Skip to content

Latest commit

 

History

History
170 lines (139 loc) · 5.53 KB

baseline.md

File metadata and controls

170 lines (139 loc) · 5.53 KB

Baseline Implementation

CRF (Conditional Random Field)

CRF 模型需要把資料 (.txt) 先轉成 CRF 能接受的格式 (.data):

Field Explain
O 代表無關 NER 的單字
B-OOO 代表某個類別 OOO 的起點
I-OOO 代表某個類別 OOO 的文字 (可能是文字途中或結尾)

程式中的 CRFFormatData() 會得到一個 .data 檔案，裡面包含所有所有 content 每個字對應的 IOB type

醫 O
師 O
： O
你 O
有 O
...
...
6 B-med_exam
8 I-med_exam
...
...

Data Preprocessing

將上面的 data 檔案透過 Dataset() 分成訓練和測試集 (train / test):

Object Format Desc
data_list (content, sentence, tuple(word, CRF-type)) 將每個 content 的 data 抓出來，每個 tuple 是每個字的 CRF 資料
train_data_list (content, sentence, tuple(word, CRF-type)) 從 data_list 拆出的 training set
test_data_list (content, sentence, tuple(word, CRF-type)) 從 data_list 拆出的 testing set
train_data_article_id_list (content_id) training set 裡面的對應 content 編號
test_data_article_id_list (content_id) testing set 裡面的對應 content 編號

data_list

# shape=(content_len, sentence_len, 1)
[[
  ('他', 'O'),
  ('說', 'O'),
  ('那', 'O'),
  ...
],
...
[
  ('兩', 'B-money'),
  ('百', 'I-money'),
  ('多', 'I-money'),
]]

Embed data_list

將每個文字 (tuple[0]) 利用 Word2Vector 轉為 512 維度的 embedding 表示:

train_embed_list = Word2Vector(traindata_list, word_vecs)
[[
  [-5.217900e-02 -1.332038e+00  8.813320e-01 -7.293170e-01 -9.045000e-01 ...] # 他
  [-2.194207e+00 -1.011986e+00  8.231450e-01 -1.624110e-01 -5.918380e-01 ...] # 說
  [-3.255907e+00 -2.022559e+00  8.332450e-02 -2.634220e-01 -6.559380e-01 ...] # 那
],
...
[
  ...
]]

X, y

  • Feature() 會將每個 512 維度的文字 embedding 轉為字典表示
  • Preprocess() 將每個 data_list 的 label 提取出來變成 list
x_train = Feature(train_embed_list)

# an example of a word token 
[
  -5.217900e-02 -1.332038e+00  8.813320e-01 -7.293170e-01 -9.045000e-01 ... 1.895004e+00
]
=>
[
  'dim_1': -0.052179, 'dim_2': -1.332038, 'dim_3': 0.881332, 'dim_4': -0.729317, 'dim_5': -0.9045, ..., 'dim_512': 1.895004
]

y_train = Preprocess(traindata_list)
['O', 'O', 'O', 'B-location', 'I-location', 'O', 'O', 'O', 'O', 'B-location', 'I-location', ...]

以下是使用 sample 資料集得到的範例資料格式大小:

Properties Shape Explain
x_train (17, 1759, 512) (content, max_len_sentence, embed)
y_train (17, 1759, 1) (content, max_len_sentence, label)
x_test (9, 2829, 512) (content, max_len_sentence, embed)
y_test (9, 2829, 1) (content, max_len_sentence, label)

Train

crf = sklearn_crfsuite.CRF(
    algorithm='lbfgs',
    c1=0.1,
    c2=0.1,
    max_iterations=100,
    all_possible_transitions=True
)
    
crf.fit(x_train, y_train)
y_pred = crf.predict(x_test)
# return predict class
# e.g., B-location

y_pred_mar = crf.predict_marginals(x_test)
# return predict distribution
# e.g., 
# {
#  'O': 0.7364770082815129, 
#  'B-location': 2.253738319713659e-07, 
#  'I-location': 0.2629652437709382, 
#  'B-time': 1.4817077986266331e-06, 
#  'I-time': 8.182176401083988e-05, 
#  'B-med_exam': 4.638799236990415e-06, 
#  'I-med_exam': 0.0003360698810705325, 
#  'B-name': 4.943808713340432e-07, 
#  'I-name': 1.1926458997191943e-06,
#  'B-money': 1.6409408539571999e-06, 
#  'I-money': 0.00013018245396879223
# }

labels = list(crf.classes_)
labels.remove('O')
f1score = metrics.flat_f1_score(
        y_test, y_pred, average='weighted', labels=labels)

Results

Properties Format Explain
y_pred (9, 2829, 1) (content, max_len_sentence, label)
y_pred_mar (9, 2829, 11) (content, max_len_sentence, label_distribution)

Upload

最終我們要輸出的格式要依照下列表格為準，我們要自己從 y_pred 和 data_list 去抓出以下格式。

article_id start_position end_position entity_text entity_type
8 52 54 前天 time
8 68 70 昨天 time
8 189 193 二十分鐘 time