Upload transformers_recognizer.py
Browse files- transformers_recognizer.py +245 -0
transformers_recognizer.py
ADDED
@@ -0,0 +1,245 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
from typing import Optional, List, Tuple, Set
|
3 |
+
|
4 |
+
from presidio_analyzer import (
|
5 |
+
RecognizerResult,
|
6 |
+
EntityRecognizer,
|
7 |
+
AnalysisExplanation,
|
8 |
+
)
|
9 |
+
from presidio_analyzer.nlp_engine import NlpArtifacts
|
10 |
+
|
11 |
+
logger = logging.getLogger("presidio-analyzer")
|
12 |
+
|
13 |
+
try:
|
14 |
+
from transformers import (
|
15 |
+
AutoTokenizer,
|
16 |
+
AutoModelForTokenClassification,
|
17 |
+
pipeline,
|
18 |
+
models,
|
19 |
+
)
|
20 |
+
from transformers.models.bert.modeling_bert import BertForTokenClassification
|
21 |
+
except ImportError:
|
22 |
+
logger.error("transformers is not installed")
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
class TransformersRecognizer(EntityRecognizer):
|
27 |
+
"""
|
28 |
+
Wrapper for a transformers model, if needed to be used within Presidio Analyzer.
|
29 |
+
|
30 |
+
:example:
|
31 |
+
>from presidio_analyzer import AnalyzerEngine, RecognizerRegistry
|
32 |
+
|
33 |
+
>transformers_recognizer = TransformersRecognizer()
|
34 |
+
|
35 |
+
>registry = RecognizerRegistry()
|
36 |
+
>registry.add_recognizer(transformers_recognizer)
|
37 |
+
|
38 |
+
>analyzer = AnalyzerEngine(registry=registry)
|
39 |
+
|
40 |
+
>results = analyzer.analyze(
|
41 |
+
> "My name is Christopher and I live in Irbid.",
|
42 |
+
> language="en",
|
43 |
+
> return_decision_process=True,
|
44 |
+
>)
|
45 |
+
>for result in results:
|
46 |
+
> print(result)
|
47 |
+
> print(result.analysis_explanation)
|
48 |
+
|
49 |
+
|
50 |
+
"""
|
51 |
+
|
52 |
+
ENTITIES = [
|
53 |
+
"LOCATION",
|
54 |
+
"PERSON",
|
55 |
+
"ORGANIZATION",
|
56 |
+
"AGE",
|
57 |
+
"ID",
|
58 |
+
"PHONE",
|
59 |
+
"EMAIL",
|
60 |
+
"DATE",
|
61 |
+
|
62 |
+
]
|
63 |
+
|
64 |
+
DEFAULT_EXPLANATION = "Identified as {} by transformers's Named Entity Recognition"
|
65 |
+
|
66 |
+
CHECK_LABEL_GROUPS = [
|
67 |
+
({"LOCATION"}, {"LOC", "HOSP"}),
|
68 |
+
({"PERSON"}, {"PER", "PERSON", "STAFF","PATIENT"}),
|
69 |
+
({"ORGANIZATION"}, {"ORGANIZATION", "ORG", "PATORG"}),
|
70 |
+
({"AGE"}, {"AGE"}),
|
71 |
+
({"ID"}, {"ID"}),
|
72 |
+
({"EMAIL"}, {"EMAIL"}),
|
73 |
+
({"DATE"}, {"DATE"}),
|
74 |
+
|
75 |
+
]
|
76 |
+
|
77 |
+
PRESIDIO_EQUIVALENCES = {
|
78 |
+
"PER": "PERSON",
|
79 |
+
"LOC": "LOCATION",
|
80 |
+
"ORG": "ORGANIZATION",
|
81 |
+
"AGE": "AGE",
|
82 |
+
"ID": "ID",
|
83 |
+
"EMAIL": "EMAIL"
|
84 |
+
}
|
85 |
+
|
86 |
+
DEFAULT_MODEL_PATH = "obi/deid_roberta_i2b2"
|
87 |
+
|
88 |
+
def __init__(
|
89 |
+
self,
|
90 |
+
supported_entities: Optional[List[str]] = None,
|
91 |
+
check_label_groups: Optional[Tuple[Set, Set]] = None,
|
92 |
+
model: Optional[BertForTokenClassification] = None,
|
93 |
+
model_path: Optional[str] = None,
|
94 |
+
):
|
95 |
+
if not model and not model_path:
|
96 |
+
model_path = self.DEFAULT_MODEL_PATH
|
97 |
+
logger.warning(
|
98 |
+
f"Both 'model' and 'model_path' arguments are None. Using default model_path={model_path}"
|
99 |
+
)
|
100 |
+
|
101 |
+
if model and model_path:
|
102 |
+
logger.warning(
|
103 |
+
f"Both 'model' and 'model_path' arguments were provided. Ignoring the model_path"
|
104 |
+
)
|
105 |
+
|
106 |
+
self.check_label_groups = (
|
107 |
+
check_label_groups if check_label_groups else self.CHECK_LABEL_GROUPS
|
108 |
+
)
|
109 |
+
|
110 |
+
supported_entities = supported_entities if supported_entities else self.ENTITIES
|
111 |
+
self.model = (
|
112 |
+
model
|
113 |
+
if model
|
114 |
+
else pipeline(
|
115 |
+
"ner",
|
116 |
+
model=AutoModelForTokenClassification.from_pretrained(model_path),
|
117 |
+
tokenizer=AutoTokenizer.from_pretrained(model_path),
|
118 |
+
aggregation_strategy="simple",
|
119 |
+
)
|
120 |
+
)
|
121 |
+
|
122 |
+
super().__init__(
|
123 |
+
supported_entities=supported_entities, name="transformers Analytics",
|
124 |
+
)
|
125 |
+
|
126 |
+
def load(self) -> None:
|
127 |
+
"""Load the model, not used. Model is loaded during initialization."""
|
128 |
+
pass
|
129 |
+
|
130 |
+
def get_supported_entities(self) -> List[str]:
|
131 |
+
"""
|
132 |
+
Return supported entities by this model.
|
133 |
+
|
134 |
+
:return: List of the supported entities.
|
135 |
+
"""
|
136 |
+
return self.supported_entities
|
137 |
+
|
138 |
+
# Class to use transformers with Presidio as an external recognizer.
|
139 |
+
def analyze(
|
140 |
+
self, text: str, entities: List[str], nlp_artifacts: NlpArtifacts = None
|
141 |
+
) -> List[RecognizerResult]:
|
142 |
+
"""
|
143 |
+
Analyze text using Text Analytics.
|
144 |
+
|
145 |
+
:param text: The text for analysis.
|
146 |
+
:param entities: Not working properly for this recognizer.
|
147 |
+
:param nlp_artifacts: Not used by this recognizer.
|
148 |
+
:return: The list of Presidio RecognizerResult constructed from the recognized
|
149 |
+
transformers detections.
|
150 |
+
"""
|
151 |
+
|
152 |
+
results = []
|
153 |
+
ner_results = self.model(text)
|
154 |
+
|
155 |
+
# If there are no specific list of entities, we will look for all of it.
|
156 |
+
if not entities:
|
157 |
+
entities = self.supported_entities
|
158 |
+
|
159 |
+
for entity in entities:
|
160 |
+
if entity not in self.supported_entities:
|
161 |
+
continue
|
162 |
+
|
163 |
+
for res in ner_results:
|
164 |
+
if not self.__check_label(
|
165 |
+
entity, res["entity_group"], self.check_label_groups
|
166 |
+
):
|
167 |
+
continue
|
168 |
+
textual_explanation = self.DEFAULT_EXPLANATION.format(
|
169 |
+
res["entity_group"]
|
170 |
+
)
|
171 |
+
explanation = self.build_transformers_explanation(
|
172 |
+
round(res["score"], 2), textual_explanation
|
173 |
+
)
|
174 |
+
transformers_result = self._convert_to_recognizer_result(
|
175 |
+
res, explanation
|
176 |
+
)
|
177 |
+
|
178 |
+
results.append(transformers_result)
|
179 |
+
|
180 |
+
return results
|
181 |
+
|
182 |
+
def _convert_to_recognizer_result(self, res, explanation) -> RecognizerResult:
|
183 |
+
|
184 |
+
entity_type = self.PRESIDIO_EQUIVALENCES.get(
|
185 |
+
res["entity_group"], res["entity_group"]
|
186 |
+
)
|
187 |
+
transformers_score = round(res["score"], 2)
|
188 |
+
|
189 |
+
transformers_results = RecognizerResult(
|
190 |
+
entity_type=entity_type,
|
191 |
+
start=res["start"],
|
192 |
+
end=res["end"],
|
193 |
+
score=transformers_score,
|
194 |
+
analysis_explanation=explanation,
|
195 |
+
)
|
196 |
+
|
197 |
+
return transformers_results
|
198 |
+
|
199 |
+
def build_transformers_explanation(
|
200 |
+
self, original_score: float, explanation: str
|
201 |
+
) -> AnalysisExplanation:
|
202 |
+
"""
|
203 |
+
Create explanation for why this result was detected.
|
204 |
+
|
205 |
+
:param original_score: Score given by this recognizer
|
206 |
+
:param explanation: Explanation string
|
207 |
+
:return:
|
208 |
+
"""
|
209 |
+
explanation = AnalysisExplanation(
|
210 |
+
recognizer=self.__class__.__name__,
|
211 |
+
original_score=original_score,
|
212 |
+
textual_explanation=explanation,
|
213 |
+
)
|
214 |
+
return explanation
|
215 |
+
|
216 |
+
@staticmethod
|
217 |
+
def __check_label(
|
218 |
+
entity: str, label: str, check_label_groups: Tuple[Set, Set]
|
219 |
+
) -> bool:
|
220 |
+
return any(
|
221 |
+
[entity in egrp and label in lgrp for egrp, lgrp in check_label_groups]
|
222 |
+
)
|
223 |
+
|
224 |
+
|
225 |
+
if __name__ == "__main__":
|
226 |
+
|
227 |
+
from presidio_analyzer import AnalyzerEngine, RecognizerRegistry
|
228 |
+
|
229 |
+
transformers_recognizer = (
|
230 |
+
TransformersRecognizer()
|
231 |
+
) # This would download a large (~500Mb) model on the first run
|
232 |
+
|
233 |
+
registry = RecognizerRegistry()
|
234 |
+
registry.add_recognizer(transformers_recognizer)
|
235 |
+
|
236 |
+
analyzer = AnalyzerEngine(registry=registry)
|
237 |
+
|
238 |
+
results = analyzer.analyze(
|
239 |
+
"My name is Christopher and I live in Irbid.",
|
240 |
+
language="en",
|
241 |
+
return_decision_process=True,
|
242 |
+
)
|
243 |
+
for result in results:
|
244 |
+
print(result)
|
245 |
+
print(result.analysis_explanation)
|