Spaces:
Build error
Build error
hellopahe
commited on
Commit
·
e7699c1
1
Parent(s):
90f83ff
add lexrank
Browse files- LexRank.py +124 -0
- app.py +50 -8
- article_extractor/tokenizers_pegasus.py +1 -1
- requirements.txt +5 -1
LexRank.py
ADDED
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
LexRank implementation
|
3 |
+
Source: https://github.com/crabcamp/lexrank/tree/dev
|
4 |
+
"""
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
from scipy.sparse.csgraph import connected_components
|
8 |
+
from scipy.special import softmax
|
9 |
+
import logging
|
10 |
+
|
11 |
+
logger = logging.getLogger(__name__)
|
12 |
+
|
13 |
+
def degree_centrality_scores(
|
14 |
+
similarity_matrix,
|
15 |
+
threshold=None,
|
16 |
+
increase_power=True,
|
17 |
+
):
|
18 |
+
if not (
|
19 |
+
threshold is None
|
20 |
+
or isinstance(threshold, float)
|
21 |
+
and 0 <= threshold < 1
|
22 |
+
):
|
23 |
+
raise ValueError(
|
24 |
+
'\'threshold\' should be a floating-point number '
|
25 |
+
'from the interval [0, 1) or None',
|
26 |
+
)
|
27 |
+
|
28 |
+
if threshold is None:
|
29 |
+
markov_matrix = create_markov_matrix(similarity_matrix)
|
30 |
+
|
31 |
+
else:
|
32 |
+
markov_matrix = create_markov_matrix_discrete(
|
33 |
+
similarity_matrix,
|
34 |
+
threshold,
|
35 |
+
)
|
36 |
+
|
37 |
+
scores = stationary_distribution(
|
38 |
+
markov_matrix,
|
39 |
+
increase_power=increase_power,
|
40 |
+
normalized=False,
|
41 |
+
)
|
42 |
+
|
43 |
+
return scores
|
44 |
+
|
45 |
+
|
46 |
+
def _power_method(transition_matrix, increase_power=True, max_iter=10000):
|
47 |
+
eigenvector = np.ones(len(transition_matrix))
|
48 |
+
|
49 |
+
if len(eigenvector) == 1:
|
50 |
+
return eigenvector
|
51 |
+
|
52 |
+
transition = transition_matrix.transpose()
|
53 |
+
|
54 |
+
for _ in range(max_iter):
|
55 |
+
eigenvector_next = np.dot(transition, eigenvector)
|
56 |
+
|
57 |
+
if np.allclose(eigenvector_next, eigenvector):
|
58 |
+
return eigenvector_next
|
59 |
+
|
60 |
+
eigenvector = eigenvector_next
|
61 |
+
|
62 |
+
if increase_power:
|
63 |
+
transition = np.dot(transition, transition)
|
64 |
+
|
65 |
+
logger.warning("Maximum number of iterations for power method exceeded without convergence!")
|
66 |
+
return eigenvector_next
|
67 |
+
|
68 |
+
|
69 |
+
def connected_nodes(matrix):
|
70 |
+
_, labels = connected_components(matrix)
|
71 |
+
|
72 |
+
groups = []
|
73 |
+
|
74 |
+
for tag in np.unique(labels):
|
75 |
+
group = np.where(labels == tag)[0]
|
76 |
+
groups.append(group)
|
77 |
+
|
78 |
+
return groups
|
79 |
+
|
80 |
+
|
81 |
+
def create_markov_matrix(weights_matrix):
|
82 |
+
n_1, n_2 = weights_matrix.shape
|
83 |
+
if n_1 != n_2:
|
84 |
+
raise ValueError('\'weights_matrix\' should be square')
|
85 |
+
|
86 |
+
row_sum = weights_matrix.sum(axis=1, keepdims=True)
|
87 |
+
|
88 |
+
# normalize probability distribution differently if we have negative transition values
|
89 |
+
if np.min(weights_matrix) <= 0:
|
90 |
+
return softmax(weights_matrix, axis=1)
|
91 |
+
|
92 |
+
return weights_matrix / row_sum
|
93 |
+
|
94 |
+
|
95 |
+
def create_markov_matrix_discrete(weights_matrix, threshold):
|
96 |
+
discrete_weights_matrix = np.zeros(weights_matrix.shape)
|
97 |
+
ixs = np.where(weights_matrix >= threshold)
|
98 |
+
discrete_weights_matrix[ixs] = 1
|
99 |
+
|
100 |
+
return create_markov_matrix(discrete_weights_matrix)
|
101 |
+
|
102 |
+
|
103 |
+
def stationary_distribution(
|
104 |
+
transition_matrix,
|
105 |
+
increase_power=True,
|
106 |
+
normalized=True,
|
107 |
+
):
|
108 |
+
n_1, n_2 = transition_matrix.shape
|
109 |
+
if n_1 != n_2:
|
110 |
+
raise ValueError('\'transition_matrix\' should be square')
|
111 |
+
|
112 |
+
distribution = np.zeros(n_1)
|
113 |
+
|
114 |
+
grouped_indices = connected_nodes(transition_matrix)
|
115 |
+
|
116 |
+
for group in grouped_indices:
|
117 |
+
t_matrix = transition_matrix[np.ix_(group, group)]
|
118 |
+
eigenvector = _power_method(t_matrix, increase_power=increase_power)
|
119 |
+
distribution[group] = eigenvector
|
120 |
+
|
121 |
+
if normalized:
|
122 |
+
distribution /= n_1
|
123 |
+
|
124 |
+
return distribution
|
app.py
CHANGED
@@ -8,6 +8,10 @@ from embed import Embed
|
|
8 |
|
9 |
import tensorflow as tf
|
10 |
|
|
|
|
|
|
|
|
|
11 |
|
12 |
class SummaryExtractor(object):
|
13 |
def __init__(self):
|
@@ -16,16 +20,51 @@ class SummaryExtractor(object):
|
|
16 |
self.tokenizer = PegasusTokenizer.from_pretrained("IDEA-CCNL/Randeng-Pegasus-523M-Summary-Chinese")
|
17 |
self.text2text_genr = Text2TextGenerationPipeline(self.model, self.tokenizer, device=self.device)
|
18 |
|
19 |
-
def extract(self, content: str
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
|
|
|
|
|
|
22 |
|
|
|
23 |
t_randeng = SummaryExtractor()
|
24 |
embedder = Embed()
|
|
|
25 |
|
26 |
|
27 |
def randeng_extract(content):
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
|
31 |
def similarity_check(inputs: list):
|
@@ -42,13 +81,13 @@ with gr.Blocks() as app:
|
|
42 |
# text_output = gr.Textbox()
|
43 |
# text_button = gr.Button("生成摘要")
|
44 |
with gr.Tab("Randeng-Pegasus-523M"):
|
45 |
-
text_input_1 = gr.Textbox()
|
46 |
-
text_output_1 = gr.Textbox()
|
47 |
text_button_1 = gr.Button("生成摘要")
|
48 |
with gr.Tab("相似度检测"):
|
49 |
with gr.Row():
|
50 |
-
text_input_query = gr.Textbox()
|
51 |
-
text_input_doc = gr.Textbox()
|
52 |
text_button_similarity = gr.Button("对比相似度")
|
53 |
text_output_similarity = gr.Textbox()
|
54 |
|
@@ -56,4 +95,7 @@ with gr.Blocks() as app:
|
|
56 |
text_button_1.click(randeng_extract, inputs=text_input_1, outputs=text_output_1)
|
57 |
text_button_similarity.click(similarity_check, inputs=[text_input_query, text_input_doc], outputs=text_output_similarity)
|
58 |
|
59 |
-
app.launch(
|
|
|
|
|
|
|
|
8 |
|
9 |
import tensorflow as tf
|
10 |
|
11 |
+
from harvesttext import HarvestText
|
12 |
+
from sentence_transformers import SentenceTransformer, util
|
13 |
+
from LexRank import degree_centrality_scores
|
14 |
+
|
15 |
|
16 |
class SummaryExtractor(object):
|
17 |
def __init__(self):
|
|
|
20 |
self.tokenizer = PegasusTokenizer.from_pretrained("IDEA-CCNL/Randeng-Pegasus-523M-Summary-Chinese")
|
21 |
self.text2text_genr = Text2TextGenerationPipeline(self.model, self.tokenizer, device=self.device)
|
22 |
|
23 |
+
def extract(self, content: str) -> str:
|
24 |
+
print(content)
|
25 |
+
return str(self.text2text_genr(content, do_sample=False, num_return_sequences=3)[0]["generated_text"])
|
26 |
+
|
27 |
+
class LexRank(object):
|
28 |
+
def __init__(self):
|
29 |
+
self.model = SentenceTransformer('paraphrase-multilingual-mpnet-base-v2')
|
30 |
+
self.ht = HarvestText()
|
31 |
+
def find_central(self, content: str):
|
32 |
+
sentences = self.ht.cut_sentences(content)
|
33 |
+
embeddings = self.model.encode(sentences, convert_to_tensor=True)
|
34 |
+
|
35 |
+
# Compute the pair-wise cosine similarities
|
36 |
+
cos_scores = util.cos_sim(embeddings, embeddings).numpy()
|
37 |
+
|
38 |
+
# Compute the centrality for each sentence
|
39 |
+
centrality_scores = degree_centrality_scores(cos_scores, threshold=None)
|
40 |
|
41 |
+
# We argsort so that the first element is the sentence with the highest score
|
42 |
+
most_central_sentence_indices = numpy.argsort(-centrality_scores)
|
43 |
+
return most_central_sentence_indices
|
44 |
|
45 |
+
# ---===--- worker instances ---===---
|
46 |
t_randeng = SummaryExtractor()
|
47 |
embedder = Embed()
|
48 |
+
lex = LexRank()
|
49 |
|
50 |
|
51 |
def randeng_extract(content):
|
52 |
+
sentences = lex.find_central(content)
|
53 |
+
|
54 |
+
num = 500
|
55 |
+
ptr = 0
|
56 |
+
for index, sentence in enumerate(sentences):
|
57 |
+
num -= len(sentence)
|
58 |
+
if num < 0 and index > 0:
|
59 |
+
ptr = index - 1
|
60 |
+
break
|
61 |
+
if num < 0 and index == 0:
|
62 |
+
ptr = index
|
63 |
+
break
|
64 |
+
print(">>>")
|
65 |
+
for ele in sentences[:ptr]:
|
66 |
+
print(ele)
|
67 |
+
return t_randeng.extract("".join(sentences[:ptr]))
|
68 |
|
69 |
|
70 |
def similarity_check(inputs: list):
|
|
|
81 |
# text_output = gr.Textbox()
|
82 |
# text_button = gr.Button("生成摘要")
|
83 |
with gr.Tab("Randeng-Pegasus-523M"):
|
84 |
+
text_input_1 = gr.Textbox(label="请输入长文本:", max_lines=1000)
|
85 |
+
text_output_1 = gr.Textbox(label="摘要文本")
|
86 |
text_button_1 = gr.Button("生成摘要")
|
87 |
with gr.Tab("相似度检测"):
|
88 |
with gr.Row():
|
89 |
+
text_input_query = gr.Textbox(label="查询文本")
|
90 |
+
text_input_doc = gr.Textbox(lines=10, label="逐行输入待比较的文本列表")
|
91 |
text_button_similarity = gr.Button("对比相似度")
|
92 |
text_output_similarity = gr.Textbox()
|
93 |
|
|
|
95 |
text_button_1.click(randeng_extract, inputs=text_input_1, outputs=text_output_1)
|
96 |
text_button_similarity.click(similarity_check, inputs=[text_input_query, text_input_doc], outputs=text_output_similarity)
|
97 |
|
98 |
+
app.launch(
|
99 |
+
# share=True,
|
100 |
+
# debug=True
|
101 |
+
)
|
article_extractor/tokenizers_pegasus.py
CHANGED
@@ -20,7 +20,7 @@ import sys
|
|
20 |
# sys.path.append("../../../../")
|
21 |
|
22 |
jieba.dt.tmp_dir = os.path.expanduser(
|
23 |
-
"
|
24 |
# jieba.enable_parallel(8)
|
25 |
jieba.initialize()
|
26 |
|
|
|
20 |
# sys.path.append("../../../../")
|
21 |
|
22 |
jieba.dt.tmp_dir = os.path.expanduser(
|
23 |
+
"tmp/")
|
24 |
# jieba.enable_parallel(8)
|
25 |
jieba.initialize()
|
26 |
|
requirements.txt
CHANGED
@@ -11,4 +11,8 @@ jieba
|
|
11 |
deepspeed
|
12 |
jieba-fast
|
13 |
protobuf
|
14 |
-
datasets
|
|
|
|
|
|
|
|
|
|
11 |
deepspeed
|
12 |
jieba-fast
|
13 |
protobuf
|
14 |
+
datasets
|
15 |
+
|
16 |
+
gradio
|
17 |
+
|
18 |
+
sentence-transformers
|