spaCy EntityRuler

spaCyで学ぶNatural Language Processing

Azadeh Mobasher

Principal Data Scientist

spaCy EntityRuler

EntityRulerはDocに固有表現を追加
単体でもEntityRecognizerと併用でも可
厳密一致のフレーズパターン（文字列）:

{"label": "ORG", "pattern": "Microsoft"}

1トークン=1辞書のトークンパターン（リスト）:

{"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "francisco"}]}

spaCy パイプラインに EntityRuler を追加

.add_pipe()で追加
パターンは.add_patterns()で一括追加

nlp = spacy.blank("en")
entity_ruler = nlp.add_pipe("entity_ruler")
patterns = [{"label": "ORG", "pattern": "Microsoft"},
            {"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "francisco"}]}]
entity_ruler.add_patterns(patterns)

spaCy パイプラインに EntityRuler を追加

.entsにEntityLinkerの結果が入る

doc = nlp("Microsoft is hiring software developer in San Francisco.")
print([(ent.text, ent.label_) for ent in doc.ents])

[('Microsoft', 'ORG'), ('San Francisco', 'GPE')]

EntityRuler の実行例

spaCyの各コンポーネントと連携
固有表現抽出器を強化
EntityRulerなしのspaCyモデル:

nlp = spacy.load("en_core_web_sm")

doc = nlp("Manhattan associates is a company in the U.S.")
print([(ent.text, ent.label_) for ent in doc.ents])

>>> [('Manhattan', 'GPE'), ('U.S.', 'GPE')]

EntityRuler の実行例

既存のnerコンポーネントの後にEntityRulerを追加:

nlp = spacy.load("en_core_web_sm")
ruler = nlp.add_pipe("entity_ruler", after='ner')
patterns = [{"label": "ORG", "pattern": [{"lower": "manhattan"}, {"lower": "associates"}]}]
ruler.add_patterns(patterns)

doc = nlp("Manhattan associates is a company in the U.S.")
print([(ent.text, ent.label_) for ent in doc.ents])

>>> [('Manhattan', 'GPE'), ('U.S.', 'GPE')]

EntityRuler の実行例

既存のnerコンポーネントの前にEntityRulerを追加:

nlp = spacy.load("en_core_web_sm")
ruler = nlp.add_pipe("entity_ruler", before='ner')
patterns = [{"label": "ORG", "pattern": [{"lower": "manhattan"}, {"lower": "associates"}]}]
ruler.add_patterns(patterns)

doc = nlp("Manhattan associates is a company in the U.S.")
print([(ent.text, ent.label_) for ent in doc.ents])

>>> [('Manhattan associates', 'ORG'), ('U.S.', 'GPE')]

練習しましょう！

spaCyで学ぶNatural Language Processing