def test_token_aligner_project_single_token_index():
source_tokens = ["abc", "def", "ghi", "jkl"]
target_tokens = ["abc", "d", "ef", "ghi", "jkl"]
ta = TokenAligner(source_tokens, target_tokens)
m = ta.project_token_idxs(1)
m_expected = np.array([1, 2])
assert (m == m_expected).all()
def test_token_aligner_project_to_empty_target_token_sequence():
source_tokens = ["abc", "def", "ghi", "jkl"]
target_tokens = []
ta = TokenAligner(source_tokens, target_tokens)
m = ta.project_tokens([1, 3])
m_expected = np.array([])
assert (m == m_expected).all()
def test_token_aligner_project_to_mismatched_token_sequence():
source_tokens = ["abc", "def", "ghi", "jkl"]
target_tokens = ["qrs", "tuv", "wxy", "z"]
ta = TokenAligner(source_tokens, target_tokens)
m = ta.project_tokens([1])
m_expected = np.array([])
assert (m == m_expected).all()
def test_token_aligner_project_span():
source_tokens = ["abc", "def", "ghi", "jkl"]
target_tokens = ["abc", "d", "ef", "ghi", "jkl"]
ta = TokenAligner(source_tokens, target_tokens)
m = ta.project_span(1, 2)
m_expected = np.array([1, 3])
assert (m == m_expected).all()
def test_moses_tok_idx_proj_3():
src_tokens = [
"Mr.",
"Immelt",
"chose",
"to",
"focus",
"on",
"the",
"incomprehensibility",
"of",
"accounting",
"rules.",
]
tgt_tokens = [
"Mr.",
"Immelt",
"chose",
"to",
"focus",
"on",
"the",
"incomprehensibility",
"of",
"accounting",
"rules",
".",
]
tgt_token_index = [[0], [1], [2], [3], [4], [5], [6], [7], [8], [9],
[10, 11]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_token_aligner_project_span_last_token_range_is_end_exclusive():
source_tokens = ["abc", "def", "ghi", "jkl"]
target_tokens = ["abc", "d", "ef", "ghi", "jkl"]
ta = TokenAligner(source_tokens, target_tokens)
m = ta.project_span(3, 4)
m_expected = np.array([4, 5])
assert (m == m_expected).all()
def test_token_aligner_project_multiple_token_indices():
source_tokens = ["abc", "def", "ghi", "jkl"]
target_tokens = ["abc", "d", "ef", "ghi", "jkl"]
ta = TokenAligner(source_tokens, target_tokens)
m = ta.project_tokens([1, 3])
m_expected = np.array([1, 2, 4])
assert (m == m_expected).all()
def test_project_span_covering_whole_sequence():
src_tokens = ["Members", "of", "the", "House", "clapped", "their", "hands"]
tgt_tokens = [
"Members", "Ġof", "Ġthe", "ĠHouse", "Ġcl", "apped", "Ġtheir", "Ġhands"
]
# reference: tgt_token_index = [[0], [1], [2], [3], [4, 5], [6], [7]]
ta = TokenAligner(src_tokens, tgt_tokens)
assert (0, 8) == ta.project_span(0, 7)
def test_bytebpe_tok_idx_proj_4():
src_tokens = ["What?"]
tgt_tokens = ["What", "?"]
tgt_token_index = [[0, 1]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_wpm_tok_idx_proj_1():
src_tokens = ["Members", "of", "the", "House", "clapped", "their", "hands"]
tgt_tokens = ["Members", "of", "the", "House", "clapped", "their", "hands"]
tgt_token_index = [[0], [1], [2], [3], [4], [5], [6]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def _create_examples(self, qa_file_path, set_type):
wiki_df = pd.read_csv(self.path_dict["wiki_dict"],
sep="\t",
names=["sent_id", "text"])
wiki_dict = {
row.sent_id: row.text
for row in wiki_df.itertuples(index=False)
}
data_df = pd.read_csv(
qa_file_path,
sep="\t",
header=None,
names=[
"sent_id",
"target_ids",
"worker_id",
"qa_index",
"qa_word",
"question",
"answer",
"response1",
"response2",
],
)
data_df["sent"] = data_df["sent_id"].apply(wiki_dict.get)
examples = []
ptb_detokenizer = nltk.tokenize.treebank.TreebankWordDetokenizer()
for i, row in enumerate(data_df.itertuples(index=False)):
# Answer indices are a space-limited list of numbers.
# We simply take the min/max of the indices
answer_idxs = list(map(int, row.answer.split()))
answer_token_start, answer_token_end = min(answer_idxs), max(
answer_idxs)
passage_ptb_tokens = row.sent.split()
passage_space_tokens = ptb_detokenizer.detokenize(
passage_ptb_tokens, convert_parentheses=True).split()
passage_space_str = " ".join(passage_space_tokens)
token_aligner = TokenAligner(source=passage_ptb_tokens,
target=passage_space_tokens)
answer_char_span = token_aligner.project_token_to_char_span(
answer_token_start, answer_token_end, inclusive=True)
answer_str = passage_space_str[
answer_char_span[0]:answer_char_span[1] + 1]
examples.append(
span_pred_template.Example(
guid="%s-%s" % (set_type, i),
passage=passage_space_str,
question=row.question,
answer=answer_str,
answer_char_span=answer_char_span,
))
return examples
def test_project_invalid_span():
src_tokens = ["Members", "of", "the", "House", "clapped", "their", "hands"]
tgt_tokens = [
"Members", "Ġof", "Ġthe", "ĠHouse", "Ġcl", "apped", "Ġtheir", "Ġhands"
]
# reference: tgt_token_index = [[0], [1], [2], [3], [4, 5], [6], [7]]
ta = TokenAligner(src_tokens, tgt_tokens)
with pytest.raises(ValueError):
ta.project_span(0, 0)
def test_sentencepiece_tok_idx_proj_3():
src_tokens = [
"Mr.",
"Immelt",
"chose",
"to",
"focus",
"on",
"the",
"incomprehensibility",
"of",
"accounting",
"rules.",
]
tgt_tokens = [
"▁Mr",
".",
"▁I",
"m",
"mel",
"t",
"▁chose",
"▁to",
"▁focus",
"▁on",
"▁the",
"▁in",
"comp",
"re",
"hen",
"s",
"ibility",
"▁of",
"▁accounting",
"▁rules",
".",
]
tgt_token_index = [
[0, 1],
[2, 3, 4, 5],
[6],
[7],
[8],
[9],
[10],
[11, 12, 13, 14, 15, 16],
[17],
[18],
[19, 20],
]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_moses_tok_idx_proj_2():
src_tokens = ["I", "look", "at", "Sarah's", "dog.", "It", "was", "cute.!"]
tgt_tokens = [
"I", "look", "at", "Sarah", "'s", "dog", ".", "It", "was", "cute",
".", "!"
]
tgt_token_index = [[0], [1], [2], [3, 4], [5, 6], [7], [8], [9, 10, 11]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_wpm_tok_idx_proj_3():
src_tokens = [
"Mr.",
"Immelt",
"chose",
"to",
"focus",
"on",
"the",
"incomprehensibility",
"of",
"accounting",
"rules.",
]
tgt_tokens = [
"Mr",
".",
"I",
"##mme",
"##lt",
"chose",
"to",
"focus",
"on",
"the",
"in",
"##com",
"##p",
"##re",
"##hen",
"##si",
"##bility",
"of",
"accounting",
"rules",
".",
]
tgt_token_index = [
[0, 1],
[2, 3, 4],
[5],
[6],
[7],
[8],
[9],
[10, 11, 12, 13, 14, 15, 16],
[17],
[18],
[19, 20],
]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_wpm_tok_idx_proj_2():
src_tokens = ["I", "look", "at", "Sarah's", "dog.", "It", "was", "cute.!"]
tgt_tokens = [
"I", "look", "at", "Sarah", "'", "s", "dog", ".", "It", "was", "cute",
".", "!"
]
tgt_token_index = [[0], [1], [2], [3, 4, 5], [6, 7], [8], [9],
[10, 11, 12]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def tokenize(self, tokenizer):
passage_tokens = tokenizer.tokenize(self.passage)
token_aligner = TokenAligner(source=self.passage, target=passage_tokens)
answer_token_span = token_aligner.project_char_to_token_span(
self.answer_char_span[0], self.answer_char_span[1], inclusive=True
)
return TokenizedExample(
guid=self.guid,
passage=passage_tokens,
question=tokenizer.tokenize(self.question),
answer_str=self.answer,
passage_str=self.passage,
answer_token_span=answer_token_span,
token_idx_to_char_idx_map=token_aligner.source_char_idx_to_target_token_idx.T,
)
def test_bpe_tok_idx_proj_2():
src_tokens = ["I", "look", "at", "Sarah's", "dog.", "It", "was", "cute.!"]
tgt_tokens = [
"i</w>",
"look</w>",
"at</w>",
"sarah</w>",
"'s</w>",
"dog</w>",
".</w>",
"it</w>",
"was</w>",
"cute</w>",
".</w>",
"!</w>",
]
tgt_token_index = [[0], [1], [2], [3, 4], [5, 6], [7], [8], [9, 10, 11]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_bpe_tok_idx_proj_3():
src_tokens = [
"Mr.",
"Immelt",
"chose",
"to",
"focus",
"on",
"the",
"incomprehensibility",
"of",
"accounting",
"rules.",
]
tgt_tokens = [
"mr.</w>",
"im",
"melt</w>",
"chose</w>",
"to</w>",
"focus</w>",
"on</w>",
"the</w>",
"in",
"comprehen",
"si",
"bility</w>",
"of</w>",
"accounting</w>",
"rules</w>",
".</w>",
]
tgt_token_index = [[0], [1, 2], [3], [4], [5], [6], [7], [8, 9, 10, 11],
[12], [13], [14, 15]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_bytebpe_tok_idx_proj_3():
src_tokens = [
"Mr.",
"Immelt",
"chose",
"to",
"focus",
"on",
"the",
"incomprehensibility",
"of",
"accounting",
"rules.",
]
tgt_tokens = [
"Mr",
".",
"ĠImm",
"elt",
"Ġchose",
"Ġto",
"Ġfocus",
"Ġon",
"Ġthe",
"Ġincomp",
"rehens",
"ibility",
"Ġof",
"Ġaccounting",
"Ġrules",
".",
]
tgt_token_index = [[0, 1], [2, 3], [4], [5], [6], [7], [8], [9, 10, 11],
[12], [13], [14, 15]]
ta = TokenAligner(src_tokens, tgt_tokens)
for src_token_idx in range(len(src_tokens)):
projected_tgt_tok_idx = ta.project_tokens(src_token_idx)
assert (tgt_token_index[src_token_idx] == projected_tgt_tok_idx).all()
def test_private_project_token_span():
mat = np.eye(5, dtype=int)
mat[0][0] = 0
mat[3][3] = 0
assert TokenAligner._project_span(mat, 1, 3, inclusive=True) == (1, 2)
assert TokenAligner._project_span(mat, 1, 3, inclusive=False) == (1, 3)
assert TokenAligner._project_span(mat, 1, 2, inclusive=True) == (1, 2)
assert TokenAligner._project_span(mat, 1, 2, inclusive=False) == (1, 2)
assert TokenAligner._project_span(mat, 1, 4, inclusive=True) == (1, 4)
assert TokenAligner._project_span(mat, 1, 4, inclusive=False) == (1, 3)
def _create_examples(self, file_path, set_type):
with gzip.open(file_path) as f:
lines = f.read().splitlines()
examples = []
ptb_detokenizer = nltk.tokenize.treebank.TreebankWordDetokenizer()
for line in lines:
datum = json.loads(line)
datum = {
"sentence_tokens":
datum["sentenceTokens"],
"entries": [{
"verb": verb_entry["verbInflectedForms"]["stem"],
"verb_idx": verb_idx,
"questions": {
question:
[[{
"tokens": datum["sentenceTokens"][span[0]:span[1]],
"span": (span[0], span[1] - 1),
} for span in answer_judgment["spans"]]
for answer_judgment in q_data["answerJudgments"]
if answer_judgment["isValid"]]
for question, q_data in
verb_entry["questionLabels"].items()
},
} for verb_idx, verb_entry in datum["verbEntries"].items()],
}
passage_ptb_tokens = datum["sentence_tokens"]
passage_space_tokens = ptb_detokenizer.detokenize(
passage_ptb_tokens, convert_parentheses=True).split()
passage_space_str = " ".join(passage_space_tokens)
token_aligner = TokenAligner(source=passage_ptb_tokens,
target=passage_space_tokens)
for entry in datum["entries"]:
for question, answer_list in entry["questions"].items():
for answer in answer_list:
for answer_span in answer:
try:
answer_char_span = token_aligner.project_token_to_char_span(
answer_span["span"][0],
answer_span["span"][1],
inclusive=True)
except ValueError:
continue
answer_str = passage_space_str[
answer_char_span[0]:answer_char_span[1] + 1]
examples.append(
span_pred_template.Example(
guid="%s-%s" % (set_type, len(examples)),
passage=passage_space_str,
question=question,
answer=answer_str,
answer_char_span=answer_char_span,
))
return examples
