app.py

import torch
import streamlit as st
from transformers import RobertaTokenizer, RobertaForSequenceClassification
import nltk
from nltk.corpus import stopwords
import re
import string

nltk.download('stopwords')
nltk.download('punkt')
stop_words = set(stopwords.words('english'))
stop_words.discard('and')


def tokenize_sentences(sentence):
    encoded_dict = tokenizer.encode_plus(
        sentence,
        add_special_tokens=True,
        max_length=128,
        padding='max_length',
        truncation=True,
        return_attention_mask=True,
        return_tensors='pt'
    )
    return torch.cat([encoded_dict['input_ids']], dim=0), torch.cat([encoded_dict['attention_mask']], dim=0)

def remove_stop_words(sentence):
    words = nltk.word_tokenize(sentence)
    custom_words = ['recommend', 'having', 'Hello', 'best', 'restaurant', 'top', 'want', 'need', 'well', 'most', 'should', 'be', 'good', 'also']
    stop_words.update(custom_words)
    words_without_stopwords = [word for word in words if word.lower() not in stop_words]
    sentence_without_stopwords = ' '.join(words_without_stopwords)
    return sentence_without_stopwords

def preprocess_query(query):
    query = str(query).lower()
    query = query.strip()
    query = remove_stop_words(query)
    query=query.translate(str.maketrans("", "", string.punctuation))
    return query

def predict_aspects(sentence, threshold):
    input_ids, attention_mask = tokenize_sentences(sentence)
    with torch.no_grad():
        outputs = aspects_model(input_ids, attention_mask=attention_mask)
    logits = outputs.logits
    predicted_aspects = torch.sigmoid(logits).squeeze().tolist()
    results = dict()
    for label, prediction in zip(LABEL_COLUMNS_ASPECTS, predicted_aspects):
        if prediction < threshold:
            continue
        precentage = round(float(prediction) * 100, 2)
        results[label] = precentage
    return results

# Load tokenizer and model
BERT_MODEL_NAME_FOR_ASPECTS_CLASSIFICATION = 'roberta-large'
tokenizer = RobertaTokenizer.from_pretrained(BERT_MODEL_NAME_FOR_ASPECTS_CLASSIFICATION, do_lower_case=True)

LABEL_COLUMNS_ASPECTS = ['FOOD-CUISINE', 'FOOD-DEALS', 'FOOD-DIET_OPTION', 'FOOD-EXPERIENCE', 'FOOD-FLAVOR', 'FOOD-GENERAL', 'FOOD-INGREDIENT', 'FOOD-KITCHEN', 'FOOD-MEAL', 'FOOD-MENU', 'FOOD-PORTION', 'FOOD-PRESENTATION', 'FOOD-PRICE', 'FOOD-QUALITY', 'FOOD-RECOMMENDATION', 'FOOD-TASTE', 'GENERAL-GENERAL', 'RESTAURANT-ATMOSPHERE', 'RESTAURANT-BUILDING', 'RESTAURANT-DECORATION', 'RESTAURANT-EXPERIENCE', 'RESTAURANT-FEATURES', 'RESTAURANT-GENERAL', 'RESTAURANT-HYGIENE', 'RESTAURANT-KITCHEN', 'RESTAURANT-LOCATION', 'RESTAURANT-OPTIONS', 'RESTAURANT-RECOMMENDATION', 'RESTAURANT-SEATING_PLAN', 'RESTAURANT-VIEW', 'SERVICE-BEHAVIOUR', 'SERVICE-EXPERIENCE', 'SERVICE-GENERAL', 'SERVICE-WAIT_TIME']

aspects_model = RobertaForSequenceClassification.from_pretrained(BERT_MODEL_NAME_FOR_ASPECTS_CLASSIFICATION, num_labels=len(LABEL_COLUMNS_ASPECTS))
aspects_model.load_state_dict(torch.load('./Aspects_Extraction_Model_updated.pth', map_location=torch.device('cpu')))
aspects_model.eval()

# Streamlit App
st.title("Implicit and Explicit Aspect Extraction")

sentence = st.text_input("Enter a sentence:")
threshold = st.slider("Threshold", min_value=0.0, max_value=1.0, step=0.01, value=0.5)

if sentence:
    processed_sentence = preprocess_query(sentence)
    results = predict_aspects(processed_sentence, threshold)
    if len(results) > 0:
        st.write("Predicted Aspects:")
        table_data = [["Aspect", "Probability"]]
        for aspect, percentage in results.items():
            aspect_parts = aspect.split("-")
            table_data.append(aspect_parts + [f"{percentage}%"])
        st.table(table_data)
    else:
        st.write("No aspects above the threshold.")