Dependency graph labels appear upside down

[urigoren]

How to reproduce the behaviour

The following code generates a dependecy graph visualisation, with some of the arc labels shown upside down.

See screenshot

from spacy import displacy

viz = {'words': [{'text': 'Balmer', 'tag': ''},
{'text': 'replaces', 'tag': ''},
{'text': 'Gates', 'tag': ''},
{'text': 'at', 'tag': ''},
{'text': 'Microsoft', 'tag': ''}],
'arcs': [{'start': 0, 'end': 1, 'dir': 'left', 'label': 'nsubj'},
{'start': 1, 'end': -1, 'dir': 'right', 'label': 'root'},
{'start': 2, 'end': 1, 'dir': 'right', 'label': 'obj'},
{'start': 3, 'end': 4, 'dir': 'left', 'label': 'case'},
{'start': 4, 'end': 1, 'dir': 'right', 'label': 'obl'}]}

html = displacy.render(viz,
jupyter=False,
style="dep",
manual=True,
options={"compact": True,"color": "blue",}
)
with open("bug.html", 'w') as f:
f.write(html)

Your Environment

• Operating System: Win10
• Python Version Used: 3.6.7
• spaCy Version Used: spacy==2.1.4
• Environment Information:

[ines]

Sorry for only getting to this now. I initially thought this was a bug, but I think it's actually related to how the data is formatted. In your arcs, you're using relative indices and things like 'end': -1. displaCy however expects absolute indices.

For instance, here's the result displaCy would internally produce for that parse:

import spacy
from spacy.displacy import parse_deps

nlp = spacy.load("en_core_web_sm")
doc = nlp("Balmer replaces Gates at Microsoft")
print(parse_deps(doc))

{'words': [{'text': 'Balmer', 'tag': 'PROPN'},
  {'text': 'replaces', 'tag': 'VERB'},
  {'text': 'Gates', 'tag': 'PROPN'},
  {'text': 'at', 'tag': 'ADP'},
  {'text': 'Microsoft', 'tag': 'PROPN'}],
 'arcs': [{'start': 0, 'end': 1, 'label': 'nsubj', 'dir': 'left'},
  {'start': 1, 'end': 2, 'label': 'dobj', 'dir': 'right'},
  {'start': 2, 'end': 3, 'label': 'prep', 'dir': 'right'},
  {'start': 3, 'end': 4, 'label': 'pobj', 'dir': 'right'}],
 'settings': {'lang': 'en', 'direction': 'ltr'}}

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.