app.py

import streamlit as st
from streamlit_tags import st_tags, st_tags_sidebar
from keytotext import pipeline
from PIL import Image

import json
from sentence_transformers import SentenceTransformer, CrossEncoder, util
import gzip
import os
import torch
import pickle
import random
import numpy as np

############
## Main page
############

st.write("# Demonstration for Etsy Query Expansion（Etsy-QE）")

st.markdown("***Idea is to build a model which will take query as inputs and generate expansion information as outputs.***")
image = Image.open('etsy-shop-LLC.png')
st.image(image)

st.sidebar.write("# Top-N Selection")
maxtags_sidebar = st.sidebar.slider('Number of query allowed?', 1, 20, 1, key='ehikwegrjifbwreuk')
#user_query = st_tags(
#    label='# Enter Query:',
#    text='Press enter to add more',
#    value=['Mother'],
#    suggestions=['gift', 'nike', 'wool'],
#    maxtags=maxtags_sidebar,
#    key="aljnf")

user_query = st.text_input("Enter a query for the generated text: e.g., gift, home decoration ...")

# Add selectbox in streamlit
option1 = st.sidebar.selectbox(
     'Which transformers model would you like to be selected?',
     ('multi-qa-MiniLM-L6-cos-v1','null','null'))

option2 = st.sidebar.selectbox(
     'Which corss-encoder model would you like to be selected?',
     ('cross-encoder/ms-marco-MiniLM-L-6-v2','null','null'))

st.sidebar.success("Load Successfully!")

#if not torch.cuda.is_available():
#    print("Warning: No GPU found. Please add GPU to your notebook")

#We use the Bi-Encoder to encode all passages, so that we can use it with sematic search
bi_encoder = SentenceTransformer(option1,device='cpu')
bi_encoder.max_seq_length = 256    #Truncate long passages to 256 tokens
top_k = 32                          #Number of passages we want to retrieve with the bi-encoder

#The bi-encoder will retrieve 100 documents. We use a cross-encoder, to re-rank the results list to improve the quality
cross_encoder = CrossEncoder(option2, device='cpu')

passages = []

# load pre-train embeedings files
embedding_cache_path = 'etsy-embeddings-cpu.pkl'
print("Load pre-computed embeddings from disc")
with open(embedding_cache_path, "rb") as fIn:
  cache_data = pickle.load(fIn)
  passages = cache_data['sentences']
  corpus_embeddings = cache_data['embeddings']

from rank_bm25 import BM25Okapi
from sklearn.feature_extraction import _stop_words
import string
from tqdm.autonotebook import tqdm
import numpy as np
import re

import yake

language = "en"
max_ngram_size = 3
deduplication_threshold = 0.9
deduplication_algo = 'seqm'
windowSize = 3
numOfKeywords = 3

custom_kw_extractor = yake.KeywordExtractor(lan=language, n=max_ngram_size, dedupLim=deduplication_threshold, dedupFunc=deduplication_algo, windowsSize=windowSize, top=numOfKeywords, features=None)

# We lower case our text and remove stop-words from indexing
def bm25_tokenizer(text):
    tokenized_doc = []
    for token in text.lower().split():
        token = token.strip(string.punctuation)

        if len(token) > 0 and token not in _stop_words.ENGLISH_STOP_WORDS:
            tokenized_doc.append(token)
    return tokenized_doc

tokenized_corpus = []
for passage in tqdm(passages):
    tokenized_corpus.append(bm25_tokenizer(passage))

bm25 = BM25Okapi(tokenized_corpus)

def word_len(s):
    return len([i for i in s.split(' ') if i])


# This function will search all wikipedia articles for passages that
# answer the query
def search(query):
    print("Input query:", query)
    total_qe = []

    ##### BM25 search (lexical search) #####
    bm25_scores = bm25.get_scores(bm25_tokenizer(query))
    top_n = np.argpartition(bm25_scores, -5)[-5:]
    bm25_hits = [{'corpus_id': idx, 'score': bm25_scores[idx]} for idx in top_n]
    bm25_hits = sorted(bm25_hits, key=lambda x: x['score'], reverse=True)
    
    #print("Top-10 lexical search (BM25) hits")
    qe_string = []
    for hit in bm25_hits[0:1000]:
      if passages[hit['corpus_id']].replace("\n", " ") not in qe_string:
        qe_string.append(passages[hit['corpus_id']].replace("\n", ""))

    sub_string = []
    for item in qe_string:
      for sub_item in item.split(","):
        sub_string.append(sub_item)
    #print(sub_string)
    total_qe.append(sub_string)

    ##### Sematic Search #####
    # Encode the query using the bi-encoder and find potentially relevant passages
    query_embedding = bi_encoder.encode(query, convert_to_tensor=True)
    hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=top_k)
    hits = hits[0]  # Get the hits for the first query

    ##### Re-Ranking #####
    # Now, score all retrieved passages with the cross_encoder
    cross_inp = [[query, passages[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)

    # Sort results by the cross-encoder scores
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]

    # Output of top-10 hits from bi-encoder
    #print("\n-------------------------\n")
    #print("Top-N Bi-Encoder Retrieval hits")
    hits = sorted(hits, key=lambda x: x['score'], reverse=True)
    qe_string = []
    for hit in hits[0:1000]:
      if passages[hit['corpus_id']].replace("\n", " ") not in qe_string:
        qe_string.append(passages[hit['corpus_id']].replace("\n", ""))
    #print(qe_string)
    total_qe.append(qe_string)

    # Output of top-10 hits from re-ranker
    #print("\n-------------------------\n")
    #print("Top-N Cross-Encoder Re-ranker hits")
    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
    qe_string = []
    for hit in hits[0:1000]:
      if passages[hit['corpus_id']].replace("\n", " ") not in qe_string:
        qe_string.append(passages[hit['corpus_id']].replace("\n", ""))
    #print(qe_string)
    total_qe.append(qe_string)

    # Total Results
    total_qe.append(qe_string)
    st.write("E-Commerce Query Expansion Results: \n")

    res = []
    for sub_list in total_qe:
      for i in sub_list:
        rs = re.sub("([^\u0030-\u0039\u0041-\u007a])", ' ', i)
        rs_final = re.sub("\x20\x20", "\n", rs)
        #st.write(rs_final.strip())
        res.append(rs_final.strip())

    #st.write(res[0:maxtags_sidebar])

    res_clean = []
    for out in res:
      if len(out) > 20:
        keywords = custom_kw_extractor.extract_keywords(out)
        for key in keywords:
          res_clean.append(key[0])
      else:
        res_clean.append(out)

    show_out = []
    for i in res_clean:
      num = word_len(i)
      if num > 1:
        show_out.append(i)

    #st.write(show_out[0:maxtags_sidebar])
    for i in show_out[0:maxtags_sidebar]:
        st.write(i)

    return show_out

def search_nolog(query):
    total_qe = []
    ##### BM25 search (lexical search) #####
    bm25_scores = bm25.get_scores(bm25_tokenizer(query))
    top_n = np.argpartition(bm25_scores, -5)[-5:]
    bm25_hits = [{'corpus_id': idx, 'score': bm25_scores[idx]} for idx in top_n]
    bm25_hits = sorted(bm25_hits, key=lambda x: x['score'], reverse=True)
    
    qe_string = []
    for hit in bm25_hits[0:1000]:
      if passages[hit['corpus_id']].replace("\n", " ") not in qe_string:
        qe_string.append(passages[hit['corpus_id']].replace("\n", ""))

    sub_string = []
    for item in qe_string:
      for sub_item in item.split(","):
        sub_string.append(sub_item)
    total_qe.append(sub_string)

    ##### Sematic Search #####
    # Encode the query using the bi-encoder and find potentially relevant passages
    query_embedding = bi_encoder.encode(query, convert_to_tensor=True)
    hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=top_k)
    hits = hits[0]  # Get the hits for the first query

    ##### Re-Ranking #####
    # Now, score all retrieved passages with the cross_encoder
    cross_inp = [[query, passages[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)

    # Sort results by the cross-encoder scores
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]

    # Output of top-10 hits from bi-encoder
    hits = sorted(hits, key=lambda x: x['score'], reverse=True)
    qe_string = []
    for hit in hits[0:1000]:
      if passages[hit['corpus_id']].replace("\n", " ") not in qe_string:
        qe_string.append(passages[hit['corpus_id']].replace("\n", ""))
    total_qe.append(qe_string)

    # Output of top-10 hits from re-ranker
    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
    qe_string = []
    for hit in hits[0:1000]:
      if passages[hit['corpus_id']].replace("\n", " ") not in qe_string:
        qe_string.append(passages[hit['corpus_id']].replace("\n", ""))
    total_qe.append(qe_string)

    # Total Results
    total_qe.append(qe_string)

    res = []
    for sub_list in total_qe:
      for i in sub_list:
        rs = re.sub("([^\u0030-\u0039\u0041-\u007a])", ' ', i)
        rs_final = re.sub("\x20\x20", "\n", rs)
        res.append(rs_final.strip())

    res_clean = []
    for out in res:
      if len(out) > 20:
        keywords = custom_kw_extractor.extract_keywords(out)
        for key in keywords:
          res_clean.append(key[0])
      else:
        res_clean.append(out)

    show_out = []
    for i in res_clean:
      num = word_len(i)
      if num > 1:
        show_out.append(i)

    return show_out

def reranking():
    rerank_list = []
    reres = []
    rerank_list = search_nolog(query = user_query)
    st.write(rerank_list[0:maxtags_sidebar])
    for i in rerank_list[0:maxtags_sidebar]:
        reres.append(i)
    np.random.seed(7)
    np.random.shuffle(reres)
    for j in reres:
        st.write(j)

st.write("## Results:")
if st.button('Generated Expansion'):
    out_res = search(query = user_query)
    #st.success(out_res)

if st.button('Rerank'):
    out_res = reranking()
    #st.success(out_res)