Notebook demonstrating how to control OpenRefine via a Python client.
Use the dbutlerdb/refine-client-py fork of PaulMakepeace/refine-client-py for Python3 support.
I'm not yet convinced this is sensible or that the python client is as useable as it might be? Things like pandas are perfectly serviceable for working with tabular data in a notebook, so why would we want to use the OpenRefine engine?
To start, ensure that the OpenRefine application server is running. You can start it from the notebook homepage (New -> OpenRefine Session.
The server connection is looked for on the default port 3333. This can be hardcoded as part of the nbopenrefineder OpneRefine start-up command.
import os
#We might have to spot these in the notebook server log - the terminal output where the notebook server is started
#If we're in MyBinder, that's not much help though?
#os.environ['OPENREFINE_HOST']='127.0.0.1'
#os.environ['OPENREFINE_PORT']='39113'
from open.refine import refine
#What is the distinction between these two?
#Can we make use of it in a Jupyter context somehow?
server = refine.RefineServer()
orefine = refine.Refine(server)
We can list any projects that currently exist:
orefine.list_projects().items()
Projects can be created via the API from a source data file.
%%capture
try:
import pandas as pd
except:
!pip install pandas
import pandas as pd
#Create a simple test data file as a CSV file
#via https://github.com/dbutlerdb/refine-client-py/blob/master/tests/data/duplicates.csv
data = '''email,name,state,gender,purchase
[email protected],Danny Baron,CA,M,TV
[email protected],Melanie White,NC,F,iPhone
[email protected],D. Baron,CA,M,Winter jacket
[email protected],Ben Tyler,NV,M,Flashlight
[email protected],Arthur Duff,OR,M,Dining table
[email protected],Daniel Baron,CA,M,Bike
[email protected],Jean Griffith,WA,F,Power drill
[email protected],Melanie White,NC,F,iPad
[email protected],Ben Morisson,FL,M,Amplifier
[email protected],Arthur Duff,OR,M,Night table'''
fn = 'test.csv'
with open(fn,'w') as f:
f.write(data)
Either copy and paste the above data into a new OpenRefine project as a clipboard item or load it in as a CSV data file.
We should now be able to see the project via the project index list:
#For some reason, the project does not appear to get named?
orefine.list_projects()
To open an OpenRefine project, we need to get it's key value from the project list. (Note that different projects may have the same name.)
dir(orefine)
KEY = '2165169842672' #Need to use a valid key
KEY = list(orefine.list_projects().keys())[0]
p = orefine.open_project(KEY)
#Inspect the column names
p.columns
#Rename a column
col = p.columns[-1]
p.rename_column(col,'{}2'.format(col))
p.columns, p.column_order
Hmmm, it looks like there's a duplicate there?
OpenRefine allows you to add a column derived from one or more other columns.
.add_column(oldColumn, newColumn, expression="value", column_insert_index=None, on_error='set-to-blank'])The value keyword denotes using the cell values from the original column.
p.add_column('name', 'name2', column_insert_index=2)
How do we access the history from the API? Or do we have to create one ourselves from command responses?
#Add a column based on the name
p.add_column('name', 'NAME', "toUppercase(value)")
For a list of supported GREL (Google Refine Expression Language) commands, see here.
The remove column OpenRefine option doesn't appear to be supported?
The export appears to work on the dataset with current filters set. This is distinct from the complere dataset.
import pandas as pd
from io import StringIO
pd.read_csv( StringIO( p.export(export_format='csv') ) )
def show_export(p):
display( pd.read_csv( StringIO( p.export(export_format='csv') ) ) )
Via https://stackoverflow.com/a/53742983/454773 it looks like there may be an options parmamter that allows for column selection. This is not currently supported by the Python client.
We can display the state of the table with current project filters set.
def show_table(p):
''' Display currently selected rows in the table. '''
cells = [ [col['v'] for col in row['cells']] for row in p.get_rows().rows.rows_response ]
df = pd.DataFrame( cells )
#The list of columns seems to include historical items
#But how do we also guarantee the current one? dicts are inherently unordered?
cols = ['Unnamed_{}'.format(i) for i in range(len(df.columns))]
for (k,v) in sorted(p.column_order.items(), key=lambda kv: kv[1]):
cols[v]=k
#Set the column names guessed at - is there a better way?
df.columns = cols
display ( df ) #columns = [n for n in p.column_order]
#How do we get the full list of column names?
show_table(p)
How can we manipulate the data table?
We can look up the current column names and column order as follows:
p.column_order
Rearrange column order:
p.reorder_columns(['email','name','state','gender','purchase2','NAME','name2'])
show_export(p)
p.reorder_columns(['email','name'])
show_export(p)
p.reorder_columns(['email','name','state','gender','purchase2','NAME','name2'])
show_export(p)
p.columns
Although the full set of columns aren't in the exportable view of the table, they do still exist:
p.get_rows().rows.rows_response[0]
So how do we reset that? We presumably need to unstep the project history?
Data in the table can be sorted on one or more columns, each ascending or descending.
from open.refine.facet import Sorting
sorting = Sorting('email')
sorting.criteria
Sorting(['name', 'gender']).criteria
p.sorting = Sorting('email')
p.reorder_rows()
show_table(p)
#Seems we can be more explicit - but are the missing items (e.g. valueType, caseSensitive) important?
Sorting([{'column':'name','reverse':True}]).criteria
p.sorting = Sorting(['email',{'column':'name','reverse':True}])
p.reorder_rows()
show_table(p)
p.get_rows().rows.rows_response[0]
It looks like we can use .engine.reset_all() to reset filters.
p.engine.reset_all()
show_table(p)
#The export filter still seems to be set though?
show_export(p)
#Also need an example with numeric type
How do we explore a project's history?
def show_history_item(h):
return h.time, h.id,h.description
show_history_item(p.history_entry)
This doesn't seem to work?
It would be useful to have a simple recipe for creating a project from a pandas dataframe.
import os
#Create an OpenRefine project from the data file
#Use the absolute path to the file
p2 = orefine.new_project(project_file=os.path.abspath(fn),
project_name='Test 1',
project_file_name=fn)
#Do we have to mediate this via a file transfer? eg could we go more directly from a pandas dataframe somehow?
#For some reason, the project does not appear to get named?
#There also can be a delay before the index listing shows that the data has been loaded?
orefine.list_projects()
Take care with this one — there is no warning...
p2.delete()
dir(refine.RefineProject)
OpenRefine supports facetting on a column, including facet counts, and several clustering algorithms.
from open.refine import facet
fr=p.compute_facets(facet.TextFacet('name'))
facets = fr.facets[0]
for k in sorted(facets.choices,
key=lambda k: facets.choices[k].count,
reverse=True):
print(facets.choices[k].count, k)
facet_value = facet.TextFacet(column='email',
selection=['[email protected]','[email protected]'])
facet_value.as_dict()
#The export doesn't appear to be affected?
p.engine.add_facet(facet_value)
p.compute_facets()
show(p)
#But the table view is?
show_table(p)
#Reset the facet filter
facet_value.reset()
facet_value.as_dict()
show_table(p)
Open Refine supports several clustering methods:
clusterer_type: binning; refine_function: fingerprint|metaphone3|cologne-phoneticclusterer_type: binning; refine_function: ngram-fingerprint; params: {'ngram-size': INT}clusterer_type: knn; refine_function: levenshtein|ppm; params: {'radius': FLOAT,'blocking-ngram-size': INT}clusters=p.compute_clusters('name',
clusterer_type='binning',
refine_function='cologne-phonetic')
for cluster in clusters:
print(cluster)
clusters=p.compute_clusters('name',
clusterer_type='knn',
refine_function='levenshtein',
params={'radius':3})
for cluster in clusters:
print(cluster)
So how do we then select and apply a correction amongst clustered values?
Reconciliation allows column values to be "reconciled" with data from third party services.
data='''Name
Diane Abbott
Boris Johnson
Boris Johnstone
Diana Abbot
Boris Johnston
Joanna Lumley
Boris Johnstone
'''
fn = 'test2.csv'
with open(fn,'w') as f:
f.write(data)
p=orefine.new_project(project_file=os.path.abspath(fn),
project_name='Test 2',
project_file_name=fn)
p.get_reconciliation_services()
p.get_reconciliation_service_by_name_or_url('Wikidata (en)')
#Get suggested types
p.guess_types_of_column('Name',
'https://tools.wmflabs.org/openrefine-wikidata/en/api')
reconciliation_config = {
'mode': 'standard-service',
'service': 'https://tools.wmflabs.org/openrefine-wikidata/en/preview?id={{id}}',
'identifierSpace': 'http://www.wikidata.org/entity/',
'schemaSpace': 'http://www.wikidata.org/prop/direct/',
'type': {
'id': 'Q5',
'name': 'human',
},
'autoMatch': True,
'columnDetails': [],
}
p.reconcile(column='Name',
service='Wikidata (en)',
reconciliation_config=reconciliation_config)
Reconciliation is asynchronous - how do we check the actual response when it arrives?
Code suggests call wait_until_idle(), but where? Perhaps requires a fork in the code to embed it in refine.py::RefineProject().reconcile() prior to providing the response?
pd.read_csv( StringIO( p.export(export_format='csv') ) )
See also: OpenRefine Style Reconciliation Containers. It would be good to weave reconciliation service recipes into Binderised demos, perhaps even using a Jupyter kernel gateway powered API to allow reconciliation service definitions via a notebook?
I need to figure some recipes for debugging in notebooks... Indeed, debugging anywhere!
#%%debug
import pdb; pdb.set_trace()
p.guess_types_of_column('Name', 'Wikidata (en)')
#c
pdb.pm()
#locals()['s']
#q