search

egerber / spaCy-entity-linker

spaCy module for linking text to Wikidata items
MIT License
215 stars 32 forks source link

readme

Tests Downloads Current Release Version pypi Version

Spacy Entity Linker

Introduction

Spacy Entity Linker is a pipeline for spaCy that performs Linked Entity Extraction with Wikidata on a given Document. The Entity Linking System operates by matching potential candidates from each sentence (subject, object, prepositional phrase, compounds, etc.) to aliases from Wikidata. The package allows to easily find the category behind each entity (e.g. "banana" is type "food" OR "Microsoft" is type "company"). It can is therefore useful for information extraction tasks and labeling tasks.

The package was written before a working Linked Entity Solution existed inside spaCy. In comparison to spaCy's linked entity system, it has the following advantages:

It also comes along with a number of disadvantages:

Installation

To install the package, run:

pip install spacy-entity-linker

Afterwards, the knowledge base (Wikidata) must be downloaded. This can be either be done by manually calling

python -m spacy_entity_linker "download_knowledge_base"

or when you first access the entity linker through spacy. This will download and extract a ~1.3GB file that contains a preprocessed version of Wikidata.

Use

import spacy  # version 3.5

# initialize language model
nlp = spacy.load("en_core_web_md")

# add pipeline (declared through entry_points in setup.py)
nlp.add_pipe("entityLinker", last=True)

doc = nlp("I watched the Pirates of the Caribbean last silvester")

# returns all entities in the whole document
all_linked_entities = doc._.linkedEntities
# iterates over sentences and prints linked entities
for sent in doc.sents:
    sent._.linkedEntities.pretty_print()

# OUTPUT:
# https://www.wikidata.org/wiki/Q194318     Pirates of the Caribbean        Series of fantasy adventure films                                                                   
# https://www.wikidata.org/wiki/Q12525597   Silvester                       the day celebrated on 31 December (Roman Catholic Church) or 2 January (Eastern Orthodox Churches)  

# entities are also directly accessible through spans
doc[3:7]._.linkedEntities.pretty_print()
# OUTPUT:
# https://www.wikidata.org/wiki/Q194318     Pirates of the Caribbean        Series of fantasy adventure films

EntityCollection

contains an array of entity elements. It can be accessed like an array but also implements the following helper functions:

doc = nlp("Elon Musk was born in South Africa. Bill Gates and Steve Jobs come from the United States")
doc._.linkedEntities.print_super_entities()
# OUTPUT:
# human (3) : Elon Musk,Bill Gates,Steve Jobs
# country (2) : South Africa,United States of America
# sovereign state (2) : South Africa,United States of America
# federal state (1) : United States of America
# constitutional republic (1) : United States of America
# democratic republic (1) : United States of America

EntityElement

each linked Entity is an object of type EntityElement. Each entity contains the methods

Usage of the get_span method with SpanInfo:

import spacy
nlp = spacy.load('en_core_web_md')
nlp.add_pipe("entityLinker", last=True)
text = 'Apple is competing with Microsoft.'
doc = nlp(text)
sents = list(doc.sents)
ent = doc._.linkedEntities[0]

# using the SpanInfo class
span = ent.get_span()
print(span.start, span.end, span.text) # behaves like a Span

# check equivalence
print(span == doc[0:1]) # True
print(doc[0:1] == span) # TypeError: Argument 'other' has incorrect type (expected spacy.tokens.span.Span, got SpanInfo)

# now get the real span
span = ent.get_span(doc) # passing the doc instance here
print(span.start, span.end, span.text)

print(span == doc[0:1]) # True
print(doc[0:1] == span) # True

Example

In the following example we will use SpacyEntityLinker to find find the mentioned Football Team in our text and explore other football teams of the same type


doc = nlp("I follow the New England Patriots")

patriots_entity = doc._.linkedEntities[0]
patriots_entity.pretty_print()
# OUTPUT:
# https://www.wikidata.org/wiki/Q193390     
# New England Patriots            
# National Football League franchise in Foxborough, Massachusetts                    

football_team_entity = patriots_entity.get_super_entities()[0]
football_team_entity.pretty_print()
# OUTPUT:
# https://www.wikidata.org/wiki/Q17156793 
# American football team          
# organization, in which a group of players are organized to compete as a team in American football   

for child in football_team_entity.get_sub_entities(limit=32):
    print(child)
    # OUTPUT:
    # New Orleans Saints
    # New York Giants
    # Pittsburgh Steelers
    # New England Patriots
    # Indianapolis Colts
    # Miami Seahawks
    # Dallas Cowboys
    # Chicago Bears
    # Washington Redskins
    # Green Bay Packers
    # ...

Entity Linking Policy

Currently the only method for choosing an entity given different possible matches (e.g. Paris - city vs Paris - firstname) is max-prior. This method achieves around 70% accuracy on predicting the correct entities behind link descriptions on wikipedia.

Note

The Entity Linker at the current state is still experimental and should not be used in production mode.

Performance

The current implementation supports only Sqlite. This is advantageous for development because it does not requirement any special setup and configuration. However, for more performance critical usecases, a different database with in-memory access (e.g. Redis) should be used. This may be implemented in the future.

Data

the knowledge base was derived from this dataset: https://www.kaggle.com/kenshoresearch/kensho-derived-wikimedia-data

It was cleaned and post-procesed, including filtering out entities of "overrepresented" categories such as

The purpose behind the knowledge base cleaning was to reduce the knowledge base size, while keeping the most useful entities for general purpose applications.

Currently, the only way to change the knowledge base is a bit hacky and requires to replace or modify the underlying sqlite database. You will find it under site_packages/data_spacy_entity_linker/wikidb_filtered.db. The database contains 3 tables:

Versions:

TODO