def main(dataset, window):
model = gensim.models.KeyedVectors.load_word2vec_format(
'./data/embedding/vec/externel_vec/wiki.en.vec', binary=False)
# 注：因为gensim版本更新的问题，如果下面这个load有问题，可以使用新的接口：model = gensim.models.word2vec.Word2Vec.load(MODEL_PATH)
# model = gensim.models.Word2Vec.load_word2vec_format("wiki.en.text.vector", binary=False)
# model.similarity("woman", "girl")
# 计算生成经典特征
data_dir = os.path.join('./data/embedding/', dataset)
file_names = read_file(os.path.join(data_dir, 'abstract_list')).split(',')
out_dir = os.path.join(data_dir, 'wiki_sim')
for file_name in file_names:
print(file_name)
filtered_text = filter_text(
read_file(os.path.join(data_dir, 'abstracts', file_name)))
edges = get_edges(filtered_text, window=window)
edge_sim = {}
for edge in edges:
word1 = edge[0]
word2 = edge[1]
try:
sim = model.similarity(word1, word2)
except:
sim = 0
e = tuple(
sorted([normalized_token(word1),
normalized_token(word2)]))
edge_sim[e] = [sim]
edgefeatures2file(os.path.join(data_dir, 'wiki_sim', file_name),
edge_sim)
print('.......wiki_sim_DONE........')
def position_sum(text, nodes):
"""
Return a dict, key is node,
value is the sum of the Reciprocal sum
of all the node positions in text.
:param text: text with no tags
:param nodes: list of nodes in word graph, stemmed
"""
def all_pos(obj, alist):
pos = []
while True:
try:
p = alist.index(obj)
pos.append(p + 1)
alist[p] = None
except:
return pos
text = text.lower()
words = [normalized_token(w) for w in word_tokenize(text) if is_word(w)]
pos_sum = {}
for n in nodes:
pos = all_pos(n, words)
weight = sum([1 / p for p in pos])
pos_sum[n] = weight
return pos_sum
def create_B(node_list, gold):
keyphrases = list(normalized_token(word) for word in gold.split())
n = len(node_list)
B = [0] * n
for g in keyphrases:
if g not in node_list:
keyphrases.pop(keyphrases.index(g))
for keyphrase in keyphrases:
try:
prefer = node_list.index(keyphrase)
except:
continue
b = [0] * n
b[prefer] = 1
B = []
for node in node_list:
if node not in keyphrases:
neg = node_list.index(node)
b[neg] = -1
c = b[:]
B.append(c)
b[neg] = 0
if B == []:
B = [0] * n
return np.matrix(B)
def read_pr(fpath, vocabulary, damping):
pr = {}
with open(fpath, encoding='utf8') as f:
f_csv = csv.DictReader(f)
for row in f_csv:
token = normalized_token(vocabulary[row['node_id']])
score = float(row[str(damping)])
pr[token] = score
return pr
def get_term_freq(text):
"""
Return a dict, key is stemmed word,
value is frequency in text.
:param text: text with no tags
"""
text = text.lower()
words = [normalized_token(w) for w in word_tokenize(text) if is_word(w)]
tf = {}
for w in words:
tf[w] = words.count(w)
return tf
def evaluate(dataset):
"""
Evaluate ranking result.
:param dataset: name of dataset
:param pr: dict, key is stemmed word, value is score
"""
method_name = 'pagerank_zf'
dataset = dataset.upper()
abstract_dir = os.path.join('./data', dataset, 'abstracts')
gold_dir = os.path.join('./data', dataset, 'gold')
extracted = os.path.join('./result', dataset, 'extracted_zf')
pr_type = 'a1b1' #alfa=1beta=1
pr_dir = os.path.join('./data', dataset, 'rank_zf', pr_type)
vocabulary_path = os.path.join('./data', dataset, 'rank_zf',
'vocabulary') #对应
damping = 0.85 #0.2 0.5 0.8 0.85
with_tag = True
topn = 4
window = 2
ngrams = 2
weight2 = 0.6
weight3 = 0.3
names = [
name for name in os.listdir(pr_dir)
if os.path.isfile(os.path.join(pr_dir, name))
]
vocabulary = id2word(vocabulary_path)
count = 0
gold_count = 0
extract_count = 0
mrr = 0
prcs_micro = 0
recall_micro = 0
for name in names:
pr = read_pr(os.path.join(pr_dir, name), vocabulary, damping)
doc_path = os.path.join(abstract_dir, name)
text = read_file(doc_path)
text_candidates = filter_text(text, with_tag=with_tag)
edge_freq = get_edge_freq(text_candidates, window=window)
edges = dict2list(edge_freq)
graph = build_graph(edges)
keyphrases = get_phrases(pr,
graph,
doc_path,
ng=ngrams,
pl2=weight2,
pl3=weight3,
with_tag=with_tag)
top_phrases = []
for phrase in keyphrases:
if phrase[0] not in str(top_phrases):
top_phrases.append(phrase[0])
if len(top_phrases) == topn:
break
if not os.path.exists(extracted):
os.makedirs(extracted)
with open(os.path.join(extracted, name), encoding='utf-8',
mode='w') as file:
file.write('\n'.join(top_phrases))
standard = read_file(os.path.join(gold_dir, name)).split('\n')
if standard[-1] == '':
standard = standard[:-1]
# 根据phrases是否取词干决定
standard = list(' '.join(list(normalized_token(w) for w in g.split()))
for g in standard)
count_micro = 0
position = []
for phrase in top_phrases:
if phrase in standard:
count += 1
count_micro += 1
position.append(top_phrases.index(phrase))
if position != []:
mrr += 1 / (position[0] + 1)
gold_count += len(standard)
extract_count += len(top_phrases)
prcs_micro += count_micro / len(top_phrases)
recall_micro += count_micro / len(standard)
prcs = count / extract_count
recall = count / gold_count
f1 = 2 * prcs * recall / (prcs + recall)
mrr /= len(names)
prcs_micro /= len(names)
recall_micro /= len(names)
f1_micro = 2 * prcs_micro * recall_micro / (prcs_micro + recall_micro)
print(dataset, method_name, count, prcs, recall, f1, mrr)
eval_result = method_name + pr_type + str(damping) + '@' + str(topn) + ',' + dataset + ',' + str(prcs) + ',' \
+ str(recall) + ',' + str(f1) + ',' + str(mrr) + ',' + str(prcs_micro) \
+ ',' + str(recall_micro) + ',' + str(f1_micro) + ',\n'
with open(os.path.join('./result', method_name + '.csv'),
mode='a',
encoding='utf8') as file:
file.write(eval_result)
def evaluate_extraction(dataset,
method_name,
ngrams=2,
damping=0.85,
omega=None,
phi=None,
alter_topn=None,
alter_edge=None,
alter_node=None):
"""
评价实验结果
omega,phi, [0]代表不适用任何特征，权重设置为1。None为所有特征的简单加和。[-1]只用最后一个特征。
"""
if dataset == 'KDD':
abstr_dir = './data/embedding/KDD/abstracts/'
out_dir = './result/embedding/'
gold_dir = './data/embedding/KDD/gold/'
edge_dir = './data/embedding/KDD/edge_features/'
node_dir = './data/embedding/KDD/node_features/'
file_names = read_file('./data/embedding/KDD/abstract_list').split(',')
topn = 4
elif dataset == 'WWW':
abstr_dir = './data/embedding/WWW/abstracts/'
out_dir = './result/embedding/'
gold_dir = './data/embedding/WWW/gold/'
edge_dir = './data/embedding/WWW/edge_features/'
node_dir = './data/embedding/WWW/node_features/'
file_names = read_file('./data/embedding/WWW/abstract_list').split(',')
topn = 5
else:
print('wrong dataset name')
if not os.path.exists(out_dir):
os.makedirs(out_dir)
if alter_edge:
edge_dir = alter_edge
if alter_node:
node_dir = alter_node
if alter_topn:
topn = alter_topn
count = 0
gold_count = 0
extract_count = 0
mrr = 0
prcs_micro = 0
recall_micro = 0
for file_name in file_names:
# print(file_name)
pr, graph = wpr(edge_dir + file_name,
node_dir + file_name,
omega=omega,
phi=phi,
d=damping)
gold = read_file(gold_dir + file_name)
pl2 = 0.62
pl3 = 0.3
if dataset == "WWW":
pl2 = 0.55
file = os.path.join(abstr_dir, file_name)
keyphrases = get_phrases(pr, graph, file, ng=ngrams, pl2=pl2, pl3=pl3)
top_phrases = []
for phrase in keyphrases:
if phrase[0] not in str(top_phrases):
top_phrases.append(phrase[0])
if len(top_phrases) == topn:
break
golds = gold.split('\n')
if golds[-1] == '':
golds = golds[:-1]
golds = list(' '.join(list(normalized_token(w) for w in g.split()))
for g in golds)
count_micro = 0
position = []
for phrase in top_phrases:
if phrase in golds:
count += 1
count_micro += 1
position.append(top_phrases.index(phrase))
if position != []:
mrr += 1 / (position[0] + 1)
gold_count += len(golds)
extract_count += len(top_phrases)
if len(top_phrases) != 0:
prcs_micro += count_micro / len(top_phrases)
recall_micro += count_micro / len(golds)
# 记录每个文档关键词提取的详细结果
# prcs_single = count_micro / len(top_phrases)
# recall_single = count_micro / len(golds)
# output_single = str(file_name) + ',' + str(prcs_single) + ',' + str(recall_single) + ','\
# + ','.join(phrase for phrase in top_phrases) + '\n'
# with open(out_dir + dataset + 'DETAILS.csv', mode='a', encoding='utf8') as f:
# f.write(output_single)
prcs = count / extract_count
recall = count / gold_count
f1 = 2 * prcs * recall / (prcs + recall)
mrr /= len(file_names)
prcs_micro /= len(file_names)
recall_micro /= len(file_names)
f1_micro = 2 * prcs_micro * recall_micro / (prcs_micro + recall_micro)
print(prcs, recall, f1, mrr)
tofile_result = method_name + ',' + str(prcs) + ',' + str(recall) + ',' + str(f1) + ',' + str(mrr) + ',' \
+ str(prcs_micro) + ',' + str(recall_micro) + ',' + str(f1_micro) + ',\n'
with open(out_dir + dataset + '_RESULTS.csv', mode='a',
encoding='utf8') as f:
f.write(tofile_result)