app.py

import joblib
from sentence_transformers import CrossEncoder, SentenceTransformer
import streamlit as st
from sklearn.metrics.pairwise import cosine_similarity
from samples import get_samples
import textdistance
from sklearn.feature_extraction.text import TfidfVectorizer
from encode_sentences import encode_sentences

model_save_path = 'trained_model_stsbenchmark_bert-base-uncased'
bi_encoder = 'Bi-Encoder'
cross_encoder = 'Cross-Encoder'
levenshtein_distance = 'Levenshtein Distance'
tf_idf = 'TF-IDF'
random_forest = 'Random Forest'
title = 'Sentence Similarity with Transformers'
st.set_page_config(page_title=title, layout='wide', initial_sidebar_state='auto')
@st.cache_data
def cache_variables():
    tfidf_vectorizer = TfidfVectorizer()
    cross_encoder_trasformer = CrossEncoder(model_save_path)
    bi_encoder_trasformer = SentenceTransformer(model_save_path)
    random_forest_model = joblib.load('trained_model_random_forest.joblib')
    return tfidf_vectorizer, cross_encoder_trasformer, bi_encoder_trasformer, random_forest_model

@st.cache_data
def compute_similarity(sentence_1, sentence_2, comparison):
    if comparison == bi_encoder:
        return cosine_similarity([bi_encoder_trasformer.encode(sentence_1)], [bi_encoder_trasformer.encode(sentence_2)])[0][0]
    return cross_encoder_trasformer.predict([sentence_1, sentence_2])

tfidf_vectorizer, cross_encoder_trasformer, bi_encoder_trasformer, random_forest_model = cache_variables()

st.title(title)
st.write("This app takes two sentences and outputs their similarity score using a fine-tuned transformer model.")

# Example sentences section
test_samples = get_samples()
st.sidebar.header("Example Sentences")
example_1 = st.sidebar.radio(
    "Sentence 1", test_samples['sentence1'].values.tolist())
example_2 = st.sidebar.radio(
    "Sentence 2", test_samples['sentence2'].values.tolist())

# Input fields
sentence_1 = st.text_input("Enter Sentence 1:", example_1)
sentence_2 = st.text_input("Enter Sentence 2:", example_2)
comparison = st.selectbox("Comparicon:", [
                          bi_encoder, cross_encoder, levenshtein_distance, tf_idf, random_forest])

if st.button("Compare"):
    # Compute similarity
    if comparison in [bi_encoder, cross_encoder]:
        similarity = compute_similarity(sentence_1, sentence_2, comparison)
    elif comparison == levenshtein_distance:
        similarity = textdistance.levenshtein.normalized_similarity(
            sentence_1, sentence_2)
    elif comparison == tf_idf:
        similarity = cosine_similarity(
            tfidf_vectorizer.fit_transform([sentence_1, sentence_2]))[0][1]
    elif comparison == random_forest:
        similarity = random_forest_model.predict(encode_sentences(
            bi_encoder_trasformer, sentence_1, sentence_2))[0]
    st.markdown(
        f"<b style='font-size: 1.5em'>{comparison}</b> similarity score: <b style='font-size: 1.5em'>:red[{similarity:.4f}]</b>", unsafe_allow_html=True)
    st.write(
        "A higher score indicates greater similarity. The score ranges from 0 to 1.")