamps ] def read_word_timestamps(word_alignments_fn): data = json.loads(open(word_alignments_fn).read()) return [ {"word": word, "start_sec": start, "end_sec": end} for word, start, end in data["timestamps"] ] class SLUE2Config(datasets.BuilderConfig): """BuilderConfig for SLUE.""" def __init__(self, **kwargs): """ Args: data_dir: `string`, the path to the folder containing the files in the downloaded .tar citation: `string`, citation for the data set url: `string`, url for information about the data set **kwargs: keyword arguments forwarded to super. """ super(SLUE2Config, self).__init__( version=datasets.Version("2.4.0", ""), **kwargs ) class SLUE2(datasets.GeneratorBasedBuilder): """Librispeech dataset.""" DEFAULT_WRITER_BATCH_SIZE = 256 DEFAULT_CONFIG_NAME = "hvb" BUILDER_CONFIGS = [ SLUE2Config( name="hvb", description="SLUE-HVB set.", ), SLUE2Config( name="sqa5", description="SLUE-SQA-5 set which includes Spoken Question Answering task.", ), SLUE2Config( name="vp_nel", description="SLUE-vp_nel set with named entity labels and time-stamps.", ), SLUE2Config( name="ted", description="SLUE-TED set which includes Speech Summarisation task", ), ] def _info(self): if self.config.name == "hvb": features = { "issue_id": datasets.Value("string"), "audio": datasets.Audio(sampling_rate=16_000), "speaker_id": datasets.Value("string"), "text": datasets.Value("string"), "utt_index": datasets.Value("int32"), "channel": datasets.Value("int32"), "role": datasets.Value("string"), "start_ms": datasets.Value("int32"), "duration_ms": datasets.Value("int32"), "intent": datasets.Value("string"), "dialog_acts": datasets.Sequence( datasets.Value("string"), ), } elif self.config.name == "sqa5": features = { "question_id": datasets.Value("string"), "question_audio": datasets.Audio(sampling_rate=16_000), "question_speaker_id": datasets.Value("string"), "raw_question_text": datasets.Value("string"), "normalized_question_text": datasets.Value("string"), "document_id": datasets.Value("string"), "document_audio": datasets.Audio(sampling_rate=16_000), "document_speaker_id": datasets.Value("string"), "raw_document_text": datasets.Value("string"), "normalized_document_text": datasets.Value("string"), "word2time": datasets.Sequence( { "word": datasets.Value("string"), "normalized_word": datasets.Value("string"), "start_second": datasets.Value("float64"), "end_second": datasets.Value("float64"), } ), "answer_spans": datasets.Sequence( { "answer": datasets.Value("string"), "start_second": datasets.Value("float64"), "end_second": datasets.Value("float64"), } ), } elif self.config.name == "vp_nel": features = { "id": datasets.Value("string"), "audio": datasets.Audio(sampling_rate=16_000), "speaker_id": datasets.Value("string"), "text": datasets.Value("string"), "word_timestamps": datasets.Sequence( { "word": datasets.Value("string"), "start_sec": datasets.Value("float64"), "end_sec": datasets.Value("float64"), } ), "ne_timestamps": datasets.Sequence( { "ne_label": datasets.Value("string"), "start_char_idx": datasets.Value("int32"), "char_offset": datasets.Value("int32"), "start_sec": datasets.Value("float64"), "end_sec": datasets.Value("float64"), } ), } elif self.config.name == "ted": features = { "id": datasets.Value("string"), "audio": datasets.Audio(sampling_rate=16_000), "speaker": datasets.Value("string"), "transcript": datasets.Value("string"), "title": datasets.Value("string"), "abstract": datasets.Value("string"), } return datasets.DatasetInfo( description=_DESCRIPTION, features=datasets.Features(features), supervised_keys=("file", "text"), homepage=_URL, citation=_CITATION, license=_LICENSE, ) def _split_generators( self, dl_manager: datasets.DownloadManager ) -> List[datasets.SplitGenerator]: config_name = f"slue-{self.config.name}" if config_name=="slue-ted": train_dl_dir = dl_manager.download_and_extract(_DL_URLS[config_name]+"_train.zip") valid_dl_dir = dl_manager.download_and_extract(_DL_URLS[config_name]+"_dev.zip") test_dl_dir = dl_manager.download_and_extract(_DL_URLS[config_name]+"_test_blind.zip") else: dl_dir = dl_manager.download_and_extract(_DL_URLS[config_name]) data_dir = os.path.join(dl_dir, config_name) splits = [] if self.config.name in ["hvb", "sqa5"]: splits.append( datasets.SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={ "filepath": os.path.join( data_dir or "", f"{config_name}_fine-tune.tsv" ), "data_dir": data_dir, }, ) ) if self.config.name in ["ted"]: splits.append( datasets.SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={ "filepath": os.path.join( os.path.join(train_dl_dir, config_name) or "", f"{config_name}_fine-tune.tsv" ), "data_dir": os.path.join(train_dl_dir, config_name), }, ) ) splits.append( datasets.SplitGenerator( name=datasets.Split.VALIDATION, gen_kwargs={ "filepath": os.path.join( os.path.join(valid_dl_dir, config_name+"_dev") or "", f"{config_name}_dev.tsv" ), "data_dir": os.path.join(valid_dl_dir, config_name+"_dev"), }, ), ) splits.append( datasets.SplitGenerator( name=datasets.Split.TEST, gen_kwargs={ "filepath": os.path.join( os.path.join(test_dl_dir, config_name+"_test") or "", f"{config_name}_test_blind.tsv" ), "data_dir": os.path.join(test_dl_dir, config_name+"_test"), }, ), ) if self.config.name in ["hvb", "sqa5", "vp_nel"]: splits.append( datasets.SplitGenerator( name=datasets.Split.VALIDATION, gen_kwargs={ "filepath": os.path.join( data_dir or "", f"{config_name}_dev.tsv" ), "data_dir": data_dir, }, ), ) splits.append( datasets.SplitGenerator( name=datasets.Split.TEST, gen_kwargs={ "filepath": os.path.join( data_dir or "", f"{config_name}_test_blind.tsv" ), "data_dir": data_dir, }, ), ) if self.config.name == "sqa5": splits.append( datasets.SplitGenerator( name="verified_test", gen_kwargs={ "filepath": os.path.join( data_dir or "", f"{config_name}_verified-test_blind.tsv" ), "data_dir": data_dir, }, ) ) return splits def _generate_examples(self, filepath, data_dir): logger.info(f"generating examples from = {filepath}") with open(filepath) as f: if self.config.name == "sqa5": reader = csv.DictReader(f, delimiter="\t", quoting=csv.QUOTE_NONE) else: reader = csv.DictReader(f, delimiter="\t") for idx, row in enumerate(reader): if self.config.name == "hvb": split = "test" if "test" in filepath else "dev" if "dev" in filepath else "fine-tune" audio_file = os.path.join( data_dir, split, f'{row["issue_id"]}_{row["start_ms"]}_{int(row["start_ms"]) + int(row["duration_ms"])}.wav' ) example = { "issue_id": row["issue_id"], "audio": audio_file, "speaker_id": row["speaker_id"], "text": row["text"], "utt_index": int(row["utt_index"]), "channel": int(row["channel"]), "role": row["role"], "start_ms": int(row["start_ms"]), "duration_ms": int(row["duration_ms"]), "intent": row["intent"], "dialog_acts": eval(row.get("dialog_acts", "[]")), } elif self.config.name == "sqa5": question_audio_file = os.path.join( data_dir, row["split"], "question", row["question_id"] + ".wav" ) document_audio_file = os.path.join( data_dir, row["split"], "document", row["document_id"] + ".wav" ) word2time_file = os.path.join( data_dir, row["split"], "word2time", row["document_id"] + ".txt" ) example = { "question_id": row["question_id"], "question_audio": question_audio_file, "question_speaker_id": row["question_speaker_id"], "raw_question_text": row["question_text"], "normalized_question_text": row["normalized_question_text"], "document_id": row["document_id"], "document_audio": document_audio_file, "document_speaker_id": row["document_speaker_id"], "raw_document_text": row["document_text"], "normalized_document_text": row["normalized_document_text"], "word2time": load_word2time(word2time_file), "answer_spans": parse_qa_answer_spans(row.get("answer_spans", "[]")), } elif self.config.name == "vp_nel": split = "test" if "test" in filepath else "dev" utt_id = row["id"] word_alignments_fn = os.path.join( data_dir, "word_timestamps", split, f"{utt_id}.json" ) audio_file = os.path.join( data_dir, 'audio', split, f"{utt_id}.ogg", ) example = { "id": utt_id, "audio": audio_file, "speaker_id": row["speaker_id"], "text": row["normalized_text"], "ne_timestamps": parse_nel_time_spans( row.get("normalized_nel", "[]") ), "word_timestamps": read_word_timestamps(word_alignments_fn), } if self.config.name == "ted": split = "test" if "test" in filepath else "dev" if "dev" in filepath else "fine-tune" audio_file = os.path.join( data_dir, split, row["id"] + ".flac" ) if "title" not in row: title_str=[] else: title_str=row["title"] if "abstract" not in row: abstract_str=[] else: abstract_str=row["abstract"] example = { "id": row["id"], "audio": audio_file, "speaker": row["speaker"], "transcript": row["transcript"], "title": title_str, "abstract": abstract_str, } yield idx, example