How to let Knowledge base works ignore upper/lower case

[aoleiReiz]

currently I have kb which all upper cased, it Entity linker failed to get kb_id when do prediction on lowered text. could you please help how to solve it

[svlandeg]

Hi!

From spaCy v3.0 onwards, it should (hopefully) be relatively straightforward to tune the candidate generation to your needs. The candidate generation step selects all relevant IDs from the Knowledge base, given a certain textual mention. More specifically, it needs to be a function that takes a KnowledgeBase and a Span, and outputs a list of Candidate instances:

The default candidate generation method is spacy.CandidateGenerator.v1 and you are correct that it works case-sensitive:

@registry.misc("spacy.CandidateGenerator.v1")
def create_candidates():
    return get_candidates

def get_candidates(kb, span):
    return kb.get_alias_candidates(span.text)

You should be able to overwrite this behaviour by defining your own candidate generator, e.g. something like this:

@registry.misc("custom_upper_candidates")
def create_candidates():
    return get_custom_candidates

def get_custom_candidates(kb, span):
    cased_results = kb.get_alias_candidates(span.text)
    if len(cased_results) > 0:
        return cased_results
    return kb.get_alias_candidates(span.text.upper())

And then you'd define this candidate generator in your config file:

[components.entity_linker]
factory = "entity_linker"
entity_vector_length = 64
get_candidates = {"@misc":"custom_upper_candidates"}
incl_context = true
incl_prior = true
labels_discard = []

Let us know if you run into issues with that!

[amirj]

Thanks @svlandeg

Assume that we're creating a more complex candidate generator function; we need to return a list of candidates as follow:

def get_custom_candidates(kb, span):
    candidate = Candidate(kb, entity_hash, entity_freq, entity_vector, alias_hash, prior_prob)
    return [candidate]

In order to create a Candidate object we need entity_hash, entity_freq, entity_vector, alias_hash, prior_prob. To the best of my knowledge, some of these items (entity_vector , prior_prob) can be retrieved from kb, but it's not clear to me how to assign entity_hash, entity_freq, alias_hash?

[svlandeg]

These are closely tied to the other methods in the KnowledgeBase as implemented internally in spaCy, so you can find the definitions in the documentation:

entity_freq and entity_vector are the two arguments you'd pass when calling add_entity. entity_freq is currently not used by the implementation of the entity_linker.

prior_prob is the probability you'd pass on for a given entity-alias combination when calling add_alias.

entity_hash and alias_hash are the hashes of the entity ID and the mention string in the vocab respectively, defined by:

# when adding the strings for the first time
entity_hash = self.vocab.strings.add(entity)
alias_hash = self.vocab.strings.add(alias)

or

# assuming the strings are already in the vocab
entity_hash = self.vocab.strings[entity]
alias_hash = self.vocab.strings[alias]

[amirj]

Thank @svlandeg

I'm using the following functions:

@registry.misc("custom_candidate_generator")
def create_candidates():
    return get_custom_candidates


def get_custom_candidates(kb, span):
    candidates = []
    for ent_id, score in candidate_generator_func():
        candidate = Candidate(
            kb=kb,
            entity_hash=kb.vocab.strings[ent_id],
            entity_freq=111,
            entity_vector=kb.get_vector(ent_id),
            alias_hash=kb.vocab.strings[span.text],
            prior_prob=score,
        )
        candidates.append(candidate)
    return candidates

Add entity linker pipe and initialize for training:

config = {
    "labels_discard": [],
    "n_sents": 0,
    "incl_prior": True,
    "incl_context": True,
    "model": DEFAULT_NEL_MODEL,
    "entity_vector_length": 200,
    "get_candidates": {"@misc": "custom_candidate_generator"},
}
el = self.nlp.add_pipe("entity_linker", config=config, last=True)
el.initialize(get_examples=self.get_examples, kb_loader=load_kb(loc_kb)) #loc_kb is the Path of KB on disk

EntityLinker is trained for one epoch, after that when I'm going to evaluate the model on a dev set, I'm getting the following error:

***/python3.8/site-packages/spacy/pipeline/entity_linker.py:362: RuntimeWarning: invalid value encountered in true_divide
  sims = xp.dot(
Traceback (most recent call last):
  File "sample_entity_linker/models/spacy_entity_linker.py", line 292, in <module>
    spacy_entity_linker.train(epoches=10)
  File "sample_entity_linker/models/spacy_entity_linker.py", line 205, in train
    score = self.evaluate(dataset=self.dev_set)
  File "sample_entity_linker/models/spacy_entity_linker.py", line 220, in evaluate
    score = self.nlp.evaluate(dataset)
  File "/home/aj896572/.conda/envs/sample-entity-linker/lib/python3.8/site-packages/spacy/language.py", line 1355, in evaluate
    for doc, eg in zip(
  File "/home/aj896572/.conda/envs/sample-entity-linker/lib/python3.8/site-packages/spacy/util.py", line 1484, in _pipe
    yield from proc.pipe(docs, **kwargs)
  File "spacy/pipeline/trainable_pipe.pyx", line 79, in pipe
  File "/home/aj896572/.conda/envs/sample-entity-linker/lib/python3.8/site-packages/spacy/util.py", line 1503, in raise_error
    raise e
  File "spacy/pipeline/trainable_pipe.pyx", line 75, in spacy.pipeline.trainable_pipe.TrainablePipe.pipe
  File "/home/aj896572/.conda/envs/sample-entity-linker/lib/python3.8/site-packages/spacy/pipeline/entity_linker.py", line 373, in predict
    final_kb_ids.append(best_candidate.entity_)
  File "spacy/kb.pyx", line 45, in spacy.kb.Candidate.entity_.__get__
  File "spacy/strings.pyx", line 132, in spacy.strings.StringStore.__getitem__
KeyError: "[E018] Can't retrieve string for hash '17224167776089022530'. This usually refers to an issue with the `Vocab` or `StringStore`."

I'm not sure to completely understand the relationship between EntityLinker and KnowledgeBase, i.e. when adding a new entity_linker pipe, it's pairing to which kb?! given the fact that we call "initialize" function only for training purpose --I see a set_kb function in the documentation but I'm not using it in my code!

[svlandeg]

This might be partially addressed by #8080, but I wanted to come back to this question:

when adding a new entity_linker pipe, it's pairing to which kb?!

Upon construction, the entity_linker gets an empty KnowledgeBase. You need to override this with either set_kb or initialize.