l = line.split("\t")
target_id = l[0]
source_id = l[1]
intent = l[2]
left_citated_text = l[3]
right_citated_text = l[4]
left_citation_tokenized = tokenizer.tokenize(
left_citated_text)[-250:]
right_citation_tokenized = tokenizer.tokenize(
right_citated_text)[:250]
input_tokens = tokenizer.convert_tokens_to_ids(
left_citation_tokenized) + [
tokenizer.sep_token_id
] + tokenizer.convert_tokens_to_ids(right_citation_tokenized)
position_citation_mark = len(left_citation_tokenized)
tokens_tensor = torch.tensor([input_tokens])
outputs = model(tokens_tensor)
emb = np.array(outputs[0][0][position_citation_mark].cpu())
if intent not in intentdict:
intentdict[intent] = intentn
intentn += 1
X.append(emb)
y.append(intentdict[intent])
return X, y
if __name__ == "__main__":
ent_vocab = build_ent_vocab(
"/home/ohagi_masaya/TransBasedCitEmb/dataset/AASC/train.csv")
load_data_SVM_with_context(ent_vocab)
def load_AASC_graph_data(args):
def extract_by_frequency(path_train, path_test, frequency):
dftrain = pd.read_csv(path_train, quotechar="'")
dftest = pd.read_csv(path_test, quotechar="'")
source_cut_train = dftrain[[
'target_id', 'source_id'
]].drop_duplicates(subset=['target_id', 'source_id'])
source_cut_test = dftest[[
'target_id', 'source_id'
]].drop_duplicates(subset=['target_id', 'source_id'])
ftrain_fre = open(
path_train[:-4] + "_frequency" + str(frequency) + ".csv", "w")
ftest_fre = open(
path_test[:-4] + "_frequency" + str(frequency) + ".csv", "w")
wtrain = csv.writer(ftrain_fre, quotechar="'")
wtest = csv.writer(ftest_fre, quotechar="'")
wtrain.writerow([
"target_id", "left_citated_text", "right_citated_text", "source_id"
])
wtest.writerow([
"target_id", "left_citated_text", "right_citated_text", "source_id"
])
source_train_keys = source_cut_train.source_id.value_counts().keys()
source_test_keys = source_cut_test.source_id.value_counts().keys()
dic1 = {}
train_counts = source_cut_train.source_id.value_counts()
test_counts = source_cut_test.source_id.value_counts()
for key in source_train_keys:
dic1[key] = train_counts[key]
for key in source_test_keys:
if key in dic1:
dic1[key] += test_counts[key]
else:
dic1[key] = test_counts[key]
frequencylist = []
for key in dic1:
if dic1[key] >= frequency:
frequencylist.append(key)
dftrain = dftrain.loc[dftrain["source_id"].isin(frequencylist)]
dftest = dftest.loc[dftest["source_id"].isin(frequencylist)]
for target_id, left_citated_text, right_citated_text, source_id in zip(
dftrain["target_id"], dftrain["left_citated_text"],
dftrain["right_citated_text"], dftrain["source_id"]):
wtrain.writerow(
[target_id, left_citated_text, right_citated_text, source_id])
ftrain_fre.close()
for target_id, left_citated_text, right_citated_text, source_id in zip(
dftest["target_id"], dftest["left_citated_text"],
dftest["right_citated_text"], dftest["source_id"]):
wtest.writerow(
[target_id, left_citated_text, right_citated_text, source_id])
ftest_fre.close()
entitylist = list(
set(
list(dftrain["source_id"].values) +
list(dftrain["target_id"].values) +
list(dftest["source_id"].values) +
list(dftest["target_id"].values)))
entitylist.sort()
ent_vocab = {"UNKNOWN": 0, "MASK": 1}
for i, entity in enumerate(entitylist):
ent_vocab[entity] = i + 2
return path_train[:-4] + "_frequency" + str(
frequency) + ".csv", path_test[:-4] + "_frequency" + str(
frequency) + ".csv", ent_vocab
path = settings.citation_recommendation_dir
path_train = os.path.join(path, "train.csv")
path_test = os.path.join(path, "test.csv")
ent_vocab = build_ent_vocab(path_train)
path_train_frequency5, path_test_frequency5, ent_vocab_frequency5 = extract_by_frequency(
path_train, path_test, args.frequency)
#randomでMASKするように一旦変更
datasetdict = {
"tail": AASCDataSet,
"random": AASCDataSetRANDOM,
"both": AASCDataSetBOTH
}
cur_dataset = datasetdict[args.mask_type]
if args.train_data == "full":
dataset_train = cur_dataset(path_train,
ent_vocab=ent_vocab,
WINDOW_SIZE=args.WINDOW_SIZE,
MAX_LEN=args.MAX_LEN,
pretrained_model=args.pretrained_model,
mode="train")
else:
dataset_train = cur_dataset(path_train_frequency5,
ent_vocab=ent_vocab,
WINDOW_SIZE=args.WINDOW_SIZE,
MAX_LEN=args.MAX_LEN,
pretrained_model=args.pretrained_model,
mode="train")
if args.test_data == "full":
dataset_test = cur_dataset(path_test,
ent_vocab=ent_vocab,
WINDOW_SIZE=args.WINDOW_SIZE,
MAX_LEN=args.MAX_LEN,
pretrained_model=args.pretrained_model,
mode="test")
else:
dataset_test = cur_dataset(path_test_frequency5,
ent_vocab=ent_vocab,
WINDOW_SIZE=args.WINDOW_SIZE,
MAX_LEN=args.MAX_LEN,
pretrained_model=args.pretrained_model,
mode="test")
print("----loading data done----")
return dataset_train, dataset_test, ent_vocab
def load_AASC_graph_data(args):
def extract_by_frequency(path_train, path_test, frequency):
dftrain = pd.read_csv(path_train, quotechar="'")
dftest = pd.read_csv(path_test, quotechar="'")
source_cut_train = dftrain[[
'target_id', 'source_id'
]].drop_duplicates(subset=['target_id', 'source_id'])
source_cut_test = dftest[[
'target_id', 'source_id'
]].drop_duplicates(subset=['target_id', 'source_id'])
ftrain_fre = open(
path_train[:-4] + "_frequency" + str(frequency) + ".csv", "w")
ftest_fre = open(
path_test[:-4] + "_frequency" + str(frequency) + ".csv", "w")
wtrain = csv.writer(ftrain_fre, quotechar="'")
wtest = csv.writer(ftest_fre, quotechar="'")
wtrain.writerow([
"target_id", "left_citated_text", "right_citated_text", "source_id"
])
wtest.writerow([
"target_id", "left_citated_text", "right_citated_text", "source_id"
])
source_train_keys = source_cut_train.source_id.value_counts().keys()
source_test_keys = source_cut_test.source_id.value_counts().keys()
dic1 = {}
train_counts = source_cut_train.source_id.value_counts()
test_counts = source_cut_test.source_id.value_counts()
for key in source_train_keys:
dic1[key] = train_counts[key]
for key in source_test_keys:
if key in dic1:
dic1[key] += test_counts[key]
else:
dic1[key] = test_counts[key]
frequencylist = []
for key in dic1:
if dic1[key] >= frequency:
frequencylist.append(key)
dftrain = dftrain.loc[dftrain["source_id"].isin(frequencylist)]
dftest = dftest.loc[dftest["source_id"].isin(frequencylist)]
for target_id, left_citated_text, right_citated_text, source_id in zip(
dftrain["target_id"], dftrain["left_citated_text"],
dftrain["right_citated_text"], dftrain["source_id"]):
wtrain.writerow(
[target_id, left_citated_text, right_citated_text, source_id])
ftrain_fre.close()
for target_id, left_citated_text, right_citated_text, source_id in zip(
dftest["target_id"], dftest["left_citated_text"],
dftest["right_citated_text"], dftest["source_id"]):
wtest.writerow(
[target_id, left_citated_text, right_citated_text, source_id])
ftest_fre.close()
entitylist = list(
set(
list(dftrain["source_id"].values) +
list(dftrain["target_id"].values) +
list(dftest["source_id"].values) +
list(dftest["target_id"].values)))
entitylist.sort()
entvocab = {"UNKNOWN": 0, "MASK": 1}
for i, entity in enumerate(entitylist):
entvocab[entity] = i + 2
return path_train[:-4] + "_frequency" + str(
frequency) + ".csv", path_test[:-4] + "_frequency" + str(
frequency) + ".csv", entvocab
path = settings.citation_recommendation_dir
path_train = os.path.join(path, "train.csv")
path_test = os.path.join(path, "test.csv")
path_emb_train = os.path.join(path, "scibert_AASCtrain.npy")
path_emb_test = os.path.join(path, "scibert_AASCtest.npy")
entvocab = build_ent_vocab(path_train)
matrix_train = makecitationmatrix_AASC(path_train, path_emb_train,
entvocab)
matrix_test = makecitationmatrix_AASC(path_test, path_emb_test, entvocab)
path_train_frequency5, path_test_frequency5, entvocab_frequency5 = extract_by_frequency(
path_train, path_test, args.frequency)
if args.train_data == "full":
dataset_train = AASCDataSet(path_train,
ent_vocab=entvocab,
MAX_LEN=args.MAX_LEN,
matrix=matrix_train)
else:
dataset_train = AASCDataSet(path_train_frequency5,
ent_vocab=entvocab,
MAX_LEN=args.MAX_LEN,
matrix=matrix_train)
if args.test_data == "full":
dataset_test = AASCDataSet(path_test,
ent_vocab=entvocab,
MAX_LEN=args.MAX_LEN,
matrix=matrix_test,
mode="test")
else:
dataset_test = AASCDataSet(path_test_frequency5,
ent_vocab=entvocab,
MAX_LEN=args.MAX_LEN,
matrix=matrix_test,
mode="test")
#return dataset_train,dataset_test,entvocab
return dataset_train, dataset_test, entvocab, matrix_train, matrix_test
