extr-ds

Library to programmatically build labeled datasets for Named-Entity Recognition (NER) and Relation Extraction (RE) Machine Learning tasks.

Install

pip install extr-ds

Command Line

see Instructions on how to use the command line utility to manage your project.

1. Init Project

extr-ds --init

2. Split and Annotate

extr-ds --split

3.a Annotate Entities or Relations Again?

extr-ds --annotate -ents
extr-ds --annotate -rels

3.b Change Relation Extraction Label

extr-ds --relate -label NO_RELATION=5,7,9

3.b Remove Relation Extraction Instance

extr-ds --relate -delete 5,6,7

3.c Recover removed Relation Extraction Instances

extr-ds --relate -recover 5,6,7

4. Save

extr-ds --save -ents
extr-ds --save -rels

5. Reset "Gold Standard" datasets

extr-ds --reset

6. Help!?

extr-ds --help

API

Example

text = 'Ted Johnson is a pitcher.'

1. Label Entities for Named-Entity Recognition Task (NER)

from extr import RegEx, RegExLabel
from extr.entities import EntityExtactor
from extr_ds.labelers import IOB

entity_extractor = EntityExtactor([
    RegExLabel('PERSON', [
        RegEx([r'(ted\s+johnson|ted)'], re.IGNORECASE)
    ]),
    RegExLabel('POSITION', [
        RegEx([r'pitcher'], re.IGNORECASE)
    ]),
])

sentence_tokenizer = ## 3rd party tokenizer ##
label = IOB(sentence_tokenizer, entity_extractor).label(text)

## label == <Label tokens=..., labels=['B-PERSON', 'I-PERSON', 'O', 'O', 'B-POSITION', 'O']>

2. Annotate for Relation Extraction Task (RE)

from extr.entities import EntityExtractor
from extr.relations import RegExRelationLabelBuilder, \
                           RelationExtractor
from extr_ds.labelers import RelationClassification
from extr_ds.labelers.relation import RelationBuilder, BaseRelationLabeler, RuleBasedRelationLabeler


person_to_position_relationship = RegExRelationLabelBuilder('is_a') \
    .add_e1_to_e2(
        'PERSON',
        [
            r'\s+is\s+a\s+',
        ],
        'POSITION'
    ) \
    .build()

base_relation_labeler = BaseRelationLabeler(
    RelationBuilder(relation_formats=[
        ('PERSON', 'POSITION', 'NO_RELATION')
    ])
)

rule_based_relation_labeler = RuleBasedRelationLabeler(
    RelationExtractor([person_to_position_relationship])
)

labeler = RelationClassification(
    EntityExtractor([
        RegExLabel('PERSON', [
            RegEx([r'(ted johnson|bob)'], re.IGNORECASE)
        ]),
        RegExLabel('POSITION', [
            RegEx([r'pitcher'], re.IGNORECASE)
        ]),
    ]),
    base_relation_labeler,
    relation_labelers=[
        rule_based_relation_labeler
    ]
)

results = labeler.label(text)

## results.relation_labels == [
##    <RelationLabel sentence="<e1>Ted Johnson</e1> is a <e2>pitcher</e2>." label="is_a">
## ]

3. Find and define the type of difference between labels

from extr_ds.validators import check_for_differences

differences_in_labels = check_for_differences(
    ['B-PERSON', 'I-PERSON', 'O', 'O', 'B-POSITION', 'O'],
    ['B-PERSON', 'O', 'O', 'O', 'B-POSITION', 'O']
)

## differences_in_labels.has_diffs == True
## differences_in_labels.diffs_between_labels == [
##      <Difference index=1, diff_type=DifferenceTypes.S2_MISSING>
## ]

differences_in_labels = check_for_differences(
    ['B-PERSON', 'I-PERSON', 'O', 'O', 'B-POSITION', 'O'],
    ['B-PERSON', 'B-PERSON', 'O', 'O', 'B-POSITION', 'O']
)

## differences_in_labels.has_diffs == True
## differences_in_labels.diffs_between_labels == [
##      <Difference index=1, diff_type=DifferenceTypes.MISMATCH>
## ]