diff --git a/bigbio/hub/hub_repos/ppr/README.md b/bigbio/hub/hub_repos/ppr/README.md new file mode 100644 index 00000000..f0e96121 --- /dev/null +++ b/bigbio/hub/hub_repos/ppr/README.md @@ -0,0 +1,43 @@ + +--- +language: +- en +bigbio_language: +- English +license: unknown +multilinguality: monolingual +bigbio_license_shortname: UNKNOWN +pretty_name: PPR +homepage: https://github.com/DMCB-GIST/PPRcorpus +bigbio_pubmed: True +bigbio_public: True +bigbio_tasks: +- NAMED_ENTITY_RECOGNITION +- RELATION_EXTRACTION +--- + + +# Dataset Card for PPR + +## Dataset Description + +- **Homepage:** https://github.com/DMCB-GIST/PPRcorpus +- **Pubmed:** True +- **Public:** True +- **Tasks:** NER,RE + +The Plant-Phenotype corpus is a text corpus with human annotations of plants, phenotypes, and their relations on a corpus in 600 PubMed abstracts. + +## Citation Information + +``` +@article{cho2022plant, + author = {Cho, Hyejin and Kim, Baeksoo and Choi, Wonjun and Lee, Doheon and Lee, Hyunju}, + title = {Plant phenotype relationship corpus for biomedical relationships between plants and phenotypes}, + journal = {Scientific Data}, + volume = {9}, + year = {2022}, + publisher = {Nature Publishing Group}, + doi = {https://doi.org/10.1038/s41597-022-01350-1}, +} +``` diff --git a/bigbio/hub/hub_repos/ppr/bigbiohub.py b/bigbio/hub/hub_repos/ppr/bigbiohub.py new file mode 100644 index 00000000..a4792b4b --- /dev/null +++ b/bigbio/hub/hub_repos/ppr/bigbiohub.py @@ -0,0 +1,592 @@ +from collections import defaultdict +from dataclasses import dataclass +from enum import Enum +import logging +from pathlib import Path +from types import SimpleNamespace +from typing import TYPE_CHECKING, Dict, Iterable, List, Tuple + +import datasets + +if TYPE_CHECKING: + import bioc + +logger = logging.getLogger(__name__) + + +BigBioValues = SimpleNamespace(NULL="") + + +@dataclass +class BigBioConfig(datasets.BuilderConfig): + """BuilderConfig for BigBio.""" + + name: str = None + version: datasets.Version = None + description: str = None + schema: str = None + subset_id: str = None + + +class Tasks(Enum): + NAMED_ENTITY_RECOGNITION = "NER" + NAMED_ENTITY_DISAMBIGUATION = "NED" + EVENT_EXTRACTION = "EE" + RELATION_EXTRACTION = "RE" + COREFERENCE_RESOLUTION = "COREF" + QUESTION_ANSWERING = "QA" + TEXTUAL_ENTAILMENT = "TE" + SEMANTIC_SIMILARITY = "STS" + TEXT_PAIRS_CLASSIFICATION = "TXT2CLASS" + PARAPHRASING = "PARA" + TRANSLATION = "TRANSL" + SUMMARIZATION = "SUM" + TEXT_CLASSIFICATION = "TXTCLASS" + + +entailment_features = datasets.Features( + { + "id": datasets.Value("string"), + "premise": datasets.Value("string"), + "hypothesis": datasets.Value("string"), + "label": datasets.Value("string"), + } +) + +pairs_features = datasets.Features( + { + "id": datasets.Value("string"), + "document_id": datasets.Value("string"), + "text_1": datasets.Value("string"), + "text_2": datasets.Value("string"), + "label": datasets.Value("string"), + } +) + +qa_features = datasets.Features( + { + "id": datasets.Value("string"), + "question_id": datasets.Value("string"), + "document_id": datasets.Value("string"), + "question": datasets.Value("string"), + "type": datasets.Value("string"), + "choices": [datasets.Value("string")], + "context": datasets.Value("string"), + "answer": datasets.Sequence(datasets.Value("string")), + } +) + +text_features = datasets.Features( + { + "id": datasets.Value("string"), + "document_id": datasets.Value("string"), + "text": datasets.Value("string"), + "labels": [datasets.Value("string")], + } +) + +text2text_features = datasets.Features( + { + "id": datasets.Value("string"), + "document_id": datasets.Value("string"), + "text_1": datasets.Value("string"), + "text_2": datasets.Value("string"), + "text_1_name": datasets.Value("string"), + "text_2_name": datasets.Value("string"), + } +) + +kb_features = datasets.Features( + { + "id": datasets.Value("string"), + "document_id": datasets.Value("string"), + "passages": [ + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "text": datasets.Sequence(datasets.Value("string")), + "offsets": datasets.Sequence([datasets.Value("int32")]), + } + ], + "entities": [ + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "text": datasets.Sequence(datasets.Value("string")), + "offsets": datasets.Sequence([datasets.Value("int32")]), + "normalized": [ + { + "db_name": datasets.Value("string"), + "db_id": datasets.Value("string"), + } + ], + } + ], + "events": [ + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + # refers to the text_bound_annotation of the trigger + "trigger": { + "text": datasets.Sequence(datasets.Value("string")), + "offsets": datasets.Sequence([datasets.Value("int32")]), + }, + "arguments": [ + { + "role": datasets.Value("string"), + "ref_id": datasets.Value("string"), + } + ], + } + ], + "coreferences": [ + { + "id": datasets.Value("string"), + "entity_ids": datasets.Sequence(datasets.Value("string")), + } + ], + "relations": [ + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "arg1_id": datasets.Value("string"), + "arg2_id": datasets.Value("string"), + "normalized": [ + { + "db_name": datasets.Value("string"), + "db_id": datasets.Value("string"), + } + ], + } + ], + } +) + + +TASK_TO_SCHEMA = { + Tasks.NAMED_ENTITY_RECOGNITION.name: "KB", + Tasks.NAMED_ENTITY_DISAMBIGUATION.name: "KB", + Tasks.EVENT_EXTRACTION.name: "KB", + Tasks.RELATION_EXTRACTION.name: "KB", + Tasks.COREFERENCE_RESOLUTION.name: "KB", + Tasks.QUESTION_ANSWERING.name: "QA", + Tasks.TEXTUAL_ENTAILMENT.name: "TE", + Tasks.SEMANTIC_SIMILARITY.name: "PAIRS", + Tasks.TEXT_PAIRS_CLASSIFICATION.name: "PAIRS", + Tasks.PARAPHRASING.name: "T2T", + Tasks.TRANSLATION.name: "T2T", + Tasks.SUMMARIZATION.name: "T2T", + Tasks.TEXT_CLASSIFICATION.name: "TEXT", +} + +SCHEMA_TO_TASKS = defaultdict(set) +for task, schema in TASK_TO_SCHEMA.items(): + SCHEMA_TO_TASKS[schema].add(task) +SCHEMA_TO_TASKS = dict(SCHEMA_TO_TASKS) + +VALID_TASKS = set(TASK_TO_SCHEMA.keys()) +VALID_SCHEMAS = set(TASK_TO_SCHEMA.values()) + +SCHEMA_TO_FEATURES = { + "KB": kb_features, + "QA": qa_features, + "TE": entailment_features, + "T2T": text2text_features, + "TEXT": text_features, + "PAIRS": pairs_features, +} + + +def get_texts_and_offsets_from_bioc_ann(ann: "bioc.BioCAnnotation") -> Tuple: + + offsets = [(loc.offset, loc.offset + loc.length) for loc in ann.locations] + + text = ann.text + + if len(offsets) > 1: + i = 0 + texts = [] + for start, end in offsets: + chunk_len = end - start + texts.append(text[i : chunk_len + i]) + i += chunk_len + while i < len(text) and text[i] == " ": + i += 1 + else: + texts = [text] + + return offsets, texts + + +def remove_prefix(a: str, prefix: str) -> str: + if a.startswith(prefix): + a = a[len(prefix) :] + return a + + +def parse_brat_file( + txt_file: Path, + annotation_file_suffixes: List[str] = None, + parse_notes: bool = False, +) -> Dict: + """ + Parse a brat file into the schema defined below. + `txt_file` should be the path to the brat '.txt' file you want to parse, e.g. 'data/1234.txt' + Assumes that the annotations are contained in one or more of the corresponding '.a1', '.a2' or '.ann' files, + e.g. 'data/1234.ann' or 'data/1234.a1' and 'data/1234.a2'. + Will include annotator notes, when `parse_notes == True`. + brat_features = datasets.Features( + { + "id": datasets.Value("string"), + "document_id": datasets.Value("string"), + "text": datasets.Value("string"), + "text_bound_annotations": [ # T line in brat, e.g. type or event trigger + { + "offsets": datasets.Sequence([datasets.Value("int32")]), + "text": datasets.Sequence(datasets.Value("string")), + "type": datasets.Value("string"), + "id": datasets.Value("string"), + } + ], + "events": [ # E line in brat + { + "trigger": datasets.Value( + "string" + ), # refers to the text_bound_annotation of the trigger, + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "arguments": datasets.Sequence( + { + "role": datasets.Value("string"), + "ref_id": datasets.Value("string"), + } + ), + } + ], + "relations": [ # R line in brat + { + "id": datasets.Value("string"), + "head": { + "ref_id": datasets.Value("string"), + "role": datasets.Value("string"), + }, + "tail": { + "ref_id": datasets.Value("string"), + "role": datasets.Value("string"), + }, + "type": datasets.Value("string"), + } + ], + "equivalences": [ # Equiv line in brat + { + "id": datasets.Value("string"), + "ref_ids": datasets.Sequence(datasets.Value("string")), + } + ], + "attributes": [ # M or A lines in brat + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "ref_id": datasets.Value("string"), + "value": datasets.Value("string"), + } + ], + "normalizations": [ # N lines in brat + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "ref_id": datasets.Value("string"), + "resource_name": datasets.Value( + "string" + ), # Name of the resource, e.g. "Wikipedia" + "cuid": datasets.Value( + "string" + ), # ID in the resource, e.g. 534366 + "text": datasets.Value( + "string" + ), # Human readable description/name of the entity, e.g. "Barack Obama" + } + ], + ### OPTIONAL: Only included when `parse_notes == True` + "notes": [ # # lines in brat + { + "id": datasets.Value("string"), + "type": datasets.Value("string"), + "ref_id": datasets.Value("string"), + "text": datasets.Value("string"), + } + ], + }, + ) + """ + + example = {} + example["document_id"] = txt_file.with_suffix("").name + with txt_file.open() as f: + example["text"] = f.read() + + # If no specific suffixes of the to-be-read annotation files are given - take standard suffixes + # for event extraction + if annotation_file_suffixes is None: + annotation_file_suffixes = [".a1", ".a2", ".ann"] + + if len(annotation_file_suffixes) == 0: + raise AssertionError( + "At least one suffix for the to-be-read annotation files should be given!" + ) + + ann_lines = [] + for suffix in annotation_file_suffixes: + annotation_file = txt_file.with_suffix(suffix) + try: + with annotation_file.open() as f: + ann_lines.extend(f.readlines()) + except Exception: + continue + + example["text_bound_annotations"] = [] + example["events"] = [] + example["relations"] = [] + example["equivalences"] = [] + example["attributes"] = [] + example["normalizations"] = [] + + if parse_notes: + example["notes"] = [] + + for line in ann_lines: + line = line.strip() + if not line: + continue + + if line.startswith("T"): # Text bound + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + ann["type"] = fields[1].split()[0] + ann["offsets"] = [] + span_str = remove_prefix(fields[1], (ann["type"] + " ")) + text = fields[2] + for span in span_str.split(";"): + start, end = span.split() + ann["offsets"].append([int(start), int(end)]) + + # Heuristically split text of discontiguous entities into chunks + ann["text"] = [] + if len(ann["offsets"]) > 1: + i = 0 + for start, end in ann["offsets"]: + chunk_len = end - start + ann["text"].append(text[i : chunk_len + i]) + i += chunk_len + while i < len(text) and text[i] == " ": + i += 1 + else: + ann["text"] = [text] + + example["text_bound_annotations"].append(ann) + + elif line.startswith("E"): + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + + ann["type"], ann["trigger"] = fields[1].split()[0].split(":") + + ann["arguments"] = [] + for role_ref_id in fields[1].split()[1:]: + argument = { + "role": (role_ref_id.split(":"))[0], + "ref_id": (role_ref_id.split(":"))[1], + } + ann["arguments"].append(argument) + + example["events"].append(ann) + + elif line.startswith("R"): + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + ann["type"] = fields[1].split()[0] + + ann["head"] = { + "role": fields[1].split()[1].split(":")[0], + "ref_id": fields[1].split()[1].split(":")[1], + } + ann["tail"] = { + "role": fields[1].split()[2].split(":")[0], + "ref_id": fields[1].split()[2].split(":")[1], + } + + example["relations"].append(ann) + + # '*' seems to be the legacy way to mark equivalences, + # but I couldn't find any info on the current way + # this might have to be adapted dependent on the brat version + # of the annotation + elif line.startswith("*"): + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + ann["ref_ids"] = fields[1].split()[1:] + + example["equivalences"].append(ann) + + elif line.startswith("A") or line.startswith("M"): + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + + info = fields[1].split() + ann["type"] = info[0] + ann["ref_id"] = info[1] + + if len(info) > 2: + ann["value"] = info[2] + else: + ann["value"] = "" + + example["attributes"].append(ann) + + elif line.startswith("N"): + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + ann["text"] = fields[2] + + info = fields[1].split() + + ann["type"] = info[0] + ann["ref_id"] = info[1] + ann["resource_name"] = info[2].split(":")[0] + ann["cuid"] = info[2].split(":")[1] + example["normalizations"].append(ann) + + elif parse_notes and line.startswith("#"): + ann = {} + fields = line.split("\t") + + ann["id"] = fields[0] + ann["text"] = fields[2] if len(fields) == 3 else BigBioValues.NULL + + info = fields[1].split() + + ann["type"] = info[0] + ann["ref_id"] = info[1] + example["notes"].append(ann) + + return example + + +def brat_parse_to_bigbio_kb(brat_parse: Dict) -> Dict: + """ + Transform a brat parse (conforming to the standard brat schema) obtained with + `parse_brat_file` into a dictionary conforming to the `bigbio-kb` schema (as defined in ../schemas/kb.py) + :param brat_parse: + """ + + unified_example = {} + + # Prefix all ids with document id to ensure global uniqueness, + # because brat ids are only unique within their document + id_prefix = brat_parse["document_id"] + "_" + + # identical + unified_example["document_id"] = brat_parse["document_id"] + unified_example["passages"] = [ + { + "id": id_prefix + "_text", + "type": "abstract", + "text": [brat_parse["text"]], + "offsets": [[0, len(brat_parse["text"])]], + } + ] + + # get normalizations + ref_id_to_normalizations = defaultdict(list) + for normalization in brat_parse["normalizations"]: + ref_id_to_normalizations[normalization["ref_id"]].append( + { + "db_name": normalization["resource_name"], + "db_id": normalization["cuid"], + } + ) + + # separate entities and event triggers + unified_example["events"] = [] + non_event_ann = brat_parse["text_bound_annotations"].copy() + for event in brat_parse["events"]: + event = event.copy() + event["id"] = id_prefix + event["id"] + trigger = next( + tr + for tr in brat_parse["text_bound_annotations"] + if tr["id"] == event["trigger"] + ) + if trigger in non_event_ann: + non_event_ann.remove(trigger) + event["trigger"] = { + "text": trigger["text"].copy(), + "offsets": trigger["offsets"].copy(), + } + for argument in event["arguments"]: + argument["ref_id"] = id_prefix + argument["ref_id"] + + unified_example["events"].append(event) + + unified_example["entities"] = [] + anno_ids = [ref_id["id"] for ref_id in non_event_ann] + for ann in non_event_ann: + entity_ann = ann.copy() + entity_ann["id"] = id_prefix + entity_ann["id"] + entity_ann["normalized"] = ref_id_to_normalizations[ann["id"]] + unified_example["entities"].append(entity_ann) + + # massage relations + unified_example["relations"] = [] + skipped_relations = set() + for ann in brat_parse["relations"]: + if ( + ann["head"]["ref_id"] not in anno_ids + or ann["tail"]["ref_id"] not in anno_ids + ): + skipped_relations.add(ann["id"]) + continue + unified_example["relations"].append( + { + "arg1_id": id_prefix + ann["head"]["ref_id"], + "arg2_id": id_prefix + ann["tail"]["ref_id"], + "id": id_prefix + ann["id"], + "type": ann["type"], + "normalized": [], + } + ) + if len(skipped_relations) > 0: + example_id = brat_parse["document_id"] + logger.info( + f"Example:{example_id}: The `bigbio_kb` schema allows `relations` only between entities." + f" Skip (for now): " + f"{list(skipped_relations)}" + ) + + # get coreferences + unified_example["coreferences"] = [] + for i, ann in enumerate(brat_parse["equivalences"], start=1): + is_entity_cluster = True + for ref_id in ann["ref_ids"]: + if not ref_id.startswith("T"): # not textbound -> no entity + is_entity_cluster = False + elif ref_id not in anno_ids: # event trigger -> no entity + is_entity_cluster = False + if is_entity_cluster: + entity_ids = [id_prefix + i for i in ann["ref_ids"]] + unified_example["coreferences"].append( + {"id": id_prefix + str(i), "entity_ids": entity_ids} + ) + return unified_example diff --git a/bigbio/hub/hub_repos/ppr/ppr.py b/bigbio/hub/hub_repos/ppr/ppr.py new file mode 100644 index 00000000..ac1d1be1 --- /dev/null +++ b/bigbio/hub/hub_repos/ppr/ppr.py @@ -0,0 +1,389 @@ +# coding=utf-8 +# Copyright 2022 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. + +import itertools as it +from typing import Dict, Generator, List, Tuple + +import datasets + +from .bigbiohub import BigBioConfig, Tasks, kb_features + +_LANGUAGES = ["English"] +_PUBMED = True +_LOCAL = False + +_CITATION = """\ +@article{cho2022plant, + author = {Cho, Hyejin and Kim, Baeksoo and Choi, Wonjun and Lee, Doheon and Lee, Hyunju}, + title = {Plant phenotype relationship corpus for biomedical relationships between plants and phenotypes}, + journal = {Scientific Data}, + volume = {9}, + year = {2022}, + publisher = {Nature Publishing Group}, + doi = {https://doi.org/10.1038/s41597-022-01350-1}, +} +""" + +_DATASETNAME = "ppr" +_DISPLAYNAME = "Plant-Phenotype-Relations" + +_DESCRIPTION = """\ +The Plant-Phenotype corpus is a text corpus with human annotations of plants, phenotypes, and their relations on a \ +corpus in 600 PubMed abstracts. +""" + +_HOMEPAGE = "https://github.com/DMCB-GIST/PPRcorpus" + +_LICENSE = "UNKNOWN" + +_URLS = { + _DATASETNAME: [ + "https://raw.githubusercontent.com/davidkartchner/PPRcorpus/main/corpus/PPR_train_corpus.txt", + "https://raw.githubusercontent.com/davidkartchner/PPRcorpus/main/corpus/PPR_dev_corpus.txt", + "https://raw.githubusercontent.com/davidkartchner/PPRcorpus/main/corpus/PPR_test_corpus.txt", + ], +} + +_SUPPORTED_TASKS = [Tasks.NAMED_ENTITY_RECOGNITION, Tasks.RELATION_EXTRACTION] + +_SOURCE_VERSION = "1.0.0" +_BIGBIO_VERSION = "1.0.0" + + +class PlantPhenotypeDataset(datasets.GeneratorBasedBuilder): + """Plant-Phenotype is dataset for NER and RE of plants and their induced phenotypes""" + + SOURCE_VERSION = datasets.Version(_SOURCE_VERSION) + BIGBIO_VERSION = datasets.Version(_BIGBIO_VERSION) + + BUILDER_CONFIGS = [ + BigBioConfig( + name="ppr_source", + version=SOURCE_VERSION, + description="Plant Phenotype Relations source schema", + schema="source", + subset_id="plant_phenotype", + ), + BigBioConfig( + name="ppr_bigbio_kb", + version=BIGBIO_VERSION, + description="Plant Phenotype Relations BigBio schema", + schema="bigbio_kb", + subset_id="plant_phenotype", + ), + ] + + DEFAULT_CONFIG_NAME = "ppr_source" + + def _info(self) -> datasets.DatasetInfo: + if self.config.schema == "source": + + features = datasets.Features( + { + "passage_id": datasets.Value("string"), + "pmid": datasets.Value("string"), + "section": datasets.Value("int32"), + "text": datasets.Value("string"), + "entities": [ + { + "offsets": datasets.Sequence(datasets.Value("int32")), + "text": datasets.Value("string"), + "type": datasets.Value("string"), + } + ], + "relations": [ + { + "relation_type": datasets.Value("string"), + "entity1_offsets": datasets.Sequence(datasets.Value("int32")), + "entity1_text": datasets.Value("string"), + "entity1_type": datasets.Value("string"), + "entity2_offsets": datasets.Sequence(datasets.Value("int32")), + "entity2_text": datasets.Value("string"), + "entity2_type": datasets.Value("string"), + } + ], + } + ) + + elif self.config.schema == "bigbio_kb": + features = kb_features + else: + raise NotImplementedError(f"Schema {self.config.schema} not supported") + + return datasets.DatasetInfo( + description=_DESCRIPTION, + features=features, + homepage=_HOMEPAGE, + license=_LICENSE, + citation=_CITATION, + ) + + def _split_generators(self, dl_manager) -> List[datasets.SplitGenerator]: + """Returns SplitGenerators.""" + + urls = _URLS[_DATASETNAME] + train, dev, test = dl_manager.download_and_extract(urls) + + return [ + datasets.SplitGenerator( + name=datasets.Split.TRAIN, + gen_kwargs={ + "filepath": train, + }, + ), + datasets.SplitGenerator( + name=datasets.Split.TEST, + gen_kwargs={ + "filepath": test, + }, + ), + datasets.SplitGenerator( + name=datasets.Split.VALIDATION, + gen_kwargs={ + "filepath": dev, + }, + ), + ] + + def _generate_examples( + self, + filepath, + ) -> Tuple[int, Dict]: + """Yields examples as (key, example) tuples.""" + + with open(filepath, "r") as f: + chunks = f.read().strip().split("\n\n") + + if self.config.schema == "source": + for id_, doc in self._generate_source_examples(chunks): + yield id_, doc + + elif self.config.schema == "bigbio_kb": + for id_, doc in self._generate_bigbio_kb_examples(chunks): + yield id_, doc + + def _generate_whole_documents(self, annotation_chunks: List[str]) -> Generator[Dict, None, None]: + """Aggregate individual sentence annotations into whole abstracts. + + Args: + annotation_chunks (List[str]): List of annotation chunks, i.e., a sentence with its annotations. + For example: + 10072339_4 OBJECTIVE: A patient with possible airborne facial dermatitis to potato is described. + 10072339 44 61 facial dermatitis Negative_phenotype + 10072339 65 71 potato Plant + + Returns: + Generator producing a dictionary containing the pmid of an article and all document chunks. + """ + prev_pmid = None + pmid = "" + doc_chunks = [] + + for chunk in annotation_chunks: + lines = chunk.split("\n") + + # The first line is the sentence (format: _\t) + passage_info, passage_text = lines[0].split("\t") + + # Then annotations in Pubtator format + annotations = [line.split("\t") for line in lines[1:]] + + # Get info on passage + pmid, section = passage_info.split("_") + if prev_pmid is None: + prev_pmid = pmid + + elif prev_pmid != pmid: + yield {"pmid": prev_pmid, "doc_chunks": doc_chunks} + + # Reset everything for next PMID + prev_pmid = pmid + doc_chunks = [] + + doc_chunks.append( + { + "passage": passage_text, + "annotations": annotations, + "sentence_id": passage_info, + } + ) + + # Take care of last document + yield {"pmid": pmid, "doc_chunks": doc_chunks} + + def _generate_source_examples(self, annotation_chunks: List[str]) -> Generator[Tuple[str, Dict], None, None]: + """Generate examples in format of source schema + + Args: + annotation_chunks (List[str]): List of annotation chunks. + + Returns: + Generator of instance tuples (, ) + """ + for chunk in annotation_chunks: + lines = chunk.split("\n") + + passage_id, passage_text = lines[0].split("\t") + annotations = [line.split("\t") for line in lines[1:]] + + # Get info on passage + pmid, section = passage_id.split("_") + section = int(section) + + # Grab entities and relations + entities = [] + relations = [] + for annotation in annotations: + if len(annotation) == 5: + # It's an entity annotation + entities.append( + { + "offsets": (int(annotation[1]), int(annotation[2])), + "text": annotation[3], + "type": annotation[4], + } + ) + + elif len(annotation) == 10: + # Relation annotation + relations.append( + { + "relation_type": annotation[1], + "entity1_offsets": (int(annotation[2]), int(annotation[3])), + "entity1_text": annotation[4], + "entity1_type": annotation[5], + "entity2_offsets": (int(annotation[6]), int(annotation[7])), + "entity2_text": annotation[8], + "entity2_type": annotation[9], + } + ) + else: + # This is a special case that occurs for a single data point + relations.append( + { + "relation_type": annotation[1], + "entity1_offsets": (int(annotation[2]), int(annotation[3])), + "entity1_text": annotation[4], + "entity1_type": annotation[5], + "entity2_offsets": (int(annotation[8]), int(annotation[9])), + "entity2_text": annotation[10], + "entity2_type": annotation[11], + } + ) + + # Consolidate into document + document = { + "passage_id": passage_id, + "pmid": pmid, + "section": section, + "text": passage_text, + "entities": entities, + "relations": relations, + } + + yield passage_id, document + + def _generate_bigbio_kb_examples(self, annotation_chunks: List[str]): + """Generator for training examples in bigbio_kb schema format. + + Args: + annotation_chunks (List[str]): List of annotation chunks. + + Returns: + Generator of instance tuples (, ) + """ + uid = it.count(1) + for document in self._generate_whole_documents(annotation_chunks): + pmid = document["pmid"] + offset_delta = 0 + id_ = str(next(uid)) + + passages = [] + entities = [] + relations = [] + + # Iterate through each section of the article + for text_section in document["doc_chunks"]: + # Extract passages + passage = text_section["passage"] + passages.append( + { + "id": str(next(uid)), + "text": [passage], + "type": "sentence", + "offsets": [(offset_delta, offset_delta + len(passage))], + } + ) + + # Extract entities + entities_sublist = [] + for annotation in text_section["annotations"]: + if len(annotation) == 5: + entities_sublist.append( + { + "id": str(next(uid)), + "type": annotation[4], + "text": [annotation[3]], + "offsets": [(int(annotation[1]) + offset_delta, int(annotation[2]) + offset_delta)], + "normalized": [], + } + ) + + # Create mapping of offsets to entity_id + ent2id = {tuple(x["offsets"]): x["id"] for x in entities_sublist} + entities.extend(entities_sublist) + + # Extract relations + for annotation in text_section["annotations"]: + if len(annotation) == 10: + e1_offsets = [(int(annotation[2]) + offset_delta, int(annotation[3]) + offset_delta)] + e2_offsets = [(int(annotation[6]) + offset_delta, int(annotation[7]) + offset_delta)] + relations.append( + { + "id": str(next(uid)), + "type": annotation[1], + "arg1_id": ent2id[tuple(e1_offsets)], + "arg2_id": ent2id[tuple(e2_offsets)], + "normalized": [], + } + ) + + # Special case for a single annotation + elif len(annotation) > 10: + e1_offsets = [(int(annotation[2]) + offset_delta, int(annotation[3]) + offset_delta)] + e2_offsets = [(int(annotation[8]) + offset_delta, int(annotation[9]) + offset_delta)] + relations.append( + { + "id": str(next(uid)), + "type": annotation[1], + "arg1_id": ent2id[tuple(e1_offsets)], + "arg2_id": ent2id[tuple(e2_offsets)], + "normalized": [], + } + ) + + offset_delta += len(passage) + 1 + + doc = { + "id": id_, + "document_id": pmid, + "passages": passages, + "entities": entities, + "relations": relations, + "events": [], + "coreferences": [], + } + + yield id_, doc