CONBERT / app.py
Shredder's picture
Update app.py
e9e58e0
from predict import run_prediction
from io import StringIO
import json
import gradio as gr
import spacy
from spacy import displacy
from transformers import AutoTokenizer, AutoModelForTokenClassification,RobertaTokenizer,pipeline
import torch
import nltk
from nltk.tokenize import sent_tokenize
from fin_readability_sustainability import BERTClass, do_predict
import pandas as pd
import en_core_web_sm
from fincat_utils import extract_context_words
from fincat_utils import bert_embedding_extract
import pickle
lr_clf = pickle.load(open("lr_clf_FiNCAT.pickle",'rb'))
nlp = en_core_web_sm.load()
nltk.download('punkt')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#SUSTAINABILITY STARTS
tokenizer_sus = RobertaTokenizer.from_pretrained('roberta-base')
model_sustain = BERTClass(2, "sustanability")
model_sustain.to(device)
model_sustain.load_state_dict(torch.load('sustainability_model.bin', map_location=device)['model_state_dict'])
def get_sustainability(text):
df = pd.DataFrame({'sentence':sent_tokenize(text)})
actual_predictions_sustainability = do_predict(model_sustain, tokenizer_sus, df)
highlight = []
for sent, prob in zip(df['sentence'].values, actual_predictions_sustainability[1]):
if prob>=4.384316:
highlight.append((sent, 'non-sustainable'))
elif prob<=1.423736:
highlight.append((sent, 'sustainable'))
else:
highlight.append((sent, '-'))
return highlight
#SUSTAINABILITY ENDS
#CLAIM STARTS
def score_fincat(txt):
li = []
highlight = []
txt = " " + txt + " "
k = ''
for word in txt.split():
if any(char.isdigit() for char in word):
if word[-1] in ['.', ',', ';', ":", "-", "!", "?", ")", '"', "'"]:
k = word[-1]
word = word[:-1]
st = txt.find(" " + word + k + " ")+1
k = ''
ed = st + len(word)
x = {'paragraph' : txt, 'offset_start':st, 'offset_end':ed}
context_text = extract_context_words(x)
features = bert_embedding_extract(context_text, word)
if(features[0]=='None'):
continue
#highlight.append((txt, ' '))
#return highlight
prediction = lr_clf.predict(features.reshape(1, 768))
prediction_probability = '{:.4f}'.format(round(lr_clf.predict_proba(features.reshape(1, 768))[:,1][0], 4))
highlight.append((word, ' In-claim' if prediction==1 else 'Out-of-Claim'))
# li.append([word,' In-claim' if prediction==1 else 'Out-of-Claim', prediction_probability])
else:
continue
# highlight.append((word, ' '))
if(len(highlight)<1):
highlight.append((txt,' '))
return highlight
##Summarization
summarizer = pipeline("summarization", model="knkarthick/MEETING_SUMMARY")
def summarize_text(text):
resp = summarizer(text)
stext = resp[0]['summary_text']
return stext
def split_in_sentences(text):
doc = nlp(text)
return [str(sent).strip() for sent in doc.sents]
def make_spans(text,results):
results_list = []
for i in range(len(results)):
results_list.append(results[i]['label'])
facts_spans = []
facts_spans = list(zip(split_in_sentences(text),results_list))
return facts_spans
##Forward Looking Statement
fls_model = pipeline("text-classification", model="yiyanghkust/finbert-fls", tokenizer="yiyanghkust/finbert-fls")
def fls(text):
results = fls_model(split_in_sentences(text))
return make_spans(text,results)
##Company Extraction
ner=pipeline('ner',model='Jean-Baptiste/camembert-ner-with-dates',tokenizer='Jean-Baptiste/camembert-ner-with-dates', aggregation_strategy="simple")
def fin_ner(text):
replaced_spans = ner(text)
new_spans=[]
for item in replaced_spans:
item['entity']=item['entity_group']
del item['entity_group']
new_spans.append(item)
return {"text": text, "entities": new_spans}
#CUAD STARTS
def load_questions():
questions = []
with open('questions.txt') as f:
questions = f.readlines()
return questions
def load_questions_short():
questions_short = []
with open('questionshort.txt') as f:
questions_short = f.readlines()
return questions_short
questions = load_questions()
questions_short = load_questions_short()
def quad(query,file):
with open(file.name) as f:
paragraph = f.read()
questions = load_questions()
questions_short = load_questions_short()
if (not len(paragraph)==0) and not (len(query)==0):
print('getting predictions')
predictions = run_prediction([query], paragraph, 'marshmellow77/roberta-base-cuad',n_best_size=5)
answer = ""
if predictions['0'] == "":
answer = 'No answer found in document'
else:
with open("nbest.json") as jf:
data = json.load(jf)
for i in range(1):
raw_answer=data['0'][i]['text']
answer += f"{data['0'][i]['text']} -- \n"
answer += f"Probability: {round(data['0'][i]['probability']*100,1)}%\n\n"
#summarizer = pipeline("summarization", model="knkarthick/MEETING_SUMMARY")
#resp = summarizer(answer)
#stext = resp[0]['summary_text']
# highlight,dff=score_fincat(answer)
return answer,summarize_text(answer),fin_ner(answer),score_fincat(answer),get_sustainability(answer),fls(answer)
# b6 = gr.Button("Get Sustainability")
#b6.click(get_sustainability, inputs = text, outputs = gr.HighlightedText())
#iface = gr.Interface(fn=get_sustainability, inputs="textbox", title="CONBERT",description="SUSTAINABILITY TOOL", outputs=gr.HighlightedText(), allow_flagging="never")
#iface.launch()
iface = gr.Interface(fn=quad, inputs=[gr.Dropdown(choices=questions_short,label='SEARCH QUERY'),gr.inputs.File(label='TXT FILE')], title="CONBERT",description="SUSTAINABILITY TOOL",article='Article', outputs=[gr.outputs.Textbox(label='Answer'),gr.outputs.Textbox(label='Summary'),gr.HighlightedText(label='NER'),gr.HighlightedText(label='CLAIM'),gr.HighlightedText(label='SUSTAINABILITY'),gr.HighlightedText(label='FLS')], allow_flagging="never")
iface.launch()