Spaces:
Sleeping
Sleeping
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# TODO: Visualizations for topic, text
|
2 |
+
# TODO: BERT model switch
|
3 |
+
# TODO: Default input
|
4 |
+
# TODO: Pre segmented sentences
|
5 |
+
# TODO: Progress bar
|
6 |
+
|
7 |
+
import streamlit as st
|
8 |
+
import pandas as pd
|
9 |
+
import re
|
10 |
+
import json
|
11 |
+
|
12 |
+
from utils import window, get_depths, get_local_maxima, compute_threshold, get_threshold_segments
|
13 |
+
|
14 |
+
st.write("loading ...")
|
15 |
+
|
16 |
+
import spacy
|
17 |
+
nlp = spacy.load('en_core_web_sm')
|
18 |
+
|
19 |
+
def print_list(lst):
|
20 |
+
for e in lst:
|
21 |
+
st.markdown("- " + e)
|
22 |
+
|
23 |
+
# Demo start
|
24 |
+
|
25 |
+
st.subheader("Topic Segmentation Demo")
|
26 |
+
|
27 |
+
uploaded_file = st.file_uploader("choose a text file", type=["txt"])
|
28 |
+
|
29 |
+
if uploaded_file is not None:
|
30 |
+
st.session_state["text"] = uploaded_file.getvalue().decode('utf-8')
|
31 |
+
|
32 |
+
st.write("OR")
|
33 |
+
|
34 |
+
input_text = st.text_area(
|
35 |
+
label="Enter text separated by newlines",
|
36 |
+
value="",
|
37 |
+
key="text",
|
38 |
+
height=150
|
39 |
+
|
40 |
+
)
|
41 |
+
|
42 |
+
button=st.button('Get Segments')
|
43 |
+
|
44 |
+
# Radio bar
|
45 |
+
# BERT or TOPIC
|
46 |
+
select_names = ["LDA Topic", "BERT"]
|
47 |
+
model = st.radio(label='Select model', options=select_names, index=0)
|
48 |
+
|
49 |
+
if (button==True) and input_text != "":
|
50 |
+
|
51 |
+
# Parse sample document and break it into sentences
|
52 |
+
texts = input_text.split('\n')
|
53 |
+
sents = []
|
54 |
+
for text in texts:
|
55 |
+
doc = nlp(text)
|
56 |
+
for sent in doc.sents:
|
57 |
+
sents.append(sent)
|
58 |
+
|
59 |
+
# Select tokens while ignoring punctuations and stopwords, and lowercase them
|
60 |
+
MIN_LENGTH = 3
|
61 |
+
tokenized_sents = [[token.lemma_.lower() for token in sent if
|
62 |
+
not token.is_stop and not token.is_punct and token.text.strip() and len(token) >= MIN_LENGTH]
|
63 |
+
for sent in sents]
|
64 |
+
|
65 |
+
|
66 |
+
st.write("building topic model ...")
|
67 |
+
|
68 |
+
# Build gensim dictionary and topic model
|
69 |
+
from gensim import corpora, models
|
70 |
+
import numpy as np
|
71 |
+
|
72 |
+
np.random.seed(123)
|
73 |
+
|
74 |
+
N_TOPICS = 5
|
75 |
+
N_PASSES = 5
|
76 |
+
|
77 |
+
dictionary = corpora.Dictionary(tokenized_sents)
|
78 |
+
bow = [dictionary.doc2bow(sent) for sent in tokenized_sents]
|
79 |
+
topic_model = models.LdaModel(corpus=bow, id2word=dictionary, num_topics=N_TOPICS, passes=N_PASSES)
|
80 |
+
|
81 |
+
###st.write(topic_model.show_topics())
|
82 |
+
|
83 |
+
|
84 |
+
st.write("inferring topics ...")
|
85 |
+
# Infer topics with minimum threshold
|
86 |
+
THRESHOLD = 0.05
|
87 |
+
doc_topics = list(topic_model.get_document_topics(bow, minimum_probability=THRESHOLD))
|
88 |
+
|
89 |
+
# st.write(doc_topics)
|
90 |
+
|
91 |
+
# get top k topics for each sentence
|
92 |
+
k = 3
|
93 |
+
top_k_topics = [[t[0] for t in sorted(sent_topics, key=lambda x: x[1], reverse=True)][:k]
|
94 |
+
for sent_topics in doc_topics]
|
95 |
+
# st.write(top_k_topics)
|
96 |
+
|
97 |
+
###st.write("apply window")
|
98 |
+
|
99 |
+
from itertools import chain
|
100 |
+
|
101 |
+
WINDOW_SIZE = 3
|
102 |
+
|
103 |
+
window_topics = window(top_k_topics, n=WINDOW_SIZE)
|
104 |
+
# assert(len(window_topics) == (len(tokenized_sents) - WINDOW_SIZE + 1))
|
105 |
+
window_topics = [list(set(chain.from_iterable(window))) for window in window_topics]
|
106 |
+
|
107 |
+
# Encode topics for similarity computation
|
108 |
+
|
109 |
+
from sklearn.preprocessing import MultiLabelBinarizer
|
110 |
+
|
111 |
+
binarizer = MultiLabelBinarizer(classes=range(N_TOPICS))
|
112 |
+
|
113 |
+
encoded_topic = binarizer.fit_transform(window_topics)
|
114 |
+
|
115 |
+
# Get similarities
|
116 |
+
|
117 |
+
st.write("generating segments ...")
|
118 |
+
|
119 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
120 |
+
|
121 |
+
sims_topic = [cosine_similarity([pair[0]], [pair[1]])[0][0] for pair in zip(encoded_topic, encoded_topic[1:])]
|
122 |
+
# plot
|
123 |
+
|
124 |
+
# Compute depth scores
|
125 |
+
depths_topic = get_depths(sims_topic)
|
126 |
+
# plot
|
127 |
+
|
128 |
+
# Get local maxima
|
129 |
+
filtered_topic = get_local_maxima(depths_topic, order=1)
|
130 |
+
# plot
|
131 |
+
|
132 |
+
###st.write("compute threshold")
|
133 |
+
# Automatic threshold computation
|
134 |
+
# threshold_topic = compute_threshold(depths_topic)
|
135 |
+
threshold_topic = compute_threshold(filtered_topic)
|
136 |
+
|
137 |
+
# topk_segments = get_topk_segments(filtered_topic, k=5)
|
138 |
+
# Select segments based on threshold
|
139 |
+
threshold_segments_topic = get_threshold_segments(filtered_topic, threshold_topic)
|
140 |
+
|
141 |
+
# st.write(threshold_topic)
|
142 |
+
|
143 |
+
###st.write("compute segments")
|
144 |
+
|
145 |
+
segment_ids = threshold_segments_topic + WINDOW_SIZE
|
146 |
+
|
147 |
+
segment_ids = [0] + segment_ids.tolist() + [len(sents)]
|
148 |
+
slices = list(zip(segment_ids[:-1], segment_ids[1:]))
|
149 |
+
|
150 |
+
segmented = [sents[s[0]: s[1]] for s in slices]
|
151 |
+
|
152 |
+
for segment in segmented[:-1]:
|
153 |
+
print_list([s.text for s in segment])
|
154 |
+
st.markdown("""---""")
|
155 |
+
print_list([s.text for s in segmented[-1]])
|