def cal_rouge_gain(src_txt, tgt_txt, c_idxs, rouge2_ratio=0.5, do_norm=True):
'''
sents: all the sentences
abstract: target ground truth text
c_idxs: collection of selected indexs
'''
sents = [sent.split() for sent in src_txt]
abstract = tgt_txt.replace("<q>", " ").split()
evaluated_1grams = [_get_word_ngrams(1, [sent], do_count=True) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract], do_count=True)
evaluated_2grams = [_get_word_ngrams(2, [sent], do_count=True) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract], do_count=True)
#print(reference_1grams, reference_2grams)
#print(evaluated_1grams, evaluated_2grams)
if c_idxs is None or len(c_idxs) == 0:
rouge_score = 0.0
candidates_1 = coll.defaultdict(int)
candidates_2 = coll.defaultdict(int)
else:
candidates_1 = [evaluated_1grams[idx] for idx in c_idxs if idx > -1] #fix a little bug
#candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c_idxs if idx > -1]
#candidates_2 = set.union(*map(set, candidates_2))
candidates_1 = union_dic(candidates_1)
candidates_2 = union_dic(candidates_2)
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 * (1-rouge2_ratio) + rouge_2 * rouge2_ratio
#print(rouge_score)
min_gain = float('inf')
max_gain = float('-inf')
rouge_gain = [float('-inf')] * len(sents)
for i in range(len(sents)):
if c_idxs is not None and i in c_idxs:
continue
unigrams = union_dic([candidates_1, evaluated_1grams[i]])
bigrams = union_dic([candidates_2, evaluated_2grams[i]])
#unigrams = candidates_1.union(set(evaluated_1grams[i]))
#bigrams = candidates_2.union(set(evaluated_2grams[i]))
#print(unigrams, bigrams)
cur_rouge_1 = cal_rouge(unigrams, reference_1grams)['f']
cur_rouge_2 = cal_rouge(bigrams, reference_2grams)['f']
#print(cur_rouge_1, cur_rouge_2)
cur_rouge_score = cur_rouge_1 * (1-rouge2_ratio) + cur_rouge_2 * rouge2_ratio
#cur_rouge_score = cur_rouge_1 + cur_rouge_2
cur_gain = cur_rouge_score - rouge_score
rouge_gain[i] = cur_gain
max_gain = max(max_gain, cur_gain)
min_gain = min(min_gain, cur_gain)
gain_gap = max_gain - min_gain
#print(gain_gap)
rouge_gain = np.asarray(rouge_gain)
#print(rouge_gain)
if do_norm and gain_gap > 0:
rouge_gain = (rouge_gain - min_gain) / gain_gap
#same after softmax
#do clip
#rouge_gain[rouge_gain<0] = float('-inf')
return rouge_gain
def combination_selection(doc_sent_list, abstract_sent_list, summary_size):
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
max_idx = (0, 0)
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
impossible_sents = []
for s in range(summary_size + 1):
combinations = itertools.combinations(
[i for i in range(len(sents)) if i not in impossible_sents], s + 1)
for c in combinations:
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
if (s == 0 and rouge_score == 0):
impossible_sents.append(c[0])
if rouge_score > max_rouge:
max_idx = c
max_rouge = rouge_score
return sorted(list(max_idx))
def cal_rouge_doc(src_txt, tgt_txt, c_idxs, c_masks=None):
# one entry of src, tgt and selected sentences
abstract = tgt_txt.replace("<q>", " ").split()
reference_1grams = _get_word_ngrams(1, [abstract], do_count=True)
reference_2grams = _get_word_ngrams(2, [abstract], do_count=True)
#the following two are same
if c_masks is None:
sents = [src_txt[i].split() for i in c_idxs if i > -1]
else:
sents = [
src_txt[idx].split() for i, idx in enumerate(c_idxs) if c_masks[i]
]
#-1 is padding id
evaluated_1grams = [
_get_word_ngrams(1, [sent], do_count=True) for sent in sents
]
evaluated_2grams = [
_get_word_ngrams(2, [sent], do_count=True) for sent in sents
]
#candidates_1 = set.union(*map(set, evaluated_1grams))
#candidates_2 = set.union(*map(set, evaluated_2grams))
candidates_1 = union_dic(evaluated_1grams)
candidates_2 = union_dic(evaluated_2grams)
rouge_1 = cal_rouge(candidates_1, reference_1grams)
rouge_2 = cal_rouge(candidates_2, reference_2grams)
#print(rouge_1['p'], rouge_1['r'], rouge_1['f'])
#print(rouge_2['p'], rouge_2['r'], rouge_2['f'])
rouge_score = rouge_1['f'] + rouge_2['f']
return rouge_score
def presumm_reward3(doc_sent_list, abstract_sent_list, mode='f'):
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
rwds = []
prev_sc = 0
# cur_max_rouge = max_rouge
# cur_id = -1
for i in range(len(sents)): #source
# if (i in selected):
# continue
c = selected + [i]
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)[mode]
rouge_2 = cal_rouge(candidates_2, reference_2grams)[mode]
rouge_score = rouge_1 + rouge_2 - prev_sc #max with it 0, use the r version, give -0.1, -0.2 ... for EOE
rwds.append(rouge_score + 1)
prev_sc = rouge_1 + rouge_2
return rwds
def get_label_orders(batch_src_sents, batch_tgt_str,
batch_labels): #, temperature=20.0):
'''batch_src_sents: batch of list of source sentences containing tokens
batch_tgt_str: batch of ground truth summary
batch_labels: batch of labels such as [0,0,0,1,0] indicating ground truth sentences
return batch of orders according to max marginal gain [[2,1,0],[6,9]] before padding
'''
# get rouge of selected sentences according to ground truth.
# for each candidate sentence, get the rouge gain of each sentence.
batch_size = len(batch_src_sents)
batch_label_seq = []
for i in range(batch_size):
label = batch_labels[i]
rel_idxs = [idx for idx in range(len(label)) if label[idx] == 1]
selected_idxs = []
src_txt, tgt_txt = batch_src_sents[i], batch_tgt_str[i]
sents = [src_txt[x].split() for x in rel_idxs]
abstract = tgt_txt.replace("<q>", " ").split()
evaluated_1grams = [
_get_word_ngrams(1, [sent], do_count=True) for sent in sents
]
reference_1grams = _get_word_ngrams(1, [abstract], do_count=True)
evaluated_2grams = [
_get_word_ngrams(2, [sent], do_count=True) for sent in sents
]
reference_2grams = _get_word_ngrams(2, [abstract], do_count=True)
#print('ref', reference_1grams)
idxs = list(range(len(rel_idxs)))
#print('eval', evaluated_1grams)
#print(rel_idxs)
while len(idxs) > 0:
cur_id = -1
cur_max_rouge = float('-inf')
#cur_max_rouge should be larger than 0, in case of any illegal case, relax it to be possiblely negative
for r_idx in idxs:
c = selected_idxs + [r_idx]
candidates_1 = [evaluated_1grams[idx] for idx in c]
#candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
#candidates_2 = set.union(*map(set, candidates_2))
candidates_1 = union_dic(candidates_1)
candidates_2 = union_dic(candidates_2)
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
#print(candidates_1)
#print(r_idx, c, rouge_score, cur_max_rouge)
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = r_idx
#print(cur_id)
assert cur_id > -1
idxs.remove(cur_id)
selected_idxs += [cur_id]
label_seq = [rel_idxs[x] for x in selected_idxs]
batch_label_seq.append(label_seq)
return batch_label_seq
def greedy_selection_given_context(src_txt, tgt_txt, c_idxs, summary_size=3):
sents = [sent.split() for sent in src_txt]
abstract = tgt_txt.replace("<q>", " ").split()
evaluated_1grams = [
_get_word_ngrams(1, [sent], do_count=True) for sent in sents
]
reference_1grams = _get_word_ngrams(1, [abstract], do_count=True)
evaluated_2grams = [
_get_word_ngrams(2, [sent], do_count=True) for sent in sents
]
reference_2grams = _get_word_ngrams(2, [abstract], do_count=True)
#print(reference_1grams, reference_2grams)
#print(evaluated_1grams, evaluated_2grams)
max_rouge = 0.0
selected = []
if c_idxs is not None:
for cid in c_idxs:
if cid > -1:
selected.append(cid)
sel_count = len(selected)
summary_sent_count = summary_size - len(selected)
#print(sel_count)
for s in range(summary_sent_count):
cur_max_rouge = max_rouge
cur_id = -1
for i in range(len(sents)):
if (i in selected):
continue
c = selected + [i]
candidates_1 = [evaluated_1grams[idx] for idx in c]
#candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
#candidates_2 = set.union(*map(set, candidates_2))
candidates_1 = union_dic(candidates_1)
candidates_2 = union_dic(candidates_2)
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = i
if (cur_id == -1):
return selected[sel_count:]
selected.append(cur_id)
max_rouge = cur_max_rouge
return selected[sel_count:]
def cal_rouge_score(pred_list, truth_list):
# one entry of src, tgt and selected sentences
rouge1_arr, rouge2_arr = [], []
for pred_txt, tgt_txt in zip(pred_list, truth_list):
gold = tgt_txt.replace("<q>", " ").split()
reference_1grams = _get_word_ngrams(1, [gold], do_count=True)
reference_2grams = _get_word_ngrams(2, [gold], do_count=True)
candi = pred_txt.replace("<q>", " ").split()
candidates_1 = _get_word_ngrams(1, [candi], do_count=True)
candidates_2 = _get_word_ngrams(2, [candi], do_count=True)
rouge_1 = cal_rouge(candidates_1, reference_1grams)
rouge1_arr.append(rouge_1)
rouge_2 = cal_rouge(candidates_2, reference_2grams)
rouge2_arr.append(rouge_2)
return rouge1_arr, rouge2_arr
def each_selection(doc_sent_list, abstract_sent_list):
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
# abstract = sum(abstract_sent_list, [])
# abstract = _rouge_clean(' '.join(abstract)).split()
# sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
# evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
# reference_1grams = _get_word_ngrams(1, [abstract])
# evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
# reference_2grams = _get_word_ngrams(2, [abstract])
abstract = sum(abstract_sent_list, [])
abstracts = [_rouge_clean(' '.join(s)).split() for s in abstract]
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = [_get_word_ngrams(1, [sent]) for sent in abstracts]
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = [_get_word_ngrams(2, [sent]) for sent in abstracts]
selected = []
for _abstract in range(len(abstracts)):
rouge_score = []
reference_1 = set(reference_1grams[_abstract])
reference_2 = set(reference_2grams[_abstract])
# print('------------------------------------------------------------')
# print(len(sents))
# print(len(abstracts))
for _sent in range(len(sents)):
if _sent in rouge_score:
continue
candidates_1 = set(evaluated_1grams[_sent])
candidates_2 = set(evaluated_2grams[_sent])
rouge_1 = cal_rouge(candidates_1, reference_1)['f']
rouge_2 = cal_rouge(candidates_2, reference_2)['f']
_rouge_score = rouge_1 + rouge_2
# print(_rouge_score)
rouge_score.append(_rouge_score)
# print('length',len(rouge_score))
if len(rouge_score) > 0:
cur_index = rouge_score.index(max(rouge_score))
if cur_index not in selected:
selected.append(cur_index)
return sorted(selected)
def greedy_selection(doc_sent_list, abstract_sent_list, summary_size):
"""
Algorithm used to create gold summaries from hand created target. It compare every sentence with hand maded one and calculate 2-ROUGE score. After that
It sort the sentence the sentence by best score.
Input: DOCUMENT, ABSTRACT GIVEN BY USER, SUMMARY MAX SIZE(number of sentence)
Output: summary_size gold sentence which maximize the 2 rouge against hand written abstract.
"""
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
for s in range(summary_size):
cur_max_rouge = max_rouge
cur_id = -1
for i in range(len(sents)):
if (i in selected):
continue #don't use duplicate
c = selected + [i]
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = i
if (cur_id == -1):
return selected
selected.append(cur_id)
max_rouge = cur_max_rouge
return sorted(selected)
def greedy_selection2(doc_sent_list, abstract_sent_list, summary_size):
max_rouge = 0.0
#日本語の場合
doc_sent_list = doc_sent_list.split('。')
doc_sent_list = [a + '。' for a in doc_sent_list]
sents = doc_sent_list[:-1]
abstract = tokenizer.tokenize(abstract_sent_list)
sents = [tokenizer.tokenize(a) for a in sents]
#abstract = _rouge_clean(' '.join(abstract_sent_list)).split()
#sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
for s in range(summary_size):
cur_max_rouge = max_rouge
cur_id = -1
for i in range(len(sents)):
if (i in selected):
continue
c = selected + [i]
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = i
if (cur_id == -1):
return selected
selected.append(cur_id)
max_rouge = cur_max_rouge
print(type(sorted(selected)))
return sorted(selected)
def greedy_selection(doc_sent_list, abstract_sent_list, summary_size):
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
for s in range(summary_size):
cur_max_rouge = max_rouge
cur_id = -1
for i in range(len(sents)): #source
if (i in selected):
continue
c = selected + [i]
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
# print(cur_max_rouge)
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = i
if (cur_id == -1):
return selected
selected.append(cur_id)
max_rouge = cur_max_rouge
# print("output", selected)
# if not selected:
# print("empty selected")
# selected.append(0)
return selected #sorted(selected)
def sentence_rouge(doc_sent_list, abstract_sent_list):
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
sent_scores = []
for i in range(len(sents)):
candidates_1 = set.union(*map(set, [evaluated_1grams[i]]))
candidates_2 = set.union(*map(set, [evaluated_2grams[i]]))
rouge_1 = cal_rouge(candidates_1, reference_1grams)
rouge_2 = cal_rouge(candidates_2, reference_2grams)
sent_scores.append({'rouge_1': rouge_1, 'rouge_2': rouge_2})
return sent_scores
def greedy_selection(doc_sent_list, abstract_sent_list, summary_size):
def _rouge_clean(s):
return re.sub(r"[^a-zA-Z0-9 ]", "", s)
max_rouge = 0.0
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(" ".join(abstract)).split()
sents = [_rouge_clean(" ".join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
for s in range(summary_size):
cur_max_rouge = max_rouge
cur_id = -1
for i in range(len(sents)):
if i in selected:
continue
c = selected + [i]
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)["f"]
rouge_2 = cal_rouge(candidates_2, reference_2grams)["f"]
rouge_score = rouge_1 + rouge_2
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = i
if cur_id == -1:
return selected
selected.append(cur_id)
max_rouge = cur_max_rouge
return sorted(selected)
def presumm_reward4(doc_sent_list, abstract_sent_list, n_sents=10):
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
rwds = [0]
# cur_max_rouge = max_rouge
# cur_id = -1
for i in range(len(sents)): #source
# if (i in selected):
# continue
# c = selected + [i]
# candidates_1 = [evaluated_1grams[idx] for idx in c]
# candidates_1 = set.union(*map(set, candidates_1))
# candidates_2 = [evaluated_2grams[idx] for idx in c]
# candidates_2 = set.union(*map(set, candidates_2))
# rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
# rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
if i < n_sents:
rouge_1 = 0 #cal_rouge(evaluated_1grams[i], reference_1grams)['f']
rouge_2 = cal_rouge(evaluated_2grams[i], reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
else:
rouge_score = 0
rwds.append(rouge_score)
return rwds[1:]
def greedy_selection(doc_sent_list, abstract_sent_list, summary_size):
""" Create pseudo extraction labels.
Args:
doc_sent_list (list[list[str]]):
source text to be processed.
abstract_sent_list (list[list[str]]):
target text to be processed.
summary_size(int) :
maximum number of extracted sentences.
Returns:
A list of extracted sentence indices in ascending order.
"""
def _rouge_clean(s):
return re.sub(r'[^a-zA-Z0-9 ]', '', s)
max_rouge = 0.0
# Clean and concat all target sentences
abstract = sum(abstract_sent_list, [])
abstract = _rouge_clean(' '.join(abstract)).split()
# Clean all source sentences
sents = [_rouge_clean(' '.join(s)).split() for s in doc_sent_list]
# Get 1 grams and 2 grams from source and target
evaluated_1grams = [_get_word_ngrams(1, [sent]) for sent in sents]
reference_1grams = _get_word_ngrams(1, [abstract])
evaluated_2grams = [_get_word_ngrams(2, [sent]) for sent in sents]
reference_2grams = _get_word_ngrams(2, [abstract])
selected = []
for s in range(summary_size):
cur_max_rouge = max_rouge
cur_id = -1
# Iterates through all sentences
for i in range(len(sents)):
if (i in selected):
continue
# Consider selected and candidate sentences together
c = selected + [i]
# Calculate ROUGE-1-F + ROUGE-2-F with target
candidates_1 = [evaluated_1grams[idx] for idx in c]
candidates_1 = set.union(*map(set, candidates_1))
candidates_2 = [evaluated_2grams[idx] for idx in c]
candidates_2 = set.union(*map(set, candidates_2))
rouge_1 = cal_rouge(candidates_1, reference_1grams)['f']
rouge_2 = cal_rouge(candidates_2, reference_2grams)['f']
rouge_score = rouge_1 + rouge_2
if rouge_score > cur_max_rouge:
cur_max_rouge = rouge_score
cur_id = i
# If no sentence exceeds current max score then stop
if (cur_id == -1):
return selected
# Record currently chosen sentence and score
selected.append(cur_id)
max_rouge = cur_max_rouge
return sorted(selected)