import json import logging from pathlib import Path import re from typing import Dict, List, Tuple import click import ir_datasets from hashlib import md5 from ir_datasets.datasets.trec_cast import ( NAME, ColonCommaDupes, WapoDupes, DownloadConfig, ) logging.basicConfig(level=logging.INFO) base_path = ir_datasets.util.home_path() / NAME dlc = DownloadConfig.context(NAME, base_path) RE_TOKEN_MATCHER = re.compile(r"\S") class Feeder: def __init__(self, base: str, mainpath: Path): self.feeders = [self.feeder(path) for path in mainpath.glob(f"{base}_*.jsonl")] self.current = [next(feeder) for feeder in self.feeders] def feeder(self, path: Path): with path.open("rt") as fp: for line in fp: data = json.loads(line) yield f"{data['id']}", [content["body"] for content in data["contents"]] while True: yield "", "" def get(self, doc_id: str): for ix, (key, body) in enumerate(self.current): if key == doc_id: self.current[ix] = next(self.feeders[ix]) return body assert False, f"Could not find {doc_id}" class Range: def __init__(self, text: str, pos: int): self.text = text self.start = pos self.end = pos + 1 def process_next(self, target: str): ix = self.end while ix < len(self.text): c = self.text[ix] ix += 1 if not c.isspace(): if c == target: self.end = ix return True return False return None @property def limits(self): return self.start, self.end def find_ranges(doc_id: str, body: str, passages: List[str]): # Due to a bug in chunking in CaST, we need a clever strategy # to recover ranges... so we find words from the passages passage_ranges: List[List[Tuple[int, int]]] = [] for passage in passages: passage = passage.strip() c_ranges: List[Range] = [] ranges: List[Tuple[int, int]] = [] for m in RE_TOKEN_MATCHER.finditer(passage): c = m.group(0) old_c_ranges = c_ranges # Try to expand ranges... if c_ranges: c_ranges = [range for range in c_ranges if range.process_next(c)] # otherwise, start new ones if not c_ranges: if old_c_ranges: # Takes the first one... might no be the best ranges.append(old_c_ranges[0].limits) c_ranges = [Range(body, m.start(0)) for m in re.finditer(re.escape(c), body)] if not c_ranges: logging.error( "Cannot find character %s in %s", c, doc_id, ) logging.error(" [passage] %s", passage) logging.error(" [body] %s", body) raise if c_ranges: ranges.append(c_ranges[0].limits) logging.debug(" ---> %s", passage) logging.debug("ranges: %s", ranges) logging.debug("extracts: %s", [body[start:end] for start, end in ranges]) p_1 = re.sub(r"\s+", "", "".join([body[start:end] for start, end in ranges])) p_2 = re.sub(r"\s+", "", passage) assert p_1 == p_2 passage_ranges.append(ranges) return passage_ranges ITERATORS = { # Generated using the official scripts, adapted to our cases (and dupes) "MARCO_v1": lambda: ir_datasets.load("trec-cast/v2/msmarco").docs_iter(), "WaPo-v2": lambda: ir_datasets.load("trec-cast/v2/wapo").docs_iter(), "KILT-nodupes": lambda: ir_datasets.load("trec-cast/v2/kilt").docs_iter(), # Using the official split "MARCO_v2": lambda: ir_datasets.load("trec-cast/v3/msmarco").docs_iter(), "WaPo": lambda: ir_datasets.load("trec-cast/v3/wapo").docs_iter(), "KILT": lambda: ir_datasets.load("trec-cast/v3/kilt").docs_iter(), } @click.argument("name", type=click.Choice(list(ITERATORS.keys()))) @click.argument("jsonlines", type=Path) @click.command() def cli(name: str, jsonlines: Path): """Computes the ranges based on the official CaST splits jsonlines: path to the folder containing the official jsonl files name: name of the collection to be processed """ core_iter = ITERATORS[name]() duplicates = set() post_duplicates = set() if name == "MARCO_v1": pass with ColonCommaDupes(dlc["v2/dupes/marco_v1"]).stream() as fin: post_duplicates = set(dupe_id.decode().strip() for dupe_id in fin) elif name == "KILT-nodupes": # no duplicates in v2 pass elif name == "WaPo-v2": duplicates = WapoDupes(dlc["v2/dupes/wapo"]).doc_ids else: with dlc["v3/dupes"].stream() as fin: duplicates = set(dupe_id.decode().strip() for dupe_id in fin) feeder = Feeder(name, jsonlines) logging.info("Starting...") for doc in core_iter: if doc.doc_id in duplicates: continue body = doc.passages[0] all_ranges = find_ranges(doc.doc_id, body, feeder.get(doc.doc_id)) computer = md5() for ranges in all_ranges: computer.update(b"\x00") for start, end in ranges: computer.update(b"\x01") computer.update(body[start:end].encode("utf-8")) if doc.doc_id in post_duplicates: # Ignore our work.... arggg continue print( json.dumps( { "id": doc.doc_id, "ranges": all_ranges, "md5": computer.digest().hex(), }, indent=None, ) ) logging.info("Finished...") if __name__ == "__main__": cli()