EventGraph is a development platform for high-level NLP applications in Japanese. The core concept of EventGraph is event, a language information unit that is closely related to predicate-argument structure but more application-oriented. Events are linked to each other based on their syntactic and semantic relations.
- Python 3.6 or later
- pyknp
- graphviz
To install pyknp-eventgraph, use pip.
$ pip install pyknp-eventgraph
An EventGraph is built on language analysis given in a KNP format.
# Add imports.
from pyknp import KNP
from pyknp_eventgraph import EventGraph
# Parse a document.
document = ['彼女は海外勤務が長いので、英語がうまいに違いない。', '私はそう確信していた。']
knp = KNP()
analysis = [knp.parse(sentence) for sentence in document]
# Create an EventGraph.
evg = EventGraph.build(analysis)
print(evg) # <EventGraph, #sentences: 2, #events: 3, #relations: 1>Users can obtain various information about language analysis via a simple interface.
# Extract sentences.
sentences = evg.sentences
print(sentences)
# [
# <Sentence, sid: 1, ssid: 0, surf: 彼女は海外勤務が長いので、英語がうまいに違いない。>,
# <Sentence, sid: 2, ssid: 1, surf: 私はそう確信していた。>
# ]
# Convert a sentence into various forms.
sentence = evg.sentences[0]
print(sentence.surf) # 彼女は海外勤務が長いので、英語がうまいに違いない。
print(sentence.mrphs) # 彼女 は 海外 勤務 が 長い ので 、 英語 が うまい に 違いない 。
print(sentence.reps) # 彼女/かのじょ は/は 海外/かいがい 勤務/きんむ が/が 長い/ながい ので/ので 、/、 英語/えいご が/が 上手い/うまい に/に 違い無い/ちがいない 。/。# Extract events.
events = evg.events
print(events)
# [
# <Event, evid: 0, surf: 海外勤務が長いので、>,
# <Event, evid: 1, surf: 彼女は英語がうまいに違いない。>,
# <Event, evid: 2, surf: 私はそう確信していた。>
# ]
# Convert an event into various forms.
event = evg.events[0]
print(event.surf) # 海外勤務が長いので、
print(event.mrphs) # 海外 勤務 が 長い ので 、
print(event.normalized_mrphs) # 海外 勤務 が 長い
print(event.reps) # 海外/かいがい 勤務/きんむ が/が 長い/ながい ので/ので 、/、
print(event.normalized_reps) # 海外/かいがい 勤務/きんむ が/が 長い/ながい
print(event.content_rep_list) # ['海外/かいがい', '勤務/きんむ', '長い/ながい']
# Extract an event's PAS.
pas = event.pas
print(pas) # <PAS, predicate: 長い/ながい, arguments: {ガ: 勤務/きんむ}>
print(pas.predicate) # <Predicate, type: 形, surf: 長い>
print(pas.arguments) # defaultdict(<class 'list'>, {'ガ': [<Argument, case: ガ, surf: 勤務が>]})
# Extract an event's features.
features = event.features
print(features) # <Features, modality: None, tense: 非過去, negation: False, state: 状態述語, complement: False># Extract event-to-event relations.
relations = evg.relations
print(relations) # [<Relation, label: 原因・理由, modifier_evid: 0, head_evid: 1>]
# Take a closer look at an event-to-event relation
relation = relations[0]
print(relation.label) # 原因・理由
print(relation.surf) # ので
print(relation.modifier) # <Event, evid: 0, surf: 海外勤務が長いので、>
print(relation.head) # <Event, evid: 1, surf: 彼女は英語がうまいに違いない。>Users can save and load an EventGraph by serializing it as a JSON object.
# Save an EventGraph as a JSON file.
evg.save('evg.json')
# Load an EventGraph from a JSON file.
with open('evg.json') as f:
evg = EventGraph.load(f)Users can visualize an EventGraph using graphviz.
from pyknp_eventgraph import make_image
make_image(evg, 'evg.svg') # Currently, only supports 'svg'.By merging a modifier event to the modifiee, users can construct a larger information unit.
from pyknp import KNP
from pyknp_eventgraph import EventGraph
document = ['もっととろみが持続する作り方をして欲しい。']
knp = KNP()
analysis = [knp.parse(sentence) for sentence in document]
evg = EventGraph.build(analysis)
print(evg) # <EventGraph, #sentences: 1, #events: 2, #relations: 1>
# Investigate the relation.
relation = evg.relations[0]
print(relation) # <Relation, label: 連体修飾, modifier_evid: 0, head_evid: 1>
print(relation.modifier) # <Event, evid: 0, surf: もっととろみが持続する>
print(relation.head) # <Event, evid: 1, surf: 作り方をして欲しい。>
# To merge modifier events, enable `include_modifiers`.
print(relation.head.surf) # 作り方をして欲しい。
print(relation.head.surf_(include_modifiers=True)) # もっととろみが持続する作り方をして欲しい。
# Other formats also support `include_modifiers`.
print(relation.head.mrphs_(include_modifiers=True)) # もっと とろみ が 持続 する 作り 方 を して 欲しい 。
print(relation.head.normalized_mrphs_(include_modifiers=True)) # もっと とろみ が 持続 する 作り 方 を して 欲しいWhen an EventGraph is serialized in a JSON format, it will lose some functionality, including access to KNP objects and modifier merging. To keep full functionality, use Python's pickle utility for serialization.
# Save an EventGraph using Python's pickle utility.
evg.save('evg.pkl', binary=True)
# Load an EventGraph using Python's pickle utility.
with open('evg.pkl', 'rb') as f:
evg_ = EventGraph.load(f, binary=True)$ echo '彼女は海外勤務が長いので、英語がうまいに違いない。' | jumanpp | knp -tab | evg -o example-eventgraph.json
$ evgviz example-eventgraph.json example-eventgraph.svg
https://pyknp-eventgraph.readthedocs.io/en/latest/
- Kurohashi-Kawahara Lab, Kyoto University.
- [email protected]