pythonnlpspacyconll

Spacy identifying blank spaces as entities


I am just starting to work with Spacy and have put a text through to test how it is working on a pdf I OCR'd with AntFileConverter.

The txt file (sample below - would attach but unsure how) seems fine, is in UTF-8. However when I output the file in CONLL format, for some reason there are various apparent gaps, which have no original word, but seem to have been identified. This happens both at the end and within sentences.

"species in many waters in the northern hemisphere. In most countries in the region pike has both commercial and recreational value (Crossman & Casselman 1987; Raat 1988). Pike is a typical sit-and-wait predator which usually hunts prey by ambushing (Webb & Skadsen 1980)."

The output us as so:

        GPE 24  
26  species specie  NNS     20  attr
27  in  in  IN      26  prep
28  many    many    JJ      29  amod
29  waters  water   NNS     27  pobj
30  in  in  IN      29  prep
31  the the DT      33  det
32  northern    northern    JJ      33  amod
33  hemisphere  hemisphere  NN      30  pobj
34  .   .   .       20  punct
1   In  in  IN      9   prep
2   

        GPE 1   
3   most    most    JJS     4   amod
4   countries   country NNS     9   nsubj
5   in  in  IN      4   prep
6   the the DT      8   det
7   region  region  NN      8   compound
8   pike    pike    NN      5   pobj
9   has have    VBZ     0   ROOT
10  both    both    DT      11  preconj
11  commercial  commercial  JJ      9   dobj
12  

        GPE 11  
13  and and CC      11  cc
14  recreational    recreational    JJ      15  amod
15  value   value   NN      11  conj
16  (   (   -LRB-       15  punct
17  Crossman    crossman    NNP ORG 15  appos
18  &   &   CC  ORG 17  cc
19  Casselman   casselman   NNP ORG 17  conj
20  1987    1987    CD  DATE    17  nummod
21  ;   ;   :       15  punct
22  

        GPE 21  
23  Raat    raat    NNP     15  appos
24  1988    1988    CD  DATE    23  nummod
25  )   )   -RRB-       15  punct
26  .   .   .       9   punct
1   Pike    pike    NNP     2   nsubj
2   is  be  VBZ     0   ROOT
3   a   a   DT      10  det
4   typical typical JJ      10  amod
5   sit sit NN      10  nmod
6   -   -   HYPH        5   punct
7   and and CC      5   cc
8   -   -   HYPH        9   punct
9   wait    wait    VB      5   conj
10  predator    predator    NN      2   attr
11  

        GPE 10  
12  which   which   WDT     14  nsubj
13  usually usually RB      14  advmod
14  hunts   hunt    VBZ     10  relcl
15  prey    prey    NN      14  dobj
16  by  by  IN      14  prep
17  ambushing   ambush  VBG     16  pcomp
18  (   (   -LRB-       17  punct
19  Webb    webb    NNP     17  conj
20  &   &   CC      19  cc
21  

I also tried without the NER print out but these gaps continue to be marked. I thought it might be related to the line breaks, so I also tried with a Linux-style EOL but that didn't make any difference.

This is the code I am using:

import spacy
import en_core_web_sm
nlp_en = en_core_web_sm.load()
input = open('./input/55_linux.txt', 'r').read()
doc = nlp_en(input)
for sent in doc.sents:
        for i, word in enumerate(sent):
              if word.head == word:
                 head_idx = 0
              else:
                 head_idx = word.head.i - sent[0].i + 1
              output = open('CONLL_output.txt', 'a')
              output.write("%d\t%s\t%s\t%s\t%s\t%s\t%s\n"%(
                 i+1, # There's a word.i attr that's position in *doc*
                  word,
                  word.lemma_,
                  word.tag_, # Fine-grained tag
                  word.ent_type_,
                  str(head_idx),
                  word.dep_ # Relation
                 ))

Has anyone else had this problem? If so, do you know how I can solve it?


Solution

  • This is a known bug in spaCy.

    Until it is fixed, you will have to do some post-processing to get rid of those "blank" entities. Fortunately, this is easy enough, this snippet posted by the author of the library shows how:

    def remove_whitespace_entities(doc):
        doc.ents = [e for e in doc.ents if not e.text.isspace()]
        return doc
    
    nlp_en.add_pipe(remove_whitespace_entities, after='ner')
    

    So, you first define a post-processing pipe that filters all entities with a text solely consisting of whitespace characters (using isspace()).

    Then you add this pipe to the NLP pipeline, set to run after NER. Then any time you use nlp_en after that, it will not return those entities.