def create_tfidf_feat_idea_general(training_map, feature_map_character, constant, easy_domain_flag, args):
"""
ラベルごとの重要語を取り出す
TFIDFスコアを文書集合から算出した後，ラベル文書ごとに閾値（足切り値）を求め，閾値以下の語は素性を作らない
これで，「あるラベルに特徴的な語」を示す素性が作れた．と思う
ラベルごとで素性選択を行う．
疑似文書（ラベル）の作成→ラベル文書ごとにTFIDF計算→ラベルごとに閾値計算→
閾値を設定する→素性選択
"""
import math;
tfidf_type=args.tfidf_type;
L2_flag=True;
num_of_discarded_feat=0;
#単語スコアを保存しておくmap
word_score_map={};
#------------------------------------------------------------
print 'TFIDF(Idea-2,3,4) score calculating'
print 'L2 flag:{}'.format(L2_flag)
#------------------------------------------------------------
#訓練コーパスからラベルごとに疑似文書を作成する
#TFIDF計算の入力にはlist documents [list document [ unicode token ] ]が必要
if tfidf_type=='normal':
print 'use tfidf(normal)'
all_training_instances=[];
tmp_document_controll_map={};
for label in training_map:
#for k, v in sorted(training_map[subdata].items()):
label_document=[t for t in training_map[label]];
if label in tmp_document_controll_map:
tmp_document_controll_map[label]+=label_document;
else:
tmp_document_controll_map[label]=label_document
for k, v in sorted(tmp_document_controll_map.items()):
all_training_instances.append(tmp_document_controll_map[k]);
w_dt_maps_list=tf_idf.tf_idf_interface(all_training_instances);
#------------------------------------------------------------
elif args.tfidf_type=='nishimura':
#TFIDF計算の入力にはlist documents [list document [ list sub-document [unicode token] ] ]が必要
#IDFの計算にはlist documents[unicode token]が必要だが，これは関数内で自動で変換してくれる
print 'use tfidf(nishimura)'
all_training_instances=[];
tmp_document_controll_map={};
for subdata in training_map:
for k, v in sorted(training_map[subdata].items()):
if k in tmp_document_controll_map:
tmp_document_controll_map[k]+=training_map[subdata][k];
else:
tmp_document_controll_map[k]=training_map[subdata][k];
for k, v in sorted(tmp_document_controll_map.items()):
all_training_instances.append(tmp_document_controll_map[k]);
w_dt_maps_list=tf_idf.tf_idf_nishimura_interface(all_training_instances);
all_training_instances=[document for sub_document in all_training_instances for document in sub_document]
#------------------------------------------------------------
#全文書での重みスコアを足す
for document_index,w_dt_map in enumerate(w_dt_maps_list):
for t in w_dt_map:
if t not in word_score_map:
word_score_map[t]=w_dt_map[t];
else:
word_score_map[t]+=w_dt_map[t];
#アルファベットのリスト
alphabet_list=[chr(i) for i in range(65,65+26)];
alphabet_list.remove('I');
alphabet_list.remove('O');
alphabet_list.remove('Y');
#全文書の重みスコアにL2正則化をかけて，閾値を算出
if L2_flag==True:
#ラベルごとの閾値を保存しておくmap
#map threshold_point_map {unicode label:float threshold_point}
threshold_point_map={};
#L2正則化をかける
weight_sum=0;
for key in word_score_map:
weight=word_score_map[key];
weight_sum+=(weight)**2;
L2_norm=math.sqrt(weight_sum);
#L2正則化された単語重みを保存しておくmap
#map L2_normalized_map {map L2_normalized_map_per_label unicode label:{unicode token:float weight}}
L2_normalized_map={};
for label_index,w_dt_map in enumerate(w_dt_maps_list):
alphabet_label=alphabet_list[label_index];
L2_normalized_map_per_label={};
threshold_point_in_label=0;
for token in w_dt_map:
L2_normalized_weight=w_dt_map[token]/L2_norm;
if constant==True:
L2_normalized_map_per_label[token]=L2_normalized_weight;
else:
L2_normalized_map_per_label[token]=1;
#L2で正規化したラベル内の重みの和を求める
threshold_point_in_label+=L2_normalized_weight;
#ラベル内の閾値
threshold_point_map[alphabet_label]=threshold_point_in_label/len(w_dt_map);
if constant==True:
#ラベル内の単語重み(L2正則化済み)の保存
L2_normalized_map[alphabet_label]=L2_normalized_map_per_label;
else:
L2_normalized_map[alphabet_label]=L2_normalized_map_per_label;
for alphabet_label in L2_normalized_map:
threshold=threshold_point_map[alphabet_label];
for t in L2_normalized_map[alphabet_label]:
#足切り
if L2_normalized_map[alphabet_label][t] < threshold:
num_of_discarded_feat+=1;
else:
if constant==True:
weight_format=u'{}_{}_{}'.format(alphabet_label, t, L2_normalized_map[alphabet_label][t]);
else:
#素性ベクトルを定数倍しないので，１をたてておく.つまりBinary素性
weight_format=u'{}_{}_{}'.format('normal', t, 1);
if t not in feature_map_character:
feature_map_character[t]=[weight_format];
elif weight_format not in feature_map_character[t]:
feature_map_character[t].append(weight_format);
#閾値で足切りしない場合
else:
for t in word_score_map:
weight_format=u'{}_{}_{}'.format('normal', t, word_score_map[t]);
print 'The number of discarded features:{}'.format(num_of_discarded_feat);
return feature_map_character, L2_normalized_map;
def create_tfidf_feat_idea1(training_map, feature_map_character, args):
"""
TFIDFにて素性選択をするidea-1
文書集合全体で素性選択を行う．
疑似文書（ラベル）の作成→ラベル文書ごとにTFIDF計算→全文書でのスコアを足す→全単語のスコアを足す→
閾値を設定する→素性選択
"""
import math;
stop=args.stop;
tfidf_type=args.tfidf_type;
L2_flag=True;
num_of_discarded_feat=0;
#TFIDF空間にペルシア語コーパスの語も含めるか？
persian_flag=False;
#単語スコアを保存しておくmap
word_score_map={};
#------------------------------------------------------------
print 'TFIDF(Idea-1) score calculating'
print 'L2 flag:{} Persian flag:{}'.format(L2_flag, persian_flag);
#------------------------------------------------------------
#訓練コーパスからラベルごとに疑似文書を作成する
#TFIDF計算の入力にはlist documents [list document [ unicode token ] ]が必要
if tfidf_type=='normal':
print 'use tfidf(normal)'
all_training_instances=[];
tmp_document_controll_map={};
for subdata in training_map:
for k, v in sorted(training_map[subdata].items()):
label_document=[t for doc in v for t in doc];
if k in tmp_document_controll_map:
tmp_document_controll_map[k]+=label_document;
else:
tmp_document_controll_map[k]=label_document
for k, v in sorted(tmp_document_controll_map.items()):
all_training_instances.append(tmp_document_controll_map[k]);
w_dt_maps_list=tf_idf.tf_idf_interface(all_training_instances);
#------------------------------------------------------------
elif args.tfidf_type=='nishimura':
#TFIDF計算の入力にはlist documents [list document [ list sub-document [unicode token] ] ]が必要
#IDFの計算にはlist documents[unicode token]が必要だが，これは関数内で自動で変換してくれる
print 'use tfidf(nishimura)'
all_training_instances=[];
tmp_document_controll_map={};
for subdata in training_map:
for k, v in sorted(training_map[subdata].items()):
if k in tmp_document_controll_map:
tmp_document_controll_map[k]+=training_map[subdata][k];
else:
tmp_document_controll_map[k]=training_map[subdata][k];
for k, v in sorted(tmp_document_controll_map.items()):
all_training_instances.append(tmp_document_controll_map[k]);
w_dt_maps_list=tf_idf.tf_idf_nishimura_interface(all_training_instances);
all_training_instances=[document for sub_document in all_training_instances for document in sub_document]
#------------------------------------------------------------
#全文書での重みスコアを足す
for document_index,w_dt_map in enumerate(w_dt_maps_list):
for t in w_dt_map:
if t not in word_score_map:
word_score_map[t]=w_dt_map[t];
else:
word_score_map[t]+=w_dt_map[t];
#全文書の重みスコアにL2正則化をかけて，閾値を算出
if L2_flag==True:
#L2正則化をかける
weight_sum=0;
for key in word_score_map:
weight=word_score_map[key];
weight_sum+=(weight)**2;
L2_norm=math.sqrt(weight_sum);
L2_normalized_map={};
L2_weightsum=0;
for key in word_score_map:
normalized_score=word_score_map[key]/L2_norm;
L2_normalized_map[key]=normalized_score;
#L2で正規化した重みの和を求める
L2_weightsum+=normalized_score;
#足切りスコアの算出
L2_average=L2_weightsum/len(L2_normalized_map);
for doc in all_training_instances:
for t in doc:
if t in L2_normalized_map:
if L2_normalized_map[t] < L2_average:
#足切り
num_of_discarded_feat+=1;
else:
weight_format=u'{}_{}_{}'.format('normal', t, L2_normalized_map[t]);
if t not in feature_map_character:
feature_map_character[t]=[weight_format];
elif weight_format not in feature_map_character[t]:
feature_map_character[t].append(weight_format);
#閾値で足切りしない場合
else:
for t in word_score_map:
weight_format=u'{}_{}_{}'.format('normal', t, word_score_map[t]);
print 'The number of discarded features :{}'.format(num_of_discarded_feat);
#コード書き換えに伴い変数名が変わったので，一時的な措置
tfidf_score_map=word_score_map;
return feature_map_character, tfidf_score_map;
