This is some snippets of code to get started with Snorkel

snorkel_example_code.ipynb

# make sure all packages import
from snorkel.labeling import labeling_function
from snorkel.utils import to_int_label_array
from snorkel.labeling import PandasLFApplier

@labeling_function()
def child_in_msg(x):
    """Postive examples (unknown otherwise) if child's name is in message.
    """
    return 1 if x.pat_first_in_msg == 1 else -1


@labeling_function()
def child_is_older(x):
    """Positive and negative examples for different cut-offs for child's age.
    """
    if x.age < 15.5:
        return 1
    elif 15.5 <= x.age < 17:
        return -1
    elif 16 <= x.age:
        return 0


@labeling_function()
def email_domain(x):
    """Gmail is a strong signal for not being a parent in text.
    """
    return 0 if "gmail" in x.email_domain else -1


@labeling_function()
def is_proxy(x):
    """Positive example (0 otherwise) if a proxy account.
    """
    return 1 if x.proxy_account_yn == "Y" else 0


@labeling_function()
def email_domain2(x):
    """Oracle is not a email domain easily accessable to teens.
    """
    return 1 if "oracle" not in x.email_domain else -1

# Assign labeling functions
lfs = [child_in_msg, child_is_older, email_domain, email_domain2, reading_fog, is_proxy]

applier = PandasLFApplier(lfs=lfs)
L_train = applier.apply(df=X_train)
L_dev = applier.apply(df=X_val)


# Check coverage for different labeling functions
c1, c2, c3, c4, c5, c6 = (L_train != -1).mean(axis=0)
print(f"check_out coverage: {c1 * 100:.1f}%")
print(f"check coverage: {c2 * 100:.1f}%")
print(f"check coverage: {c3 * 100:.1f}%")
print(f"check coverage: {c4 * 100:.1f}%")
print(f"check coverage: {c5 * 100:.1f}%")
print(f"check coverage: {c6 * 100:.1f}%")

# Full analysis (including coverage) of how well labeling functions are performing
from snorkel.labeling import LFAnalysis

LFAnalysis(L_train, lfs).lf_summary()



#####################
# Labeling functions looking at text
#####################
from snorkel.labeling import LabelingFunction


def keyword_lookup(x, keywords, label):
    if any(word in x.msg_txt.lower() for word in keywords):
        return label
    return -1


def make_keyword_lf(keywords, label=1):
    return LabelingFunction(
        name=f"keyword_{keywords[0]}",
        f=keyword_lookup,
        resources=dict(keywords=keywords, label=label),
    )


#######################################################################
# Positive or unknown keywords
# In my case I was interested in giveaways that a parent was communicating
# with a clinician versus the teen (who should have been communicating)
#######################################################################
keyword_my_son = make_keyword_lf(keywords=["son"])

keyword_my_daughter = make_keyword_lf(keywords=["my daughter"])

keyword_my_child = make_keyword_lf(keywords=["my child", "your child"])


keyword_son = make_keyword_lf(keywords=["son", "he", "him"])

keyword_referring = make_keyword_lf(keywords=["for us", "our"])

keyword_daughter = make_keyword_lf(keywords=["daughter", "she", "her"])

keyword_child = make_keyword_lf(keywords=["child", "parent", "child's"])

#######################################################################
# Negative or unknown keywords
#######################################################################
keyword_my_parent = make_keyword_lf(
    keywords=["father", "mother", "mom", "dad", "my mom", "my dad"], label=0
)

lfs = [
    child_in_msg,
    # keyword_my_son,
    keyword_my_daughter,
    keyword_my_child,
    keyword_son,
    keyword_daughter,
    keyword_child,
    child_is_older,
    reading_fog,
    # Negative examples
    keyword_my_parent,
    email_domain,
    email_domain2,
    keyword_referring,
    is_proxy,
]

applier = PandasLFApplier(lfs=lfs)
L_train = applier.apply(df=X_train)
L_dev = applier.apply(df=X_val)
L_valid = applier.apply(df=X_test)