overcoming ✘ invalid whitespace entity spans

[DSLituiev]

What is the recommended way to solve the above issue?

I tried this, but it doesn't seem to do the job:

def trim_whitespaces(spans):
    doc = spans[0].doc
    for sp in spans:
        if (doc[sp.end].is_space) or doc[sp.end].text in {",",":", ".", ";", "\n"}:
            for ii in range(sp.end, sp.start-1, -1):
                if not (doc[ii].is_space):
                    sp.end = ii + 1
                    break
        if (doc[sp.start].is_space):
            for ii in range( sp.start, sp.end):
                if not (doc[ii].is_space):
                    sp.start = ii
                    break
    return spans

doc.set_ents(doc.spans)

[polm]

I would look at what kind of entities you have that are including spaces and why it's happening. Did your annotators actually put labels on spaces?

Usually the issue is going to be that your annotations are off by some offset due to preprocessing or a miscalculation somewhere.

Also there's nothing obviously wrong with your code, but you might want to use the alignment_mode="contract" argument to char_spans, which should do what you're doing manually.

[DSLituiev]

Hi @polm thank you for giving it a try. Here is a ZIP file with a minimal working example. Sorry it takes several files to make it all work.

minimal-example.zip

Warning I am getting with python -m spacy debug data config.cfg --code ./functions.py is:

============================== Vocab & Vectors ==============================
ℹ 40 total word(s) in the data (28 unique)
⚠ 4 misaligned tokens in the training data
ℹ No word vectors present in the package

[DSLituiev]

the training and validation data are the same in previous example:

[
        {"text": "A. First specimen (1/8) lymph nodes",
         "ents":[{"start": 19, "end":20, "label":"n_ln_involved"},
                    {"start": 21, "end": 22,  "label":"n_ln_examined"}]},
          {'text': 'Five  local lymph nodes without metastatic carcinoma (0/5).',
           'ents': [{'start': 0, 'end': 4, 'label': 'n_ln_examined'},
                      {'start': 24, 'end': 31, 'label': 'n_ln_involved'},
                      {'start': 54, 'end': 55, 'label': 'n_ln_involved'},
                      {'start': 56, 'end': 57, 'label': 'n_ln_examined'}],},
       {'text': 'adenocarcinoma involving two of ten regional\n lymph nodes (2/10).',
        'ents': [{'start': 25, 'end': 28, 'label': 'n_ln_involved'},
                    {'start': 32, 'end': 35, 'label': 'n_ln_examined'},
                    {'start': 59, 'end': 60, 'label': 'n_ln_involved'},
                    {'start': 61, 'end': 63, 'label': 'n_ln_examined'}],}
        ]

[DSLituiev]

This is the script used to generate the minimal example:

#!/usr/bin/env python
# coding: utf-8
import sys
import os
import json
from collections import defaultdict, Counter

import spacy
from spacy.tokens import Doc, DocBin, Span
from spacy.matcher import Matcher
from spacy import displacy

from functions import make_customize_tokenizer_pathology

fn_model = "en_core_web_sm"
nlp = spacy.load(fn_model)
make_customize_tokenizer_pathology()(nlp)


data = [{"ents": [{"start": 19, "end":20, "label":"n_ln_involved"},
    {"start": 21, "end": 22,  "label":"n_ln_examined"}],
    "text": "A. First specimen (1/8) lymph nodes"},
         {'text': 'Five  local lymph nodes without metastatic carcinoma (0/5).',
 'ents': [{'start': 0, 'end': 4, 'label': 'n_ln_examined'},
  {'start': 24, 'end': 31, 'label': 'n_ln_involved'},
  {'start': 54, 'end': 55, 'label': 'n_ln_involved'},
  {'start': 56, 'end': 57, 'label': 'n_ln_examined'}],},
         {'text': 'adenocarcinoma involving two of ten regional\n lymph nodes (2/10).',
 'ents': [{'start': 25, 'end': 28, 'label': 'n_ln_involved'},
  {'start': 32, 'end': 35, 'label': 'n_ln_examined'},
  {'start': 59, 'end': 60, 'label': 'n_ln_involved'},
  {'start': 61, 'end': 63, 'label': 'n_ln_examined'}],}
        ]

docs = []
for ii, entry in enumerate((data)):
    text = entry["text"]
    doc = nlp(text)
    

    spans_list = (entry["ents"])
    spans = [doc.char_span(ent["start"], ent["end"], label=ent["label"], vector=None) for ent in spans_list]
    if any((len(sp)==0 for sp in spans)):
        raise ValueError("zero-lenght span")
    doc.set_ents(spans)
    docs.append(doc)

docbins = defaultdict(lambda: DocBin())

for ii, doc in (enumerate(docs)):
    docbins["train"].add(doc)
    docbins["val"].add(doc)
    
for set_, docbin in docbins.items():
    docbin.to_disk(f"./general-onco/minimal-example/pathology-example-{set_}.spacy")

[polm]

Thanks for the examples! Note that #8774 looks like it's probably related.

[polm]

So I used your minimal example and:

1. I was able to verify that with the latest spaCy release debug data reports the misalignments but gives no error, which is wrong.
2. With master (which includes Custom tokenizer not initialized correctly on debug data #8774) everything works as it should.

So this is the same issue as that addressed in #8774, and should be fixed in the next 3.1 release.

Thanks for your patience, and sorry it took so long to get to it.

[BramVanroy]

AFAIK the flag is --code, not --code-path.

[DSLituiev]

@BramVanroy

python -m spacy debug data --help
Usage: python -m spacy debug data [OPTIONS] CONFIG_PATH

  Analyze, debug and validate your training and development data. Outputs
  useful stats, and can help you find problems like invalid entity
  annotations, cyclic dependencies, low data labels and more.

  DOCS: https://spacy.io/api/cli#debug-data

Arguments:
  CONFIG_PATH  Path to config file  [required]

Options:
  -c, --code-path, --code PATH  Path to Python file with additional code
                                (registered functions) to be imported

[BramVanroy]

My bad, that's missing in the docs then.