Spaces:
Runtime error
Runtime error
Create new file
Browse files- appStore/keyword_search.py +208 -0
appStore/keyword_search.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set path
|
2 |
+
import glob, os, sys; sys.path.append('../udfPreprocess')
|
3 |
+
|
4 |
+
#import helper
|
5 |
+
import udfPreprocess.docPreprocessing as pre
|
6 |
+
import udfPreprocess.cleaning as clean
|
7 |
+
|
8 |
+
#import needed libraries
|
9 |
+
import seaborn as sns
|
10 |
+
from pandas import DataFrame
|
11 |
+
from sentence_transformers import SentenceTransformer, CrossEncoder, util
|
12 |
+
# from keybert import KeyBERT
|
13 |
+
from transformers import pipeline
|
14 |
+
import matplotlib.pyplot as plt
|
15 |
+
import numpy as np
|
16 |
+
import streamlit as st
|
17 |
+
import pandas as pd
|
18 |
+
from rank_bm25 import BM25Okapi
|
19 |
+
from sklearn.feature_extraction import _stop_words
|
20 |
+
import string
|
21 |
+
from tqdm.autonotebook import tqdm
|
22 |
+
import numpy as np
|
23 |
+
|
24 |
+
import tempfile
|
25 |
+
import sqlite3
|
26 |
+
|
27 |
+
def app():
|
28 |
+
|
29 |
+
with st.container():
|
30 |
+
st.markdown("<h1 style='text-align: center; color: black;'> Keyword Search</h1>", unsafe_allow_html=True)
|
31 |
+
st.write(' ')
|
32 |
+
st.write(' ')
|
33 |
+
|
34 |
+
with st.expander("ℹ️ - About this app", expanded=True):
|
35 |
+
|
36 |
+
st.write(
|
37 |
+
"""
|
38 |
+
The *Keyword Search* app is an easy-to-use interface built in Streamlit for doing keyword search in policy document - developed by GIZ Data and the Sustainable Development Solution Network.
|
39 |
+
"""
|
40 |
+
)
|
41 |
+
|
42 |
+
st.markdown("")
|
43 |
+
|
44 |
+
st.markdown("")
|
45 |
+
st.markdown("## 📌 Step One: Upload document ")
|
46 |
+
|
47 |
+
with st.container():
|
48 |
+
|
49 |
+
file = st.file_uploader('Upload PDF File', type=['pdf', 'docx', 'txt'])
|
50 |
+
|
51 |
+
if file is not None:
|
52 |
+
|
53 |
+
|
54 |
+
with tempfile.NamedTemporaryFile(mode="wb") as temp:
|
55 |
+
bytes_data = file.getvalue()
|
56 |
+
temp.write(bytes_data)
|
57 |
+
|
58 |
+
st.write("Filename: ", file.name)
|
59 |
+
|
60 |
+
# load document
|
61 |
+
docs = pre.load_document(temp.name, file)
|
62 |
+
|
63 |
+
# preprocess document
|
64 |
+
haystackDoc, dataframeDoc, textData, paraList = clean.preprocessing(docs)
|
65 |
+
|
66 |
+
# testing
|
67 |
+
# st.write(len(all_text))
|
68 |
+
# for i in par_list:
|
69 |
+
# st.write(i)
|
70 |
+
|
71 |
+
keyword = st.text_input("Please enter here what you want to search, we will look for similar context in the document.",
|
72 |
+
value="floods",)
|
73 |
+
|
74 |
+
@st.cache(allow_output_mutation=True)
|
75 |
+
def load_sentenceTransformer(name):
|
76 |
+
return SentenceTransformer(name)
|
77 |
+
|
78 |
+
bi_encoder = load_sentenceTransformer('msmarco-distilbert-cos-v5') # multi-qa-MiniLM-L6-cos-v1
|
79 |
+
bi_encoder.max_seq_length = 64 #Truncate long passages to 256 tokens
|
80 |
+
top_k = 32
|
81 |
+
|
82 |
+
#@st.cache(allow_output_mutation=True)
|
83 |
+
#def load_crossEncoder(name):
|
84 |
+
# return CrossEncoder(name)
|
85 |
+
|
86 |
+
# cross_encoder = load_crossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
|
87 |
+
document_embeddings = bi_encoder.encode(paraList, convert_to_tensor=True, show_progress_bar=False)
|
88 |
+
|
89 |
+
def bm25_tokenizer(text):
|
90 |
+
tokenized_doc = []
|
91 |
+
for token in text.lower().split():
|
92 |
+
token = token.strip(string.punctuation)
|
93 |
+
|
94 |
+
if len(token) > 0 and token not in _stop_words.ENGLISH_STOP_WORDS:
|
95 |
+
tokenized_doc.append(token)
|
96 |
+
return tokenized_doc
|
97 |
+
|
98 |
+
def bm25TokenizeDoc(paraList):
|
99 |
+
tokenized_corpus = []
|
100 |
+
for passage in tqdm(paraList):
|
101 |
+
if len(passage.split()) >256:
|
102 |
+
temp = " ".join(passage.split()[:256])
|
103 |
+
tokenized_corpus.append(bm25_tokenizer(temp))
|
104 |
+
temp = " ".join(passage.split()[256:])
|
105 |
+
tokenized_corpus.append(bm25_tokenizer(temp))
|
106 |
+
else:
|
107 |
+
tokenized_corpus.append(bm25_tokenizer(passage))
|
108 |
+
|
109 |
+
return tokenized_corpus
|
110 |
+
|
111 |
+
tokenized_corpus = bm25TokenizeDoc(paraList)
|
112 |
+
|
113 |
+
|
114 |
+
document_bm25 = BM25Okapi(tokenized_corpus)
|
115 |
+
|
116 |
+
|
117 |
+
def search(keyword):
|
118 |
+
##### BM25 search (lexical search) #####
|
119 |
+
bm25_scores = document_bm25.get_scores(bm25_tokenizer(keyword))
|
120 |
+
top_n = np.argpartition(bm25_scores, -10)[-10:]
|
121 |
+
bm25_hits = [{'corpus_id': idx, 'score': bm25_scores[idx]} for idx in top_n]
|
122 |
+
bm25_hits = sorted(bm25_hits, key=lambda x: x['score'], reverse=True)
|
123 |
+
|
124 |
+
##### Sematic Search #####
|
125 |
+
# Encode the query using the bi-encoder and find potentially relevant passages
|
126 |
+
#query = "Does document contain {} issues ?".format(keyword)
|
127 |
+
question_embedding = bi_encoder.encode(keyword, convert_to_tensor=True)
|
128 |
+
|
129 |
+
hits = util.semantic_search(question_embedding, document_embeddings, top_k=top_k)
|
130 |
+
hits = hits[0] # Get the hits for the first query
|
131 |
+
|
132 |
+
|
133 |
+
##### Re-Ranking #####
|
134 |
+
# Now, score all retrieved passages with the cross_encoder
|
135 |
+
#cross_inp = [[query, paraList[hit['corpus_id']]] for hit in hits]
|
136 |
+
#cross_scores = cross_encoder.predict(cross_inp)
|
137 |
+
|
138 |
+
# Sort results by the cross-encoder scores
|
139 |
+
#for idx in range(len(cross_scores)):
|
140 |
+
# hits[idx]['cross-score'] = cross_scores[idx]
|
141 |
+
|
142 |
+
|
143 |
+
return bm25_hits, hits
|
144 |
+
|
145 |
+
|
146 |
+
if st.button("Find them."):
|
147 |
+
bm25_hits, hits = search(keyword)
|
148 |
+
|
149 |
+
st.markdown("""
|
150 |
+
We will provide with 2 kind of results. The 'lexical search' and the semantic search.
|
151 |
+
""")
|
152 |
+
# In the semantic search part we provide two kind of results one with only Retriever (Bi-Encoder) and other the ReRanker (Cross Encoder)
|
153 |
+
st.markdown("Top few lexical search (BM25) hits")
|
154 |
+
for hit in bm25_hits[0:5]:
|
155 |
+
if hit['score'] > 0.00:
|
156 |
+
st.write("\t Score: {:.3f}: \t{}".format(hit['score'], paraList[hit['corpus_id']].replace("\n", " ")))
|
157 |
+
|
158 |
+
|
159 |
+
|
160 |
+
|
161 |
+
|
162 |
+
# st.table(bm25_hits[0:3])
|
163 |
+
|
164 |
+
st.markdown("\n-------------------------\n")
|
165 |
+
st.markdown("Top few Bi-Encoder Retrieval hits")
|
166 |
+
|
167 |
+
hits = sorted(hits, key=lambda x: x['score'], reverse=True)
|
168 |
+
for hit in hits[0:5]:
|
169 |
+
# if hit['score'] > 0.45:
|
170 |
+
st.write("\t Score: {:.3f}: \t{}".format(hit['score'], paraList[hit['corpus_id']].replace("\n", " ")))
|
171 |
+
#st.table(hits[0:3]
|
172 |
+
|
173 |
+
#st.markdown("-------------------------")
|
174 |
+
|
175 |
+
#hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
|
176 |
+
#st.markdown("Top few Cross-Encoder Re-ranker hits")
|
177 |
+
#for hit in hits[0:3]:
|
178 |
+
# st.write("\t Score: {:.3f}: \t{}".format(hit['cross-score'], paraList[hit['corpus_id']].replace("\n", " ")))
|
179 |
+
#st.table(hits[0:3]
|
180 |
+
|
181 |
+
|
182 |
+
|
183 |
+
|
184 |
+
|
185 |
+
#for hit in bm25_hits[0:3]:
|
186 |
+
# print("\t{:.3f}\t{}".format(hit['score'], paraList[hit['corpus_id']].replace("\n", " ")))
|
187 |
+
|
188 |
+
|
189 |
+
|
190 |
+
|
191 |
+
|
192 |
+
|
193 |
+
|
194 |
+
# Output of top-5 hits from bi-encoder
|
195 |
+
#print("\n-------------------------\n")
|
196 |
+
#print("Top-3 Bi-Encoder Retrieval hits")
|
197 |
+
#hits = sorted(hits, key=lambda x: x['score'], reverse=True)
|
198 |
+
#for hit in hits[0:3]:
|
199 |
+
# print("\t{:.3f}\t{}".format(hit['score'], paraList[hit['corpus_id']].replace("\n", " ")))
|
200 |
+
|
201 |
+
# Output of top-5 hits from re-ranker
|
202 |
+
# print("\n-------------------------\n")
|
203 |
+
#print("Top-3 Cross-Encoder Re-ranker hits")
|
204 |
+
# hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
|
205 |
+
# for hit in hits[0:3]:
|
206 |
+
# print("\t{:.3f}\t{}".format(hit['cross-score'], paraList[hit['corpus_id']].replace("\n", " ")))
|
207 |
+
|
208 |
+
|