ValueError: [E177] Ill-formed IOB input detected #5753

sbushmanov · 2020-07-12T14:44:36Z

sbushmanov
Jul 12, 2020

When I try co convert simple conllu file (link to project page, link to full corpus, 2Gb) with ner the following error appears:

spacy convert -c 'ner' try.conllu > try.json

Traceback (most recent call last):
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/site-packages/spacy/__main__.py", line 33, in <module>
    plac.call(commands[command], sys.argv[1:])
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/site-packages/plac_core.py", line 328, in call
    cmd, result = parser.consume(arglist)
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/site-packages/plac_core.py", line 207, in consume
    return cmd, self.func(*(args + varargs + extraopts), **kwargs)
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/site-packages/spacy/cli/convert.py", line 106, in convert
    no_print=no_print,
  File "/home/sergey/anaconda3/envs/rasa/lib/python3.7/site-packages/spacy/cli/converters/conll_ner2json.py", line 111, in conll_ner2json
    biluo_ents = iob_to_biluo(iob_ents)
  File "gold.pyx", line 570, in spacy.gold.iob_to_biluo
  File "gold.pyx", line 592, in spacy.gold._consume_ent
ValueError: [E177] Ill-formed IOB input detected: =

The simplified source file is:

# newdoc id = 0
# sent_id = 0_0
# text = Вице-премьер по социальным вопросам Татьяна Голикова рассказала, в каких регионах России зафиксирована наиболее высокая смертность от рака, сообщает РИА Новости.
1	Вице-премьер	_	NOUN	_	Animacy=Anim|Case=Nom|Gender=Masc|Number=Sing	7	nsubj	_	Tag=O
2	по	_	ADP	_	_	4	case	_	Tag=O
3	социальным	_	ADJ	_	Case=Dat|Degree=Pos|Number=Plur	4	amod	_	Tag=O
4	вопросам	_	NOUN	_	Animacy=Inan|Case=Dat|Gender=Masc|Number=Plur	1	nmod	_	Tag=O
5	Татьяна	_	PROPN	_	Animacy=Anim|Case=Nom|Gender=Fem|Number=Sing	1	appos	_	Tag=B-PER
6	Голикова	_	PROPN	_	Animacy=Anim|Case=Nom|Gender=Fem|Number=Sing	5	flat:name	_	Tag=I-PER

Ner tags are given in the form Tag=O ,Tag=B-PER, Tag=I-PER etc.

It would be really helpful if:

you provide some description as to what input file format is expected (perhaps, with examples)
what ner input format is expected.

if the first most probably is CoNLL-U, it would be helpful to explicitly state so, with a link to the format definition. The second needs an explicit specification.

Any help with this error is appreciated.

Which page or section is this issue related to?

https://spacy.io/api/cli#convert

Answered by adrianeboyd

Jul 13, 2020

I'm not sure what version of spacy you're using? You don't have to use it for training, but try the latest v2 (currently v2.3.1) for converting. I get the error:

File "/path/spacy/cli/converters/conllu2json.py", line 69, in read_conllx
    id_, word, lemma, pos, tag, morph, head, dep, _1, iob = parts
ValueError: not enough values to unpack (expected 10, got 9)

This line is malformed:

13	зафиксирована	_	VERB	_	Aspect=Perf|Gender=Fem|Number=Sing|Tense=Past|Variant=Short|VerbForm=Part|Voice=Pass	7ccomp	_	O

I suspect your best option is to use tools from universal dependencies to validate the CoNLL-U format.

View full answer

adrianeboyd · 2020-07-13T08:50:00Z

adrianeboyd
Jul 13, 2020

I think you're accidentally using the wrong converter here. -c ner doesn't support CoNLL-U format. The converter docs are here: https://spacy.io/api/cli#convert with links to the examples for the supported NER formats with -c ner here: https://github.com/explosion/spaCy/tree/27a1cd3c630055802513a20c1e75d0b37943cc39/examples/training/ner_example_data

For CoNLL-U you can leave off the -c option (it checks by file ending) or use -c conllu. The problem with CoNLL-U is that there is no standard format for NER annotation. The converter is intended to convert most of the standard CoNLL-U columns (currently in spacy v2 this is columns FORM/UPOS/XPOS/HEAD/DEP). For treebanks that we use in some of the provided models, it also supports a few formats for including NER in the MISC column, but none of these are an official supported part of CoNLL-U or our converter.

Let's see, I think the simplest option with spacy v2 is to remove the Tag= to have plain NER tags in the MISC column. If your data had SpaceAfter=No or other features in the MISC column the spacy v2 converter isn't going to work correctly, but if MISC is just plain NER tags, I think the converter will detect them correctly. I usually recommend the following command to group sentences into small paragraphs for better training (this is mainly so the parser learns to split sentences and you may want to experiment with it for your data):

spacy convert -n 10 file.conllu .

If you need support for SpaceAfter=No or other features in MISC, let me know and I can explain how to use the converter from the current alpha version of spacy v3. (I wouldn't recommend using the alpha v3 version for training yet, but the converter itself is stable and has a lot of improvements over v2. I used it internally for all the updated v2.3 models. It supports more NER formats in MISC, all of the CoNLL-U columns, and SpaceAfter=No in the MISC column to create raw texts.)

0 replies

sbushmanov · 2020-07-13T09:29:07Z

sbushmanov
Jul 13, 2020
Author

Thanks a lot for a quick response!

spacy convert file.conllu . works. spacy convert -n 3 file.conllu . works too. spacy convert -n 10 file.conllu . produces empty file.
In the json output I have unesaped unicode: "\u0422\u0430\u0442\u044c\u044f\u043d\u0430". Is there a way to have readable unicode text (like in json.dumps(...., ensure_ascii=False))?

Any chance to get help with a related issue while training?

ValueError: [E069] Invalid gold-standard parse tree. Found cycle between word IDs: {59, 51, 54}

(Strangely enough, I'm just training ner:

spacy train ru /home/sergey/Py_Spacy_RU/test_del /home/sergey/Py_Spacy_RU/data/nerus/nerus_json/try.json /home/sergey/Py_Spacy_RU/data/nerus/nerus_json/try.json --base-model /home/sergey/Py_Spacy_RU/ru2 --n-iter 20 --n-early-stopping 5 --pipeline 'ner'

)

Is there a way to have a list of misbehaving sentences (ids, or just starting text like in 2.3.1), so that sentences with cycles be removed at once? Or maybe you may advise a different way of treating cycles like ignoring them?

0 replies

adrianeboyd · 2020-07-13T10:18:55Z

adrianeboyd
Jul 13, 2020

Thanks a lot for a quick response!

1. `spacy convert file.conllu .` works. `spacy convert -n 3 file.conllu .` works too. `spacy convert -n 10 file.conllu .` produces empty file.

That is pretty unexpected behavior for the -n option. I'd check that your output contains the same number of sentences and the same texts as the input in case it's due to a problem with the input and something more serious is going wrong.

2. In the json output I have unesaped unicode: `"\u0422\u0430\u0442\u044c\u044f\u043d\u0430"`. Is there a way to have readable unicode text (like in `json.dumps(...., ensure_ascii=False)`)?

No, not directly with spacy convert, we just don't have an option for this built in. You'd have to read in the file with json and export it again with the ensure_ascii option to reformat it.

Any chance to get help with a related issue while training?

ValueError: [E069] Invalid gold-standard parse tree. Found cycle between word IDs: {59, 51, 54}

(Strangely enough, I'm just training ner:
spacy train ru /home/sergey/Py_Spacy_RU/test_del /home/sergey/Py_Spacy_RU/data/nerus/nerus_json/try.json /home/sergey/Py_Spacy_RU/data/nerus/nerus_json/try.json --base-model /home/sergey/Py_Spacy_RU/ru2 --n-iter 20 --n-early-stopping 5 --pipeline 'ner'
)

Is there a way to have a list of misbehaving sentences (ids, or just starting text like in 2.3.1), so that sentences with cycles be removed at once? Or maybe you may advise a different way of treating cycles like ignoring them?

Hmm, I think it tries to read in all the available annotation in case you might be training a parser later. I would just remove the "dep" and "head" values for all the tokens from the corpus in the JSON files to get around this if you're only training an NER model.

0 replies

sbushmanov · 2020-07-13T11:42:24Z

sbushmanov
Jul 13, 2020
Author

That is pretty unexpected behavior for the -n option. I'd check that your output contains the same number of sentences and the same texts as the input in case it's due to a problem with the input and something more serious is going wrong.

Source file:

(rasa) sergey@sergey-Bionic:~/Py_Spacy_RU/data/nerus$ cat try_.conllu
# newdoc id = 0
# sent_id = 0_0
# text = Вице-премьер по социальным вопросам Татьяна Голикова рассказала, в каких регионах России зафиксирована наиболее высокая смертность от рака, сообщает РИА Новости.
1	Вице-премьер	_	NOUN	_	Animacy=Anim|Case=Nom|Gender=Masc|Number=Sing	7	nsubj	_	O
2	по	_	ADP	_	_	4	case	_	O
3	социальным	_	ADJ	_	Case=Dat|Degree=Pos|Number=Plur	4	amod	_	O
4	вопросам	_	NOUN	_	Animacy=Inan|Case=Dat|Gender=Masc|Number=Plur	1	nmod	_	O
5	Татьяна	_	PROPN	_	Animacy=Anim|Case=Nom|Gender=Fem|Number=Sing	1	appos	_	B-PER
6	Голикова	_	PROPN	_	Animacy=Anim|Case=Nom|Gender=Fem|Number=Sing	5	flat:name	_	I-PER
7	рассказала	_	VERB	_	Aspect=Perf|Gender=Fem|Mood=Ind|Number=Sing|Tense=Past|VerbForm=Fin|Voice=Act	0	root	_	O
8	,	_	PUNCT	_	_	13	punct	_	O
9	в	_	ADP	_	_	11	case	_	O
10	каких	_	DET	_	Case=Loc|Number=Plur	11	det	_	O
11	регионах	_	NOUN	_	Animacy=Inan|Case=Loc|Gender=Masc|Number=Plur	13	obl	_	O
12	России	_	PROPN	_	Animacy=Inan|Case=Gen|Gender=Fem|Number=Sing	11	nmod	_	B-LOC
13	зафиксирована	_	VERB	_	Aspect=Perf|Gender=Fem|Number=Sing|Tense=Past|Variant=Short|VerbForm=Part|Voice=Pass	7ccomp	_	O
14	наиболее	_	ADV	_	Degree=Pos	15	advmod	_	O
15	высокая	_	ADJ	_	Case=Nom|Degree=Pos|Gender=Fem|Number=Sing	16	amod	_	O
16	смертность	_	NOUN	_	Animacy=Inan|Case=Nom|Gender=Fem|Number=Sing	13	nsubj:pass	_	O
17	от	_	ADP	_	_	18	case	_	O
18	рака	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Masc|Number=Sing	16	nmod	_	O
19	,	_	PUNCT	_	_	20	punct	_	O
20	сообщает	_	VERB	_	Aspect=Imp|Mood=Ind|Number=Sing|Person=3|Tense=Pres|VerbForm=Fin|Voice=Act	0	root	_	O
21	РИА	_	PROPN	_	Animacy=Inan|Case=Nom|Gender=Neut|Number=Sing	20	nsubj	_	B-ORG
22	Новости	_	PROPN	_	Animacy=Inan|Case=Nom|Gender=Fem|Number=Plur	21	appos	_	I-ORG
23	.	_	PUNCT	_	_	20	punct	_	O

# sent_id = 0_1
# text = По словам Голиковой, чаще всего онкологические заболевания становились причиной смерти в Псковской, Тверской, Тульской и Орловской областях, а также в Севастополе.
1	По	_	ADP	_	_	2	case	_	O
2	словам	_	NOUN	_	Animacy=Inan|Case=Dat|Gender=Neut|Number=Plur	9	parataxis	_	O
3	Голиковой	_	PROPN	_	Animacy=Anim|Case=Gen|Gender=Fem|Number=Sing	2	nmod	_	B-PER
4	,	_	PUNCT	_	_	2	punct	_	O
5	чаще	_	ADV	_	Degree=Cmp	9	advmod	_	O
6	всего	_	PRON	_	Animacy=Inan|Case=Gen|Gender=Neut|Number=Sing	5	obl	_	O
7	онкологические	_	ADJ	_	Case=Nom|Degree=Pos|Number=Plur	8	amod	_	O
8	заболевания	_	NOUN	_	Animacy=Inan|Case=Nom|Gender=Neut|Number=Plur	9	nsubj	_	O
9	становились	_	VERB	_	Aspect=Imp|Mood=Ind|Number=Plur|Tense=Past|VerbForm=Fin|Voice=Mid	0	root	_O
10	причиной	_	NOUN	_	Animacy=Inan|Case=Ins|Gender=Fem|Number=Sing	9	xcomp	_	O
11	смерти	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Fem|Number=Sing	10	nmod	_	O
12	в	_	ADP	_	_	20	case	_	O
13	Псковской	_	ADJ	_	Case=Loc|Degree=Pos|Gender=Fem|Number=Sing	20	amod	_	B-LOC
14	,	_	PUNCT	_	_	15	punct	_	O
15	Тверской	_	ADJ	_	Case=Loc|Degree=Pos|Gender=Fem|Number=Sing	13	conj	_	B-LOC
16	,	_	PUNCT	_	_	17	punct	_	O
17	Тульской	_	ADJ	_	Case=Loc|Degree=Pos|Gender=Fem|Number=Sing	13	conj	_	B-LOC
18	и	_	CCONJ	_	_	19	cc	_	O
19	Орловской	_	ADJ	_	Case=Loc|Degree=Pos|Gender=Fem|Number=Sing	13	conj	_	B-LOC
20	областях	_	NOUN	_	Animacy=Inan|Case=Loc|Gender=Fem|Number=Plur	9	obl	_	I-LOC
21	,	_	PUNCT	_	_	25	punct	_	O
22	а	_	CCONJ	_	_	25	cc	_	O
23	также	_	ADV	_	Degree=Pos	22	fixed	_	O
24	в	_	ADP	_	_	25	case	_	O
25	Севастополе	_	PROPN	_	Animacy=Inan|Case=Loc|Gender=Masc|Number=Sing	20	conj	_	B-LOC
26	.	_	PUNCT	_	_	9	punct	_	O

# sent_id = 0_2
# text = Вице-премьер напомнила, что главные факторы смертности в России — рак и болезни системы кровообращения.
1	Вице-премьер	_	NOUN	_	Animacy=Anim|Case=Nom|Gender=Masc|Number=Sing	2	nsubj	_	O
2	напомнила	_	VERB	_	Aspect=Perf|Gender=Fem|Mood=Ind|Number=Sing|Tense=Past|VerbForm=Fin|Voice=Act	0	root	_	O
3	,	_	PUNCT	_	_	11	punct	_	O
4	что	_	SCONJ	_	_	11	mark	_	O
5	главные	_	ADJ	_	Case=Nom|Degree=Pos|Number=Plur	6	amod	_	O
6	факторы	_	NOUN	_	Animacy=Inan|Case=Nom|Gender=Masc|Number=Plur	11	nsubj	_	O
7	смертности	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Fem|Number=Sing	6	nmod	_	O
8	в	_	ADP	_	_	9	case	_	O
9	России	_	PROPN	_	Animacy=Inan|Case=Loc|Gender=Fem|Number=Sing	7	nmod	_	B-LOC
10	—	_	PUNCT	_	_	11	punct	_	O
11	рак	_	NOUN	_	Animacy=Inan|Case=Nom|Gender=Masc|Number=Sing	6	nsubj	_	O
12	и	_	CCONJ	_	_	13	cc	_	O
13	болезни	_	NOUN	_	Animacy=Inan|Case=Nom|Gender=Fem|Number=Plur	11	conj	_	O
14	системы	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Fem|Number=Sing	13	nmod	_	O
15	кровообращения	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Neut|Number=Sing	14	nmod	_	O
16	.	_	PUNCT	_	_	2	punct	_	O

# sent_id = 0_3
# text = В начале года стало известно, что смертность от онкологических заболеваний среди россиян снизилась впервые за три года.
1	В	_	ADP	_	_	2	case	_	O
2	начале	_	NOUN	_	Animacy=Inan|Case=Loc|Gender=Neut|Number=Sing	4	obl	_	O
3	года	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Masc|Number=Sing	2	nmod	_	O
4	стало	_	VERB	_	Aspect=Perf|Gender=Neut|Mood=Ind|Number=Sing|Tense=Past|VerbForm=Fin|Voice=Act	0	root	_O
5	известно	_	ADJ	_	Degree=Pos|Gender=Neut|Number=Sing|Variant=Short	4	xcomp	_	O
6	,	_	PUNCT	_	_	14	punct	_	O
7	что	_	SCONJ	_	_	14	mark	_	O
8	смертность	_	NOUN	_	Animacy=Inan|Case=Nom|Gender=Fem|Number=Sing	14	nsubj	_	O
9	от	_	ADP	_	_	11	case	_	O
10	онкологических	_	ADJ	_	Case=Gen|Degree=Pos|Number=Plur	11	amod	_	O
11	заболеваний	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Neut|Number=Plur	8	nmod	_	O
12	среди	_	ADP	_	_	13	case	_	O
13	россиян	_	NOUN	_	Animacy=Anim|Case=Gen|Gender=Masc|Number=Plur	8	nmod	_	O
14	снизилась	_	VERB	_	Aspect=Perf|Gender=Fem|Mood=Ind|Number=Sing|Tense=Past|VerbForm=Fin|Voice=Mid	4	ccomp	_	O
15	впервые	_	ADV	_	Degree=Pos	14	advmod	_	O
16	за	_	ADP	_	_	18	case	_	O
17	три	_	NUM	_	Animacy=Inan|Case=Acc	18	nummod:gov	_	O
18	года	_	NOUN	_	Animacy=Inan|Case=Gen|Gender=Masc|Number=Sing	15	obl	_	O
19	.	_	PUNCT	_	_	4	punct	_	O

Convert command:

(rasa) sergey@sergey-Bionic:~/Py_Spacy_RU/data/nerus$ spacy convert -n 10 try_.conllu .
✔ Generated output file (0 documents)
try_.json

Output:

(rasa) sergey@sergey-Bionic:~/Py_Spacy_RU/data/nerus$ cat try_.json 
[

]

Thanks in advance!

0 replies

adrianeboyd · 2020-07-13T12:55:29Z

adrianeboyd
Jul 13, 2020

I'm not sure what version of spacy you're using? You don't have to use it for training, but try the latest v2 (currently v2.3.1) for converting. I get the error:

File "/path/spacy/cli/converters/conllu2json.py", line 69, in read_conllx
    id_, word, lemma, pos, tag, morph, head, dep, _1, iob = parts
ValueError: not enough values to unpack (expected 10, got 9)

This line is malformed:

13	зафиксирована	_	VERB	_	Aspect=Perf|Gender=Fem|Number=Sing|Tense=Past|Variant=Short|VerbForm=Part|Voice=Pass	7ccomp	_	O

I suspect your best option is to use tools from universal dependencies to validate the CoNLL-U format.

0 replies

sbushmanov · 2020-07-14T07:59:33Z

sbushmanov
Jul 14, 2020
Author

spacy convert -n 10 file.conllu . works indeed in the latest spacy (2.3.1).

The empty file I mentioned was due to spacy 2.1.8.

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ValueError: [E177] Ill-formed IOB input detected #5753

{{title}}

{{editor}}'s edit

{{editor}}'s edit

Replies: 6 comments

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

Select a reply

ValueError: [E177] Ill-formed IOB input detected #5753

sbushmanov Jul 12, 2020

Which page or section is this issue related to?

Replies: 6 comments

adrianeboyd Jul 13, 2020

sbushmanov Jul 13, 2020 Author

adrianeboyd Jul 13, 2020

sbushmanov Jul 13, 2020 Author

adrianeboyd Jul 13, 2020

sbushmanov Jul 14, 2020 Author

sbushmanov
Jul 12, 2020

adrianeboyd
Jul 13, 2020

sbushmanov
Jul 13, 2020
Author

adrianeboyd
Jul 13, 2020

sbushmanov
Jul 13, 2020
Author

adrianeboyd
Jul 13, 2020

sbushmanov
Jul 14, 2020
Author