SpursGoZmy
commented
7 months ago 因为相关代码之前有产品里用过，我后面尽快完整地整理出来。
基于中文单元格内容构建手工特征的类如下，
import numpy as np
import paddle
import paddle.nn as nn
import pgl
from paddlenlp import Taskflow
from collections import defaultdict
class Manual_CTC_Feats(object):
def __init__(self) -> None:
super(Manual_CTC_Feats,self).__init__()
"""
This class is used to create manual cell feats from the paper "A Machine Learning Approach for Layout Inference in Spreadsheets"
We adopt 24 feats from the paper's selected feats(section 5.1),
note that we do not use some feats like cell style and font because we want to process general tables rather than particular spreadsheet tables,
and we adapt some feats for chinese language like punctuations
"""
self.module_name='Manual-CTC-Feats-Module'
# word_segmentation and pos_tagging tools from paddlenlp Taskflow
self.seg_model = Taskflow("word_segmentation",batch_size=4)
self.pos_model=Taskflow("pos_tagging",batch_size=4)
self.punctuation_list=['，','。','-','——','？','?','!','！','.','“','”','；']
self.special_token_list=['#','$','￥','%','&','*','~','/','<','>']
#we select some chinese words expressing the 'total' meaning in the OpenHowNet toolkit.
self.words_like_total=['共计','合计','总数','总量','全部','总共有','总额度','总额','总值',
'总和','共有','总共','累计','累计为','总合','合达','total','sum','all']
self.content_feature_num=15
self.spatial_feature_num=9
def build_content_features(self,cell_values):
cell_num=len(cell_values)
content_features=np.zeros((cell_num,self.content_feature_num)) # dtype = 'float64'
processed_cell_values=["Empty_cell" if cell_text=="" or str.isspace(cell_text) else cell_text for cell_text in cell_values] ##处理空cell_values
assert len(processed_cell_values)==len(cell_values)
# segmentation results of cell_values (a nested list)
segmented_cell_values=self.seg_model(processed_cell_values)
# pos tagging results of cell_values (a nested list)
cell_tag_results=self.pos_model(processed_cell_values)
if len(segmented_cell_values)!=len(cell_values):
print("len(segmented_cell_values): ",len(segmented_cell_values))
print("len(cell_values): ",len(cell_values))
print(cell_values)
print(segmented_cell_values)
#considered content feat:
for cell_id,cell_text in enumerate(cell_values):
# LENGTH#
text_len=len(cell_text)
if text_len==0:
continue
segmented_cell_text=segmented_cell_values[cell_id]
pos_tag_results=cell_tag_results[cell_id]
# NUM OF TOKENS#
if cell_text=="":
tokens_num=0
else:
tokens_num=len(segmented_cell_text)
# LEADING SPACES#
leading_spaces_num=0
for char in cell_text:
if char==' ':
leading_spaces_num+=1
else:
break
# IS NUMERIC?
if str.isnumeric(cell_text):
is_numeric=1
else:
is_numeric=0
#STARTS WITH NUMBER?
if str.isnumeric(cell_text[0]):
starts_with_number=1
else:
starts_with_number=0
# STARTS WITH SPECIAL?
if cell_text[0] in self.special_token_list:
starts_with_special=1
else:
starts_with_special=0
#IS CAPITALIZED?
if str.istitle(cell_text):
is_capitalized=1
else:
is_capitalized=0
#IS UPPER CASE?
if str.isupper(cell_text):
is_upper=1
else:
is_upper=0
#IS ALPHABETIC?
if str.isalpha(cell_text):
is_alpha=1
else:
is_alpha=0
#CONTAINS SPECIAL CHARS?
contain_special_char=0
for char in cell_text:
if char in self.special_token_list:
contain_special_char=1
break
#CONTAINS PUNCTUATIONS?
contain_punctuation=0
for char in cell_text:
if char in self.punctuation_list:
contain_punctuation=1
break
#CONTAINS COLON?
if cell_text.find(':')>=0:
contain_colon=1
elif cell_text.find('：')>=0:
contain_colon=1
else:
contain_colon=0
#WORDS LIKE TOTAL?
words_like_total=0
for word in segmented_cell_text:
if word in self.words_like_total:
words_like_total=1
break
#WORDS LIKE TABLE?
words_like_table=0
for word in segmented_cell_text:
if word in ['表','表格']:
words_like_table=1
break
#IN YEAR RANGE?
in_year_range=0
for tag_result in pos_tag_results:
if tag_result[1] in ['t','TIME']:
in_year_range=1
break
#build feat vectors
feat_vector=[]
feat_vector.append(text_len)
feat_vector.append(tokens_num)
feat_vector.append(leading_spaces_num)
feat_vector.append(is_numeric)
feat_vector.append(starts_with_number)
feat_vector.append(starts_with_special)
feat_vector.append(is_capitalized)
feat_vector.append(is_upper)
feat_vector.append(is_alpha)
feat_vector.append(contain_special_char)
feat_vector.append(contain_punctuation)
feat_vector.append(contain_colon)
feat_vector.append(words_like_total)
feat_vector.append(words_like_table)
feat_vector.append(in_year_range)
feat_vector=np.array(feat_vector)
content_features[cell_id]=feat_vector
return content_features
def build_spatial_feats(self,layout,cell_values):
cell_num=len(cell_values)
spatial_features=np.zeros((cell_num,self.spatial_feature_num)) # dtype = 'float64'
# considered spatial feat: row_id, col_id,has_1_2_3_4_neighbour,more than 5 neighbor, is_merged_cell,num_of_cells,
cell_id_to_row_and_col_id={}
cell_id_to_neighbor_set=defaultdict(set)
cell_id_to_cell_num=defaultdict(int)
row_num=len(layout)
col_num=len(layout[0])
for row_id in range(row_num):
for col_id in range(col_num):
cell_id=layout[row_id][col_id]
if cell_id not in cell_id_to_row_and_col_id:
cell_id_to_row_and_col_id[cell_id]=(row_id,col_id)
cell_id_to_cell_num[cell_id]+=1
#top neighbour
if row_id-1>=0:
top_neighbour=layout[row_id-1][col_id]
cell_id_to_neighbor_set[cell_id].add(top_neighbour)
#left neighbour
if col_id-1>=0:
left_neighbour=layout[row_id][col_id-1]
cell_id_to_neighbor_set[cell_id].add(left_neighbour)
#right neighbour
if col_id+1<col_num:
right_neighbour=layout[row_id][col_id+1]
cell_id_to_neighbor_set[cell_id].add(right_neighbour)
# down neighbour
if row_id+1<row_num:
down_neighbour=layout[row_id+1][col_id]
cell_id_to_neighbor_set[cell_id].add(down_neighbour)
for cell_id in range(cell_num):
row_id=cell_id_to_row_and_col_id[cell_id][0]
col_id=cell_id_to_row_and_col_id[cell_id][1]
neighbour_num=len(cell_id_to_neighbor_set[cell_id])
neighbour_num_feats=[0,0,0,0,0]
if neighbour_num<=4 and neighbour_num>0:
neighbour_num_feats[neighbour_num-1]=1
else:
neighbour_num_feats[4]=1
num_cells=cell_id_to_cell_num[cell_id]
if num_cells>1:
is_merged_cell=1
else:
is_merged_cell=0
feat_vector=[]
feat_vector.append(row_id)
feat_vector.append(col_id)
feat_vector.extend(neighbour_num_feats)
feat_vector.append(num_cells)
feat_vector.append(is_merged_cell)
feat_vector=np.array(feat_vector)
spatial_features[cell_id]=feat_vector
return spatial_features
def build_features(self,layout,cell_values):
content_feats=self.build_content_features(cell_values)
spatial_feats=self.build_spatial_feats(layout,cell_values)
cell_feats=np.concatenate((content_feats,spatial_feats),axis=1)
return cell_feats
可以通过类似如下的代码获取单元格的手工特征：
Manual_CTC_Feats_Module=Manual_CTC_Feats()
layout = cell_ID_matrix # read 'cell_ID_matrix' of one table
cell_values = chinese_cell_value_list # read 'chinese_cell_value_list' of one table
cell_feats=Manual_CTC_Feats_Module.build_features(layout,cell_values) # np.array, [cell_num,24]
Czy-family
WeixuanXiong
是基于中文的单元格内容生成一些手工特征，参考了A Machine Learning Approach for Layout Inference in Spreadsheets这篇论文采用的一些手工特征。