Almost-nonsense output with rnnmorph

[adamdrazsky]

Hi, I'm using rnnmorph on python3.9 to attach tags to the input to russian-g2p for improved accenting.
Output is very strange, multiple pluses as well as accents on top of letters. See below...

The code:

from russian_g2p.Accentor import Accentor
from rnnmorph.predictor import RNNMorphPredictor

accentor = Accentor()
predictor = RNNMorphPredictor(language="ru")

testing_corpus = [
    # testing дома
    'Я сегодня остался дома',  # до+ма
    'рабочие строят высокие бетонные дома',  # дома+
    # testing козы
    'На поляне пасутся козы',  # ко+зы
    'У моей козы сломана нога',  # козы+
    # testing уже
    'я уже пришёл из школы',  # уже+
    'эта юбка уже, чем полоска'  #у+же
] 

sentences_tagged = [predictor.predict(testing_sentence.replace(",", "").split(' ')) for testing_sentence in testing_corpus]

for sent in sentences_tagged:
  sent_tagged = [[word.word, f"{word.pos} {word.tag}"] for word in sent]
  sent_accented = [accentor.do_accents([[word[0], word[1]]]) for word in sent_tagged]
  print(sent_accented)

Output:

[['я+']], [['сего+дня']], [['оста+лся']], [['+д+о+́+м+а+']]]
[[['рабо+чие']], [['стро+ят']], [['высо+кие']], [['бето+нные']], [['дома+']]]
[[['на']], [['поля+не']], [['пасу+тся']], [['+к+о+́+з+ы+']]]
[[['у']], [['мое+й']], [['+к+о+з+ы+́+']], [['сло+мана']], [['нога+']]]
[[['я+']], [['у+же']], [['пришё+л']], [['из']], [['шко+лы']]]
[[['э+та']], [['ю+бка']], [['у+же']], [['че+м']], [['поло+ска']]]

As you can see, some words have multiple accent markers; I'm expecting only one per word at most.

What's going on here?

[adamdrazsky]

For quick testing, do:

accentor.do_accents([['дома', 'ADV Degree=Pos']])
output:
[['+д+о+́+м+а+']]
expected:
[['до+ма']]