def load_sentence(sent_dict):
tokens = load_tokens(sent_dict["tokens"])
pub_time = utils.strip_to_date(arrow.get(sent_dict["pub_time"]))
time = Sentence.get_time(tokens)
time_level = None
if time:
time = arrow.get(time)
time_format = Sentence.get_time_format(tokens)
time_level = None
if "d" in time_format:
time = datetime.datetime(time.year, time.month, time.day)
time_level = "d"
elif ("m" in time_format) or ("y" in time_format):
if "m" in time_format:
start, end = time.span("month")
time_level = "m"
else:
start, end = time.span("year")
time_level = "y"
start = datetime.datetime(start.year, start.month, start.day)
end = datetime.datetime(end.year, end.month, end.day)
time = (start, end)
return Sentence(sent_dict["raw"], tokens, pub_time, time, time_level)
def load_sentence(sent_dict):
tokens = load_tokens(sent_dict['tokens'])
pub_time = utils.strip_to_date(arrow.get(sent_dict['pub_time']))
time = Sentence.get_time(tokens)
time_level = None
if time:
time = arrow.get(time)
time_format = Sentence.get_time_format(tokens)
time_level = None
if 'd' in time_format:
time = datetime.datetime(time.year, time.month, time.day)
time_level = 'd'
elif ('m' in time_format) or ('y' in time_format):
if 'm' in time_format:
start, end = time.span('month')
time_level = 'm'
else:
start, end = time.span('year')
time_level = 'y'
start = datetime.datetime(start.year, start.month, start.day)
end = datetime.datetime(end.year, end.month, end.day)
time = (start, end)
return Sentence(sent_dict['raw'], tokens, pub_time, time, time_level)
def temporal_graph(self, X, times):
times = [utils.strip_to_date(t) for t in times]
time_to_ixs = collections.defaultdict(list)
for i in range(len(times)):
time_to_ixs[times[i]].append(i)
n_items = X.shape[0]
S = sparse.lil_matrix((n_items, n_items))
start, end = min(times), max(times)
total_days = (end - start).days + 1
for n in range(total_days + 1):
t = start + datetime.timedelta(days=n)
window_size = min(self.max_days + 1, total_days + 1 - n)
window = [
t + datetime.timedelta(days=k) for k in range(window_size)
]
if n == 0 or len(window) == 1:
indices = [i for t in window for i in time_to_ixs[t]]
if len(indices) == 0:
continue
if sparse.issparse(X):
X_n = sparse.vstack([X[i] for i in indices])
else:
X_n = np.vstack([X[i] for i in indices])
S_n = cosine_similarity(X_n)
n_items = len(indices)
for i_x, i_n in zip(indices, range(n_items)):
for j_x, j_n in zip(indices, range(i_n + 1, n_items)):
S[i_x, j_x] = S_n[i_n, j_n]
else:
# prev is actually prev + new
prev_indices = [i for t in window for i in time_to_ixs[t]]
new_indices = time_to_ixs[window[-1]]
if len(new_indices) == 0:
continue
if sparse.issparse(X):
X_prev = sparse.vstack([X[i] for i in prev_indices])
X_new = sparse.vstack([X[i] for i in new_indices])
else:
X_prev = np.vstack([X[i] for i in prev_indices])
X_new = np.vstack([X[i] for i in new_indices])
S_n = cosine_similarity(X_prev, X_new)
n_prev, n_new = len(prev_indices), len(new_indices)
for i_x, i_n in zip(prev_indices, range(n_prev)):
for j_x, j_n in zip(new_indices, range(n_new)):
S[i_x, j_x] = S_n[i_n, j_n]
return sparse.csr_matrix(S)
def get_input_time_span(ref_dates, extension):
ref_start = utils.strip_to_date(min(ref_dates))
ref_end = utils.strip_to_date(max(ref_dates))
input_start = ref_start - datetime.timedelta(days=extension)
input_end = ref_end + datetime.timedelta(days=extension)
return input_start, input_end