def convert_single_file(input_paragraph_list: List[str],
output_xmi_file: str) -> None:
document_text = '\n'.join(input_paragraph_list)
cas = Cas(typesystem=cassis.load_dkpro_core_typesystem())
cas.sofa_string = document_text
print("----")
print(document_text)
print("----")
token_type: Type = cas.typesystem.get_type(
'de.tudarmstadt.ukp.dkpro.core.api.segmentation.type.Token')
paragraph_type: Type = cas.typesystem.get_type(
'de.tudarmstadt.ukp.dkpro.core.api.segmentation.type.Paragraph')
sentence_type: Type = cas.typesystem.get_type(
'de.tudarmstadt.ukp.dkpro.core.api.segmentation.type.Sentence')
total_doc_offset: int = 0
for paragraph_str in input_paragraph_list:
this_paragraph_total_offset = total_doc_offset
doc: Doc = nlp(paragraph_str)
for token in doc:
assert isinstance(token, Token)
# print(token.text, token.idx, len(token), token.idx + len(token), token.is_space)
begin: int = total_doc_offset + token.idx
end: int = total_doc_offset + token.idx + len(token)
# annotate token -- only if it is not a space!
if not token.is_space:
cas.add_annotation(token_type.__call__(begin=begin, end=end))
total_doc_offset += len(paragraph_str)
# annotate paragraph
this_paragraph_annotation = paragraph_type.__call__(
begin=this_paragraph_total_offset, end=total_doc_offset)
cas.add_annotation(this_paragraph_annotation)
# and for paragraph too; but how about the '\n' char? maybe +1?
total_doc_offset += 1
# add sentences aligned exactly to paragraphs
cas.add_annotation(
sentence_type.__call__(begin=this_paragraph_annotation.begin,
end=this_paragraph_annotation.end))
print([x.get_covered_text() for x in cas.select(paragraph_type.name)])
print([x.get_covered_text() for x in cas.select(sentence_type.name)])
print([x.get_covered_text() for x in cas.select(token_type.name)])
# create parent folder if not exists
Path(output_xmi_file).parent.mkdir(parents=True, exist_ok=True)
cas.to_xmi(output_xmi_file)
def write_sentence_documents(sentences: List[str],
labels: List[str],
path: Path,
labeled=True):
typesystem = TypeSystem()
cas = Cas(typesystem=typesystem)
SentenceType = typesystem.create_type(
"de.tudarmstadt.ukp.dkpro.core.api.segmentation.type.Sentence")
SentimentType = typesystem.create_type("webanno.custom.Sentiment")
typesystem.add_feature(type_=SentimentType,
name="value",
rangeTypeName="uima.cas.String")
cas.sofa_string = " ".join(sentences)
begin = 0
for sentence, label in zip(sentences, labels):
end = begin + len(sentence)
cas_sentence = SentenceType(begin=begin, end=end)
sentiment_annotation = SentimentType(begin=begin, end=end, value=label)
begin = end + 1
cas.add_annotation(cas_sentence)
if labeled:
cas.add_annotation(sentiment_annotation)
cas.to_xmi(path, pretty_print=True)
for sentence in cas.select(SENTENCE_TYPE):
print(cas.get_covered_text(sentence))
def _generate_candidates(self, cas: Cas, n: int):
# We generate token n-grams
for tokens in mit.windowed(cas.select(TOKEN_TYPE), n):
begin = tokens[0].begin
end = tokens[-1].end
text = cas.sofa_string[begin:end]
yield (begin, end, text)
def iter_tokens(self, cas: Cas) -> Iterator[FeatureStructure]:
""" Returns an iterator over all tokens in the given document.
Args:
cas:
Returns:
"""
return cas.select(TOKEN_TYPE)
def iter_sentences(self, cas: Cas) -> Iterator[FeatureStructure]:
""" Returns an iterator over all sentences in the given document.
Args:
cas:
Returns:
"""
return cas.select(SENTENCE_TYPE)
def predict(self, cas: Cas, layer: str, feature: str, project_id: str,
document_id: str, user_id: str):
# Extract the tokens from the CAS and create a spacy doc from it
words = [
cas.get_covered_text(cas_token)
for cas_token in cas.select(TOKEN_TYPE)
]
doc = Doc(self._model.vocab, words=words)
# Get the pos tags
self._model.get_pipe("tok2vec")(doc)
self._model.get_pipe("tagger")(doc)
# For every token, extract the POS tag and create an annotation in the CAS
for cas_token, spacy_token in zip(cas.select(TOKEN_TYPE), doc):
prediction = create_prediction(cas, layer, feature,
cas_token.begin, cas_token.end,
spacy_token.tag_)
cas.add_annotation(prediction)
def predict(self, cas: Cas, layer: str, feature: str, project_id: str,
document_id: str, user_id: str):
stemmer = nltk.PorterStemmer()
# For every token, steam it and create an annotation in the CAS
for cas_token in cas.select(TOKEN_TYPE):
stem = stemmer.stem(cas_token.get_covered_text())
begin = cas_token.begin
end = begin + len(stem)
prediction = create_prediction(cas, layer, feature, begin, end,
stem)
cas.add_annotation(prediction)
def predict(self, cas: Cas, layer: str, feature: str, project_id: str, document_id: str, user_id: str):
for i, sentence in enumerate(cas.select(SENTENCE_TYPE)):
token_ids = self._tokenizer.convert_tokens_to_ids(self._tokenizer.tokenize(sentence.get_covered_text()))
input_tensor = torch.tensor([token_ids])
# predict output tensor
outputs = self._model(input_tensor, adapter_names=[self._adapter_internal_name])
# retrieve the predicted class label
label_id = torch.argmax(outputs[0]).item()
label = self._label_map[label_id]
prediction = self.create_prediction(cas, layer, feature, sentence.begin, sentence.end, label)
cas.add_annotation(prediction)
def predict(self, cas: Cas, layer: str, feature: str, project_id: str,
document_id: str, user_id: str):
model: Optional[Pipeline] = self._load_model(user_id)
if model is None:
logger.debug("No trained model ready yet!")
return
for sentence in cas.select(SENTENCE_TYPE):
predicted = model.predict([sentence.get_covered_text()])[0]
prediction = create_prediction(cas, layer, feature, sentence.begin,
sentence.end, predicted)
cas.add_annotation(prediction)
def predict(self, cas: Cas, layer: str, feature: str, project_id: str, document_id: str, user_id: str):
model = self._load_model(user_id)
if model is None:
logger.debug("No trained model ready yet!")
return
featurizer = self._get_featurizer()
sentences = cas.select(SENTENCE_TYPE)
featurized_sentences = featurizer.featurize([s.get_covered_text() for s in sentences])
predictions = model.predict(featurized_sentences)
for sentence, featurized_sentence, label in zip(sentences, featurized_sentences, predictions):
prediction = create_prediction(cas, layer, feature, sentence.begin, sentence.end, label)
cas.add_annotation(prediction)
def predict(self, cas: Cas, layer: str, feature: str, project_id: str, document_id: str, user_id: str):
for sentence in cas.select(SENTENCE_TYPE):
cas_tokens = list(cas.select_covered(TOKEN_TYPE, sentence))
tokens = [t.get_covered_text() for t in cas_tokens]
grouped_bert_tokens = self._tokenize_bert(tokens)
predictions = self._predict(grouped_bert_tokens)
grouped_predictions = self._align_tokens(tokens, grouped_bert_tokens, predictions)
for token, grouped_prediction in zip(cas_tokens, grouped_predictions):
begin = token.begin
end = token.end
label = Counter([self._label_map[pred] for pred in grouped_prediction]).most_common(1)[0][0]
prediction = self.create_prediction(cas, layer, feature, begin, end, label)
cas.add_annotation(prediction)
def featurize_cas(fg: FeatureGenerator, cas: Cas) -> List:
features = get_features()
results = []
for qid, entity in enumerate(cas.select("webanno.custom.EntityLinking")):
candidates = list(
cas.select_covered("inception.internal.KbHandle", entity))
if len(candidates) == 0:
continue
for i, candidate in enumerate(candidates):
if entity.iri == candidate.iri:
gold_idx = i
break
else:
continue
sentences = list(
cas.select_covering(
"de.tudarmstadt.ukp.dkpro.core.api.segmentation.type.Sentence",
entity))
assert len(sentences) == 1
sentence = sentences[0]
mention = entity.get_covered_text().lower()
context = sentence.get_covered_text().lower()
l = len(context)
# context = context[int(l * 0.25):int(l * 0.75)]
for cid, candidate in enumerate(candidates):
score = float(entity.iri == candidate.iri)
query = candidate.query
label = candidate.label.lower()
result = fg.featurize_candidate(qid, cid, "inception_rank", score,
mention, context, label or "",
candidate.description or "",
entity.iri, gold_idx,
candidate.iri, features)
result.update(fg.featurize_query(mention, query, label))
results.append(result)
return results
def predict(self, cas: Cas, layer: str, feature: str, project_id: str,
document_id: str, user_id: str):
model: Optional[sklearn_crfsuite.CRF] = self._load_model(user_id)
if model is None:
logger.debug("No trained model ready yet!")
return
all_tokens = []
featurized_sentences = []
for sentence in cas.select(SENTENCE_TYPE):
tokens = list(cas.select_covered(TOKEN_TYPE, sentence))
words = [token.get_covered_text() for token in tokens]
all_tokens.append(tokens)
featurized_sentences.append(self._sent2features(words))
all_predictions = model.predict(featurized_sentences)
assert len(all_predictions) == len(all_tokens)
for predictions, tokens in zip(all_predictions, all_tokens):
assert len(predictions) == len(tokens)
begin = None
end = None
prev_tag = "O"
for tag, token in zip(predictions, tokens):
if begin is not None and end is not None:
if tag == "O" or (tag.startswith("B")
and prev_tag.startswith("I")):
prediction = create_prediction(cas, layer, feature,
begin, end, "X")
cas.add_annotation(prediction)
if tag.startswith("B"):
begin = token.begin
end = token.end
elif tag.startswith("I"):
end = token.end
else:
begin = None
end = None
prev_tag = tag
def predict(self, cas: Cas, layer: str, feature: str, project_id: str,
document_id: str, user_id: str):
# Extract the tokens from the CAS and create a spacy doc from it
cas_tokens = cas.select(TOKEN_TYPE)
words = [cas.get_covered_text(cas_token) for cas_token in cas_tokens]
doc = Doc(self._model.vocab, words=words)
# Find the named entities
self._model.get_pipe("ner")(doc)
# For every entity returned by spacy, create an annotation in the CAS
for named_entity in doc.ents:
begin = cas_tokens[named_entity.start].begin
end = cas_tokens[named_entity.end - 1].end
label = named_entity.label_
prediction = create_prediction(cas, layer, feature, begin, end,
label)
cas.add_annotation(prediction)
def get_perc_of_mapping_type(self, cas: Cas,
alignment: AlignmentLabel) -> float:
overallMatchesCount = 0
itemsOfGivenTypeCount = 0
for t in cas.select(FeatureExtractor.TOKEN_TYPE):
item = self.get_mappable_ann(cas, t)
# check for matches/alignment
if item.match is not None and item.match.target is not None:
overallMatchesCount += 1
# check for types
if item.match.label == alignment.name:
itemsOfGivenTypeCount += 1
# if nothing has been matched at all, the result is 0
if overallMatchesCount == 0:
return 0.0
else:
return itemsOfGivenTypeCount / overallMatchesCount
def predict(self, cas: Cas, layer: str, feature: str, project_id: str,
document_id: str, user_id: str):
sentences = cas.select(SENTENCE_TYPE)
src_tokens = cas.select_covered("webanno.custom.Base", sentences[0])
trg_tokens = cas.select_covered("webanno.custom.Base", sentences[1])
src_sentence = [e.get_covered_text() for e in src_tokens]
trg_sentence = [e.get_covered_text() for e in trg_tokens]
print(src_sentence)
print(trg_sentence)
alignments = self._aligner.get_word_aligns(src_sentence, trg_sentence)
Relation = cas.typesystem.get_type(layer)
print(list(Relation.all_features))
for matching_method in alignments:
for source_idx, target_idx in alignments[matching_method]:
src = src_tokens[source_idx]
target = trg_tokens[target_idx]
prediction = Relation(
Governor=src,
Dependent=target,
begin=target.begin,
end=target.end,
inception_internal_predicted=True,
)
# setattr(prediction, feature, f"{src.get_covered_text()} -> {target.get_covered_text()}")
setattr(prediction, feature, "")
print(source_idx, target_idx, prediction)
cas.add_annotation(prediction)
break
