Provenance tracing
Contents
Provenance tracing#
Warning
Provenance tracing is still under development and may be changed in the future.
One of the main features of medkit is the tracing of provenance information. When used, medkit is able to tell how each annotation was created, that is to say:
the operation that generated it;
the input data that was used by the operation to generate the annotation;
This is true for the whole processing chain, including intermediate steps and annotations.
The goal is to retain enough information to later output in the PROV-O format. More practically, it can also be useful to know how an annotation was generated in order to know if it is trustworthy or not.
This tutorial will teach you how to gather provenance information in medkit. Before you read it, you should be familiar with the medkit components exposed in the first steps and pipeline tutorials.
A minimalistic provenance graph#
Let’s start with the simplest use case possible and take a look at provenance
for a single annotation, generated by a single operation. We are going to create
a very simple TextDocument
containing just one sentence, and run a
RegexpMatcher
on it that will match a single Entity
:
from medkit.core.text import TextDocument
from medkit.text.ner import RegexpMatcher, RegexpMatcherRule
text = "Je souffre d'asthme."
doc = TextDocument(text=text)
regexp_rule = RegexpMatcherRule(regexp=r"\basthme\b", label="problem")
regexp_matcher = RegexpMatcher(rules=[regexp_rule])
Before we actually call the run()
method of our regexp matcher, we will
activate the tracing of provenance for the entities it creates. This is done by
assigning it a ProvTracer
object. The ProvTracer
is in
charge of gathering all provenance info across all the operations. Operations
need to know it because they will inform it of each annotation they create.
from medkit.core import ProvTracer
prov_tracer = ProvTracer()
regexp_matcher.set_prov_tracer(prov_tracer)
We may now run the regexp matcher which will, as expected, match one entity:
entities = regexp_matcher.run([doc.raw_segment])
for entity in entities:
print(f"text={entity.text!r}, label={entity.label}")
text='asthme', label=problem
Let’s retrieve and inspect the provenance info concerning this entity:
def print_prov(prov):
# data item
print(f"data_item={prov.data_item.text!r}")
# operation description (if available)
op_desc = prov.op_desc
print(f"op={op_desc.name if op_desc is not None else None}")
# source data items
print(f"source_items={[d.text for d in prov.source_data_items]}")
# derived data items
print(f"derived_items={[d.text for d in prov.derived_data_items]}", end="\n\n")
entity = entities[0]
prov = prov_tracer.get_prov(entity.uid)
print_prov(prov)
data_item='asthme'
op=RegexpMatcher
source_items=["Je souffre d'asthme."]
derived_items=[]
The get_prov()
method of ProvTracer
returns a simple
Prov
object containing all the provenance info related to
a specific object. It has the following attributes:
data_item
contains the object to which the provenance info refers. Here, it is our entity. Note that it doesn’t have to be anAnnotation
subclass. For instance, it could also be anAttribute
;op_desc
holds anOperationDescription
object, that describes the operation that created the data item, in our case the regexp matcher. TheOperationDescription
will contain the name of the operation and the init parameters that were used;source_data_items
contains the objects that were used by the operation to create the new data item. Here there is only one source, the raw text segment, because the entity was found in this particular segment by the regexp matcher. But it is possible to have more than one data item in the sources;reciprocally,
derived_data_items
contains the objects that were derived from the data item by further operations. In this simple example, there are none.
If we are interested in all the provenance info gathered by our ProvTracer
instance rather than the provenance of a specific item, then we can call the
get_provs()
method:
for prov in prov_tracer.get_provs():
print_prov(prov)
data_item='asthme'
op=RegexpMatcher
source_items=["Je souffre d'asthme."]
derived_items=[]
data_item="Je souffre d'asthme."
op=None
source_items=[]
derived_items=['asthme']
We can see that we have another Prov
object with partial provenance info about
the raw text segment: we know how it was used (the entity was derived from it)
but we don’t know how it was created. This is expected, as the raw segment is a
data item that was fed at the input of our processing flow, it was not created
by any operation.
Our provenance info has a graph structure, each Prov
object representing a
node. For visualization, medkit provides a
save_prov_to_dot()
helper function that generates
graphviz-compatible .dot
files:
Note
graphviz is a graph visualization tool that defines a
simple text-based format for describing graphs, the .dot
file format, and that
provides a dot
command-line executable to generate images from such files. You
will need to install graphviz on your system to be able to run the following
code. On an ubuntu system, apt install graphviz
should do the trick.
from pathlib import Path
import subprocess
from IPython.display import Image
from medkit.tools import save_prov_to_dot
def display_dot(dot_file):
png_file = dot_file.with_suffix(".png")
subprocess.run(["dot", "-Tpng", dot_file, "-o", png_file])
return Image(png_file)
output_dir = Path("_out")
output_dir.mkdir(exist_ok=True)
dot_file = output_dir / "prov.dot"
save_prov_to_dot(prov_tracer, dot_file)
display_dot(dot_file)
Provenance composition#
Let’s move on to a slightly more complex example: before using the
RegexpMatcher
matcher, we will split our document into sentences with a
SentenceTokenizer
. We will also wrap our SentenceTokenizer
and our
RegexpMatcher
in a pipeline:
from medkit.text.segmentation import SentenceTokenizer
from medkit.core.pipeline import PipelineStep, Pipeline
text = "Je souffre d'asthme. Je n'ai pas de diabète."
doc = TextDocument(text=text)
sent_tokenizer = SentenceTokenizer(output_label="sentence")
steps = [
PipelineStep(sent_tokenizer, input_keys=["full_text"], output_keys=["sentences"]),
PipelineStep(regexp_matcher, input_keys=["sentences"], output_keys=["entities"]),
]
pipeline = Pipeline(steps=steps, input_keys=["full_text"], output_keys=["entities"])
A pipeline being itself an operation, it also has a set_prov_tracer()
method,
and calling it will automatically enable provenance tracing for all the
operations in the pipeline.
Note
In this tutorial, we always use a new ProvTracer
instance for each example.
This is because the provenance tracer accumulates provenance information, but we
don’t want to keep the provenance information from the previous examples, so we
create a new one.
prov_tracer = ProvTracer()
pipeline.set_prov_tracer(prov_tracer)
entities = pipeline.run([doc.raw_segment])
for entity in entities:
print(f"text={entity.text!r}, label={entity.label}")
text='asthme', label=problem
As expected, the result is identical to the first example: we have matched one entity. However the provenance is structured differently:
for prov in prov_tracer.get_provs():
print_prov(prov)
data_item='asthme'
op=Pipeline
source_items=["Je souffre d'asthme. Je n'ai pas de diabète."]
derived_items=[]
data_item="Je souffre d'asthme. Je n'ai pas de diabète."
op=None
source_items=[]
derived_items=['asthme']
We can see that now, the operation that created the entity is not the
RegexpMatcher
anymore, but the Pipeline
. It might seem surprising but it
does make sense: the pipeline is a processing operation itself, it received as
input the raw segment, and used it to create an entity. The sentences are
considered internal intermediary results and are not listed.
However, if we are interested in the details about what happened inside the
Pipeline
, the information is still available through a sub-provenance tracer
that can be retrieved with get_sub_prov_tracer()
:
pipeline_prov_tracer = prov_tracer.get_sub_prov_tracer(pipeline.uid)
for prov in pipeline_prov_tracer.get_provs():
print_prov(prov)
data_item="Je souffre d'asthme"
op=SentenceTokenizer
source_items=["Je souffre d'asthme. Je n'ai pas de diabète."]
derived_items=['asthme']
data_item="Je souffre d'asthme. Je n'ai pas de diabète."
op=None
source_items=[]
derived_items=["Je souffre d'asthme", "Je n'ai pas de diabète"]
data_item="Je n'ai pas de diabète"
op=SentenceTokenizer
source_items=["Je souffre d'asthme. Je n'ai pas de diabète."]
derived_items=[]
data_item='asthme'
op=RegexpMatcher
source_items=["Je souffre d'asthme"]
derived_items=[]
Although the order of each Prov
returned by get_provs()
isn’t the order of
creation of the annotations themselves, we can see the details of what happened
in the pipeline: 2 sentences were derived from the raw text by the
SentenceTokenizer
, then one entity was derived from one of the sentences by
the RegexpMatcher
.
In other words, the provenance information held by the main ProvTracer
is
composed: it is a graph, but some part of the graph have corresponding nested
sub-graphs, that can be expanded if desired. The save_prov_to_dot()
helper is
able to leverage this structure. By default, it will expand and display all
sub-provenance info recursively, but it has a optional max_sub_prov_depth
parameter that allows to limit the depth of the sub-provenance to show:
# show only outer-most provenance
save_prov_to_dot(prov_tracer, dot_file, max_sub_prov_depth=0)
display_dot(dot_file)
# expand next level of sub-provenance
save_prov_to_dot(prov_tracer, dot_file, max_sub_prov_depth=1)
display_dot(dot_file)
The same way that pipeline can contain sub-pipelines recursively, the provenance tracer can contain sub-provenance tracers recursively for the corresponding sub-pipelines.
Composed provenance makes it possible to preserve exhaustive provenance information about our data but to chose the appropriate level of detail when inspecting it. The structure of the provenance will reflect the structure of the processing flow: if it is built in a composed way, with pipelines containing sub-pipelines dealing with specific sub-tasks, then the provenance information will be composed the same way.
A more complete provenance example#
To demonstrate a bit more the potential of provenance tracing in medkit, let’s build a more complicated pipeline involving a sub-pipeline and an operation that creates attributes:
from medkit.text.context import NegationDetector, NegationDetectorRule
# segmentation
sent_tokenizer = SentenceTokenizer(output_label="sentence")
# negation detection
neg_detector = NegationDetector(output_label="is_negated")
# entity recognition
regexp_rules = [
RegexpMatcherRule(regexp=r"\basthme\b", label="problem"),
RegexpMatcherRule(regexp=r"\bdiabète\b", label="problem"),
]
regexp_matcher = RegexpMatcher(rules=regexp_rules, attrs_to_copy=["is_negated"])
# context sub pipeline handling segmentation and negation detection
sub_pipeline_steps = [
PipelineStep(sent_tokenizer, input_keys=["full_text"], output_keys=["sentences"]),
PipelineStep(neg_detector, input_keys=["sentences"], output_keys=[]), # no output
]
sub_pipeline = Pipeline(
sub_pipeline_steps,
name="ContextPipeline",
input_keys=["full_text"],
output_keys=["sentences"],
)
# main pipeline
pipeline_steps = [
PipelineStep(sub_pipeline, input_keys=["full_text"], output_keys=["sentences"]),
PipelineStep(regexp_matcher, input_keys=["sentences"], output_keys=["entities"]),
]
pipeline = Pipeline(
pipeline_steps,
name="MainPipeline",
input_keys=["full_text"],
output_keys=["entities"],
)
Note that since we have 2 pipelines, we pass an optional name
parameter to
each of them that will be used in the operation description and will help us to
distinguish them.
Running the pipeline gives us 2 entities with negation attributes:
prov_tracer = ProvTracer()
pipeline.set_prov_tracer(prov_tracer)
entities = pipeline.run([doc.raw_segment])
for entity in entities:
is_negated = entity.attrs.get(label="is_negated")[0].value
print(f"text={entity.text!r}, label={entity.label}, is_negated={is_negated}")
text='asthme', label=problem, is_negated=False
text='diabète', label=problem, is_negated=True
At the outer-most level, the provenance tells us that the main pipeline created
2 entities and 2 attributes. Intermediary data items (sentences) and operations
(SentenceTokenizer
, NegationDetector
, RegexpMatcher
) are hidden .
save_prov_to_dot(prov_tracer, dot_file, max_sub_prov_depth=0)
display_dot(dot_file)
You can see dotted arrow showing which attribute relates to which annotation.
While this is not strictly speaking provenance information, it is displayed
nonetheless to avoid any confusion, especially in the case where attributes
created by one operation are afterwards copied to new annotations (cf
attrs_to_copy
as explained in the First steps tutorial).
Expanding one more level of sub-provenance gives us the following graph:
save_prov_to_dot(prov_tracer, dot_file, max_sub_prov_depth=1)
display_dot(dot_file)
We now see the details of the operations and data items handled in our main
pipeline: a sub-pipeline created sentence segments and negation
attributes, then the RegexpMatcher
created entities, using the sentences
segments. The negation attributes were attached to both the sentences and the
entities derived from the sentences.
To have more details about the processing inside the context sub-pipeline, we have to go one step deeper:
save_prov_to_dot(prov_tracer, dot_file, max_sub_prov_depth=2)
display_dot(dot_file)
Wrapping it up#
In this tutorial, we have seen how we can use ProvTracer
to keep information
about how annotations and attributes were generated, ie. which operation created
them, using which data as input.
We have also seen how, when using pipelines and sub-pipelines, the provenance
information in a ProvTracer
will be composed, the same way that our processing
graph is. This allows us to later display the level of details that we want to
see when inspecting provenance.
Finally, we have seen how the save_prov_to_dot()
helper function can be used
to quickly visualize the captured provenance information. For more advanced
provenance usage, you may want to look at the provenance API
docs. The source code of save_prov_to_dot()
can also
serve as a reference on how to use it.