"""

Building a local shallow syntax based language model.

If the positive ratio is expected to be high enough classify some texts.

That will give the ability to estimate the positive rate, recall and improve the model's recall.

In many cases, the positive rate is very low and therefore this method is not applicable.

Guess some terms related to the concept. Brainstorms are helpful.

Look for synonyms of the terms.

https://www.thesaurus.com/

Look of common tokens in text containing your terms (use terms_related_to_concept.sql).

Sort the terms by alphabetical order to group them by semantic meaning.

It will be valuable in maintaining.

Sample some hits containing the terms.

Browse them, see that you are OK and fine tune them.

Add build_non_positive_linguistic

Add a list of supersets of your terms to be excluded

Samples hits of core that are not hits of your model to improve model recall

Use tests and not just data sets in order to make sure that the classifier classify correctly cases of importance.

Archimedes

Use rules that disagree with your classifier concept (e.g., good vs. rejected PRs, low quality files), and use the cases

in which both your classifier and they agree as a possible rich source for false positives.

Similarly, rules that agree with the classifier that hit when it is not are a rich source of false negatives.

https://patents.google.com/patent/US20190164086A1/en

https://arxiv.org/pdf/2007.10912.pdf

Term evaluation

In low positive rate, random sampling is not feasible.

One can in an sample cases not covered by the current classifier yet identified by the new term.

This way one can get the precision and additional recall.

Maintain a list of cases that you are not certain about. Time helps.

"""

import re

from os.path import join

import pandas as pd

from configuration import DATA_PATH

from language_utils import regex_to_big_query, generate_bq_function, match, SCHEMA_NAME, print_logic_to_bq\

, build_separated_terms, build_non_positive_linguistic

from model_evaluation import classifiy_commits_df, evaluate_performance, evaluate_concept_classifier

# Not sure list

"""

superior, first-string

"""

improvement = ['better', 'improved', 'fitter', 'preferred', 'finer', 'greater'

, 'higher quality'

]

positive_terms =['A number 1',

'world-class']

excluded_terms = ['for good', 'good way', 'good news', 'great time']

def build_positive_regex():

return build_separated_terms(positive_terms)

def build_excluded_regex():

return build_separated_terms(excluded_terms)

def build_not_positive_regex():

return build_non_positive_linguistic(build_positive_regex())

def is_good(commit_text):

- len(re.findall(build_not_positive_regex(), commit_text))) > 0

def good_to_bq():

print("# end - " + concept)

def print_concepts_functions_for_bq(commit: str = 'XXX'):

print()

def evaluate_good_classifier():

, samples_file=join(DATA_PATH, 'good_texts_tests.csv'))

if __name__ == '__main__':

print_concepts_functions_for_bq(commit='fedd454d2bf47de43b2bc80d52172ab8aac33bc7')

evaluate_good_classifier()