nlc-explorer / WNgen.py
Nathan Butters
attempt to remediate
401217e
#Import necessary libraries.
import re, nltk, pandas as pd, numpy as np, ssl, streamlit as st
from nltk.corpus import wordnet
import spacy
nlp = spacy.load("en_core_web_lg")
#Import necessary parts for predicting things.
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TextClassificationPipeline
import torch
import torch.nn.functional as F
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english")
model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english")
pipe = TextClassificationPipeline(model=model, tokenizer=tokenizer, return_all_scores=True)
#If an error is thrown that the corpus "omw-1.4" isn't discoverable you can use this code. (https://stackoverflow.com/questions/38916452/nltk-download-ssl-certificate-verify-failed)
'''try:
_create_unverified_https_context = ssl._create_unverified_context
except AttributeError:
pass
else:
ssl._create_default_https_context = _create_unverified_https_context
nltk.download('omw-1.4')'''
# A simple function to pull synonyms and antonyms using spacy's POS
def syn_ant(word,POS=False,human=True):
pos_options = ['NOUN','VERB','ADJ','ADV']
synonyms = []
antonyms = []
#WordNet hates spaces so you have to remove them
if " " in word:
word = word.replace(" ", "_")
if POS in pos_options:
for syn in wordnet.synsets(word, pos=getattr(wordnet, POS)):
for l in syn.lemmas():
current = l.name()
if human:
current = re.sub("_"," ",current)
synonyms.append(current)
if l.antonyms():
for ant in l.antonyms():
cur_ant = ant.name()
if human:
cur_ant = re.sub("_"," ",cur_ant)
antonyms.append(cur_ant)
else:
for syn in wordnet.synsets(word):
for l in syn.lemmas():
current = l.name()
if human:
current = re.sub("_"," ",current)
synonyms.append(current)
if l.antonyms():
for ant in l.antonyms():
cur_ant = ant.name()
if human:
cur_ant = re.sub("_"," ",cur_ant)
antonyms.append(cur_ant)
synonyms = list(set(synonyms))
antonyms = list(set(antonyms))
return synonyms, antonyms
def process_text(text):
doc = nlp(text.lower())
result = []
for token in doc:
if (token.is_stop) or (token.is_punct) or (token.lemma_ == '-PRON-'):
continue
result.append(token.lemma_)
return " ".join(result)
def clean_definition(syn):
#This function removes stop words from sentences to improve on document level similarity for differentiation.
if type(syn) is str:
synset = wordnet.synset(syn).definition()
elif type(syn) is nltk.corpus.reader.wordnet.Synset:
synset = syn.definition()
definition = nlp(process_text(synset))
return definition
def check_sim(a,b):
if type(a) is str and type(b) is str:
a = nlp(a)
b = nlp(b)
similarity = a.similarity(b)
return similarity
# Builds a dataframe dynamically from WordNet using NLTK.
def wordnet_df(word,POS=False,seed_definition=None):
pos_options = ['NOUN','VERB','ADJ','ADV']
synonyms, antonyms = syn_ant(word,POS,False)
#print(synonyms, antonyms) #for QA purposes
words = []
cats = []
#WordNet hates spaces so you have to remove them
m_word = word.replace(" ", "_")
#Allow the user to pick a seed definition if it is not provided directly to the function. Currently not working so it's commented out.
'''#commented out the way it was designed to allow for me to do it through Streamlit (keeping it for posterity, and for anyone who wants to use it without streamlit.)
for d in range(len(seed_definitions)):
print(f"{d}: {seed_definitions[d]}")
#choice = int(input("Which of the definitions above most aligns to your selection?"))
seed_definition = seed_definitions[choice]'''
try:
definition = seed_definition
except:
st.write("You did not supply a definition.")
if POS in pos_options:
for syn in wordnet.synsets(m_word, pos=getattr(wordnet, POS)):
if check_sim(process_text(seed_definition),process_text(syn.definition())) > .7:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
if len(synonyms) > 0:
for w in synonyms:
w = w.replace(" ","_")
for syn in wordnet.synsets(w, pos=getattr(wordnet, POS)):
if check_sim(process_text(seed_definition),process_text(syn.definition())) > .6:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
if len(antonyms) > 0:
for a in antonyms:
a = a.replace(" ","_")
for syn in wordnet.synsets(a, pos=getattr(wordnet, POS)):
if check_sim(process_text(seed_definition),process_text(syn.definition())) > .26:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
else:
for syn in wordnet.synsets(m_word):
if check_sim(process_text(seed_definition),process_text(syn.definition())) > .7:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
if len(synonyms) > 0:
for w in synonyms:
w = w.replace(" ","_")
for syn in wordnet.synsets(w):
if check_sim(process_text(seed_definition),process_text(syn.definition())) > .6:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
if len(antonyms) > 0:
for a in antonyms:
a = a.replace(" ","_")
for syn in wordnet.synsets(a):
if check_sim(process_text(seed_definition),process_text(syn.definition())) > .26:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
df = {"Categories":cats, "Words":words}
df = pd.DataFrame(df)
df = df.drop_duplicates().reset_index()
df = df.drop("index", axis=1)
return df
def eval_pred_test(text, return_all = False):
'''A basic function for evaluating the prediction from the model and turning it into a visualization friendly number.'''
preds = pipe(text)
neg_score = -1 * preds[0][0]['score']
sent_neg = preds[0][0]['label']
pos_score = preds[0][1]['score']
sent_pos = preds[0][1]['label']
prediction = 0
sentiment = ''
if pos_score > abs(neg_score):
prediction = pos_score
sentiment = sent_pos
elif abs(neg_score) > pos_score:
prediction = neg_score
sentiment = sent_neg
if return_all:
return prediction, sentiment
else:
return prediction
def get_parallel(word, seed_definition, QA=False):
cleaned = nlp(process_text(seed_definition))
root_syns = wordnet.synsets(word)
hypers = []
new_hypos = []
for syn in root_syns:
hypers.extend(syn.hypernyms())
for syn in hypers:
new_hypos.extend(syn.hyponyms())
hypos = list(set([syn for syn in new_hypos if cleaned.similarity(nlp(process_text(syn.definition()))) >=.75]))[:25]
# with st.sidebar:
# st.write(f"The number of hypos is {len(hypos)} during get Parallel at Similarity >= .75.") #QA
if len(hypos) <= 1:
hypos = root_syns
elif len(hypos) < 3:
hypos = list(set([syn for syn in new_hypos if cleaned.similarity(nlp(process_text(syn.definition()))) >=.5]))[:25] # added a cap to each
elif len(hypos) < 10:
hypos = list(set([syn for syn in new_hypos if cleaned.similarity(nlp(process_text(syn.definition()))) >=.66]))[:25]
elif len(hypos) >= 10:
hypos = list(set([syn for syn in new_hypos if cleaned.similarity(nlp(process_text(syn.definition()))) >=.8]))[:25]
if QA:
print(hypers)
print(hypos)
return hypers, hypos
else:
return hypos
# Builds a dataframe dynamically from WordNet using NLTK.
def wordnet_parallel_df(word,seed_definition=None):
words = []
cats = []
#WordNet hates spaces so you have to remove them
m_word = word.replace(" ", "_")
# add synonyms and antonyms for diversity
synonyms, antonyms = syn_ant(word)
words.extend(synonyms)
cats.extend(["synonyms" for n in range(len(synonyms))])
words.extend(antonyms)
cats.extend(["antonyms" for n in range(len(antonyms))])
try:
hypos = get_parallel(m_word,seed_definition)
except:
st.write("You did not supply a definition.")
#Allow the user to pick a seed definition if it is not provided directly to the function.
'''if seed_definition is None:
if POS in pos_options:
seed_definitions = [syn.definition() for syn in wordnet.synsets(m_word, pos=getattr(wordnet, POS))]
else:
seed_definitions = [syn.definition() for syn in wordnet.synsets(m_word)]
for d in range(len(seed_definitions)):
print(f"{d}: {seed_definitions[d]}")
choice = int(input("Which of the definitions above most aligns to your selection?"))
seed_definition = seed_definitions[choice]'''
#This is a QA section
# with st.sidebar:
# st.write(f"The number of hypos is {len(hypos)} during parallel df creation.") #QA
#Transforms hypos into lemmas
for syn in hypos:
cur_lemmas = syn.lemmas()
hypos = syn.hyponyms()
for hypo in hypos:
cur_lemmas.extend(hypo.lemmas())
for lemma in cur_lemmas:
ll = lemma.name()
cats.append(re.sub("_"," ", syn.name().split(".")[0]))
words.append(re.sub("_"," ",ll))
# with st.sidebar:
# st.write(f'There are {len(words)} words in the dataframe at the beginning of df creation.') #QA
df = {"Categories":cats, "Words":words}
df = pd.DataFrame(df)
df = df.drop_duplicates("Words").reset_index()
df = df.drop("index", axis=1)
return df
#@st.experimental_singleton(suppress_st_warning=True)
def cf_from_wordnet_df(seed,text,seed_definition=False):
seed_token = nlp(seed)
seed_POS = seed_token[0].pos_
#print(seed_POS) QA
try:
df = wordnet_parallel_df(seed,seed_definition)
except:
st.write("You did not supply a definition.")
df["text"] = df.Words.apply(lambda x: re.sub(r'\b'+seed+r'\b',x,text))
df["similarity"] = df.Words.apply(lambda x: seed_token[0].similarity(nlp(x)[0]))
df = df[df["similarity"] > 0].reset_index()
df.drop("index", axis=1, inplace=True)
df["pred"] = df.text.apply(eval_pred_test)
# added this because I think it will make the end results better if we ensure the seed is in the data we generate counterfactuals from.
df['seed'] = df.Words.apply(lambda x: 'seed' if x.lower() == seed.lower() else 'alternative')
return df