found a bug when the input text's length exceeds max_seq_length

[flyangovoyang]

hello, i am coming again, (nice work deserves deep reading~~)

yesterday's statement has been reorganized.

when the input text's length exceeds max_seq_length, targets in the extra part will be left out since all_doc_token has been truncated according to max_seq_length, however, start/end indexes and polarities in the extra part has been kept in tok_start_positions/tok_end_positions and polarity_labels respectively

SpanABSA/absa/utils.py

Lines 101 to 115 in 66369af

	tok_start_positions = []
	tok_end_positions = []
	for start_position, end_position in \
	zip(example.start_positions, example.end_positions):
	tok_start_position = orig_to_tok_index[start_position]
	if end_position < len(example.sent_tokens) - 1:
	tok_end_position = orig_to_tok_index[end_position + 1] - 1
	else:
	tok_end_position = len(all_doc_tokens) - 1
	tok_start_positions.append(tok_start_position)
	tok_end_positions.append(tok_end_position)
	# Account for [CLS] and [SEP] with "- 2"
	if len(all_doc_tokens) > max_seq_length - 2:
	all_doc_tokens = all_doc_tokens[0:(max_seq_length - 2)]

later, initialize tokens from all_doc_token with the [CLS] and [SEP] added.

SpanABSA/absa/utils.py

Lines 117 to 128 in 66369af

	tokens = []
	token_to_orig_map = {}
	segment_ids = []
	tokens.append("[CLS]")
	segment_ids.append(0)
	for index, token in enumerate(all_doc_tokens):
	token_to_orig_map[len(tokens)] = tok_to_orig_index[index]
	tokens.append(token)
	segment_ids.append(0)
	tokens.append("[SEP]")
	segment_ids.append(0)

convert tokens to input_ids, and pad it to max_seq_length

SpanABSA/absa/utils.py

Lines 130 to 136 in 66369af

	input_ids = tokenizer.convert_tokens_to_ids(tokens)
	input_mask = [1] * len(input_ids)
	while len(input_ids) < max_seq_length:
	input_ids.append(0)
	input_mask.append(0)
	segment_ids.append(0)

at this moment, the code initializes start_position and end_positions with the length of input_ids which has beed truncated, but the for-loop ranges from tok_start_positions/tok_end_positions and example.polarities, which has not been truncated, this may lead to index_out_of_bound exception.

SpanABSA/absa/utils.py

Lines 142 to 163 in 66369af

	# For distant supervision, we annotate the positions of all answer spans
	start_positions = [0] * len(input_ids)
	end_positions = [0] * len(input_ids)
	bio_labels = [0] * len(input_ids)
	polarity_positions = [0] * len(input_ids)
	start_indexes, end_indexes = [], []
	for tok_start_position, tok_end_position, polarity in zip(tok_start_positions, tok_end_positions, example.polarities):
	if (tok_start_position >= 0 and tok_end_position <= (max_seq_length - 1)):
	start_position = tok_start_position + 1 # [CLS]
	end_position = tok_end_position + 1 # [CLS]
	start_positions[start_position] = 1
	end_positions[end_position] = 1
	start_indexes.append(start_position)
	end_indexes.append(end_position)
	term_length = tok_end_position - tok_start_position + 1
	max_term_length = term_length if term_length > max_term_length else max_term_length
	bio_labels[start_position] = 1 # 'B'
	if start_position < end_position:
	for idx in range(start_position + 1, end_position + 1):
	bio_labels[idx] = 2 # 'I'
	for idx in range(start_position, end_position + 1):
	polarity_positions[idx] = label_to_id[polarity]

[huminghao16]

Thank you for your interests in our work.
This is indeed a potential bug. Since we set the max_seq_length to be larger than the maximum length of sentence in the dataset, this bug will not occur in the ABSA task.
However, once the sentence length exceeds max_seq_length, the may lead to index_out_of_bound exception.
We will fix this bug as soon as possible.

[bilalghanem]

A solution for this:

in SpanABSA>absa>utils.py in function: wrapped_get_final_text:

add:
if end_index not in feature.token_to_orig_map:
feature.token_to_orig_map[end_index] = feature.token_to_orig_map[end_index-1]+1

before the line:
orig_doc_end = feature.token_to_orig_map[end_index]