1
0
mirror of https://github.com/ilri/csv-metadata-quality.git synced 2024-11-23 06:10:19 +01:00
csv-metadata-quality/csv_metadata_quality/experimental.py
Alan Orth 8435ee242d
Experimental language detection using langid
Works decenty well assuming the title, abstract, and citation fields
are an accurate representation of the language as identified by the
language field. Handles ISO 639-1 (alpha 2) and ISO 639-3 (alpha 3)
values seamlessly.

This includes updated pipenv environment, test data, pytest tests
for both correct and incorrect ISO 639-1 and ISO 639-3 languages,
and a new command line option "-e".
2019-09-26 13:46:32 +03:00

96 lines
3.5 KiB
Python

import pandas as pd
def correct_language(row):
"""Analyze the text used in the title, abstract, and citation fields to pre-
dict the language being used and compare it with the item's dc.language.iso
field.
Function prints an error if the language field does not match the detected
language and returns the value in the language field if it does match.
"""
from pycountry import languages
import langid
import re
# Initialize some variables at global scope so that we can set them in the
# loop scope below and still be able to access them afterwards.
language = ""
sample_strings = list()
title = None
# Iterate over the labels of the current row's values. Before we transposed
# the DataFrame these were the columns in the CSV, ie dc.title and dc.type.
for label in row.axes[0]:
# Skip fields with missing values
if pd.isna(row[label]):
continue
# Check if current row has multiple language values (separated by "||")
match = re.match(r"^.*?language.*$", label)
if match is not None:
# Skip fields with multiple language values
if "||" in row[label]:
return
language = row[label]
# Extract title if it is present
match = re.match(r"^.*?title.*$", label)
if match is not None:
title = row[label]
# Append title to sample strings
sample_strings.append(row[label])
# Extract abstract if it is present
match = re.match(r"^.*?abstract.*$", label)
if match is not None:
sample_strings.append(row[label])
# Extract citation if it is present
match = re.match(r"^.*?citation.*$", label)
if match is not None:
sample_strings.append(row[label])
# Make sure language is not blank and is valid ISO 639-1/639-3 before proceeding with language prediction
if language != "":
# Check language value like "es"
if len(language) == 2:
if not languages.get(alpha_2=language):
return
# Check language value like "spa"
elif len(language) == 3:
if not languages.get(alpha_3=language):
return
# Language value is something else like "Span", do not proceed
else:
return
# Language is blank, do not proceed
else:
return
# Concatenate all sample strings into one string
sample_text = " ".join(sample_strings)
# Restrict the langid detection space to reduce false positives
langid.set_languages(
["ar", "de", "en", "es", "fr", "hi", "it", "ja", "ko", "pt", "ru", "vi", "zh"]
)
langid_classification = langid.classify(sample_text)
# langid returns an ISO 639-1 (alpha 2) representation of the detected language, but the current item's language field might be ISO 639-3 (alpha 3) so we should use a pycountry Language object to compare both represenations and give appropriate error messages that match the format used by in the input file.
detected_language = languages.get(alpha_2=langid_classification[0])
if len(language) == 2 and language != detected_language.alpha_2:
print(
f"Possibly incorrect language {language} (detected {detected_language.alpha_2}): {title}"
)
elif len(language) == 3 and language != detected_language.alpha_3:
print(
f"Possibly incorrect language {language} (detected {detected_language.alpha_3}): {title}"
)
else:
return language