def test_read_proj_ds_tree(self):
src_t = get_ds(self.inst2, trans(self.inst2))
tgt_w = lang(self.inst2)
aln = get_trans_gloss_alignment(self.inst2)
tgt_t = DepTree.fromstring("""
(ROOT[0]
(glaubst[2]
(Was[1])
(Du[3])
(wer[4])
(angerufen[5] (hat[6]))
))
""", stype=DEPSTR_PTB)
proj_t = project_ds(src_t, tgt_w, aln)
self.assertTrue(proj_t.structurally_eq(tgt_t))
def split_instances(instances, train=0, dev=0, test=0):
"""
:type instances: list[RGIgt]
"""
# -- 0) Initialize the counter to keep track of which word index
# is in which sentence.
instances = list(instances)
wc = WordCount()
for i, inst in enumerate(instances):
num_words = len(lang(inst))
SPLIT_LOG.debug('{} words in sentence {} (id {})'.format(num_words, i, inst.id))
wc.add(i, num_words)
# -- 2) Figure out the number of words.
num_train_words = round(train * wc.num_words)
num_dev_words = round(dev * wc.num_words)
num_test_words = round(test * wc.num_words)
# -- 3) Get the word indices.
train_word_index = num_train_words
dev_word_index = train_word_index + num_dev_words
test_word_index = dev_word_index + num_test_words
# -- 4) Figure out which sentence indices these refer to.
train_sent_index = wc.get_snt_from_wordnum(train_word_index)
dev_sent_index = wc.get_snt_from_wordnum(dev_word_index)
test_sent_index = wc.get_snt_from_wordnum(test_word_index)
# -- 4) Now, split up the data.
train_instances = instances[0:train_sent_index]
dev_instances = instances[train_sent_index:dev_sent_index]
test_instances = instances[dev_sent_index:test_sent_index+1]
# And return...
return train_instances, dev_instances, test_instances
def naacl_to_xigt(naacl_path):
"""
Convert the NAACL format to XIGT.
:param naacl_path:
"""
content = open(naacl_path, 'r').read()
# First, collect all the instances.
instances = re.findall('Igt_id[\s\S]+?Q6.*Answer', content)
xc = XigtCorpus()
for instance_txt in instances:
# id = re.search('Igt_id=([\S]+)', instance_txt).group(1)
inst = Igt(id='i{}'.format(len(xc)))
lang_raw, gloss_raw, trans_raw = instance_txt.split('\n')[1:4]
# Now, create the raw tier...
raw_tier = Tier(id=gen_tier_id(inst, 'r'), type='odin', attributes={STATE_ATTRIBUTE:RAW_STATE})
raw_tier.append(Item(id=ask_item_id(raw_tier), text=lang_raw, attributes={ODIN_TAG_ATTRIBUTE:ODIN_LANG_TAG}))
raw_tier.append(Item(id=ask_item_id(raw_tier), text=gloss_raw, attributes={ODIN_TAG_ATTRIBUTE:ODIN_GLOSS_TAG}))
raw_tier.append(Item(id=ask_item_id(raw_tier), text=trans_raw, attributes={ODIN_TAG_ATTRIBUTE:ODIN_TRANS_TAG}))
inst.append(raw_tier)
xc.append(inst)
# Generate the clean/normal tiers, but without any cleaning.
generate_normal_tier(inst, clean=False)
# Lang Dependency representation handling...
lang_ds_str = re.search('Q6:([\s\S]+?)Q6:', instance_txt).group(1)
lang_ds_lines = lang_ds_str.split('\n')[5:-3]
try:
lang_dt = parse_naacl_dep(lang(inst), lang_ds_lines)
create_dt_tier(inst, lang_dt, lang(inst), parse_method=INTENT_POS_MANUAL)
except TreeError as te:
pass
except IndexError as ie:
pass
# Eng DS handling...
eng_ds_str = re.search('Q3:([\s\S]+?)Q3:', instance_txt).group(1)
eng_ds_lines = eng_ds_str.split('\n')[2:-3]
try:
eng_dt = parse_naacl_dep(trans(inst), eng_ds_lines)
create_dt_tier(inst, eng_dt, trans(inst), parse_method=INTENT_POS_MANUAL)
except TreeError as te:
pass
except IndexError as ie:
pass
except ValueError as ve:
pass
# Add Alignment...
biling_aln_str = re.search('Q5:([\s\S]+?)Q5:', instance_txt).group(1)
biling_aln_lines = biling_aln_str.split('\n')[4:-3]
trans_offset = trans_raw.startswith(' ')
gloss_offset = gloss_raw.startswith(' ')
try:
a = Alignment()
for line in biling_aln_lines:
gloss_s, trans_s = line.split()[0:2]
if '.' in gloss_s:
continue
gloss_i = int(gloss_s)
for trans_token in trans_s.split(','):
trans_i = int(trans_token)
if trans_i == 0:
continue
else:
if trans_offset:
trans_i -= 1
if gloss_offset:
gloss_i -= 1
a.add((trans_i, gloss_i))
except:
pass
set_bilingual_alignment(inst, trans(inst), gloss(inst), a, aln_method=INTENT_ALN_MANUAL)
return xc
def test_gloss_projection_unaligned(self):
xc = xc_load(os.path.join(testfile_dir, "xigt/project_gloss_lang_tests.xml"))
igt = xc[0]
project_gloss_pos_to_lang(igt, tag_method=INTENT_POS_PROJ, unk_handling='keep')
self.assertEqual('UNK', pos_tag_tier(igt, lang(igt).id, INTENT_POS_PROJ)[-1].value())
def convert_pml(aln_path, out_path, hindi=True):
if hindi:
igt_data = retrieve_hindi()
else:
igt_data = retrieve_naacl()
a_root = load_xml(aln_path)
doc_a = a_root.find(".//reffile[@name='document_a']").get('href')
doc_b = a_root.find(".//reffile[@name='document_b']").get('href')
doc_a = os.path.join(os.path.join(os.path.dirname(aln_path), doc_a))
doc_b = os.path.join(os.path.join(os.path.dirname(aln_path), doc_b))
# Load the sentences for each document.
a_sents, a_glossed = load_sents(doc_a)
b_sents, b_glossed = load_sents(doc_b)
sent_alignments = a_root.findall(".//body/LM")
assert (a_glossed and not b_glossed) or (b_glossed and not a_glossed), "Only one file should have glosses"
xc = XigtCorpus()
for sent_alignment in sent_alignments:
# Get the sentence id...
aln_id = sent_alignment.attrib.get('id')
a_snt_id = re.search('^.+?-(.*)$', aln_id).group(1)
if a_snt_id not in igt_data:
continue
# Get the text and tokens from the naacl data.
pre_txt, lang_txt, gloss_txt, trans_txt = igt_data[a_snt_id]
lang_tokens = lang_txt.split()
gloss_tokens = gloss_txt.split()
trans_tokens = trans_txt.split()
a_snt_ref = sent_alignment.find('./tree_a.rf').text.split('#')[1]
b_snt_ref = sent_alignment.find('./tree_b.rf').text.split('#')[1]
word_alignments = sent_alignment.findall('./node_alignments/LM')
a_snt, a_edges = a_sents[a_snt_ref]
b_snt, b_edges = b_sents[b_snt_ref]
assert isinstance(a_snt, Sentence)
assert isinstance(b_snt, Sentence)
# -------------------------------------------
# Skip sentences if they are not found for whatever reason
# -------------------------------------------
if not a_snt or not b_snt:
continue
# -------------------------------------------
# Start constructing the IGT Instance.
# -------------------------------------------
trans_snt, trans_indices = a_snt, a_edges
gloss_snt, gloss_indices = b_snt, b_edges
if a_glossed:
trans_snt, trans_indices = b_snt, b_edges
gloss_snt, gloss_indices = a_snt, a_edges
# Hindi stuff...
if hindi:
lang_tokens = [w.text for w in gloss_snt]
lang_postags = [w.pos for w in gloss_snt]
lang_txt = ' '.join(lang_tokens)
trans_tokens = [w.text for w in trans_snt]
trans_postags = [w.pos for w in trans_snt]
trans_txt = ' '.join(trans_tokens)
gloss_tokens = [w.gloss if w.gloss else 'NULL' for w in gloss_snt]
gloss_postags = lang_postags
gloss_txt = ' '.join(gloss_tokens)
inst = Igt(id=re.sub('s-', 'igt', a_snt_ref))
nt = Tier(type=ODIN_TIER_TYPE, id=NORM_ID, attributes={STATE_ATTRIBUTE:NORM_STATE})
ll = Item(id='n1', attributes={ODIN_TAG_ATTRIBUTE:ODIN_LANG_TAG}, text=lang_txt)
gl = Item(id='n2', attributes={ODIN_TAG_ATTRIBUTE:ODIN_GLOSS_TAG}, text=gloss_txt)
tl = Item(id='n3', attributes={ODIN_TAG_ATTRIBUTE:ODIN_TRANS_TAG}, text=trans_txt)
nt.extend([ll,gl,tl])
inst.append(nt)
# -------------------------------------------
# Handle the phrase tiers
# -------------------------------------------
generate_lang_phrase_tier(inst)
generate_trans_phrase_tier(inst)
def process_postags(sent, tokens):
postags = []
for i, token in enumerate(tokens):
word = sent.getorder(i+1)
if word is None:
postags.append(None)
else:
postags.append(word.pos)
return postags
# -------------------------------------------
# Now, handle the translation words.
# -------------------------------------------
tt = create_word_tier(ODIN_TRANS_TAG, trans_tokens, trans_phrase(inst)[0])
inst.append(tt)
if not hindi:
trans_postags = process_postags(trans_snt, trans_tokens)
add_pos_tags(inst, tt.id, trans_postags, tag_method=INTENT_POS_MANUAL)
# -------------------------------------------
# Handle the words tiers...
# -------------------------------------------
wt = create_word_tier(ODIN_LANG_TAG, lang_tokens, lang_phrase(inst)[0])
gwt= create_word_tier(ODIN_GLOSS_TAG, gloss_tokens, gl)
inst.extend([wt, gwt])
# Quickly set the alignment for the gloss words.
for w, gw in zip(wt, gwt):
gw.alignment = w.id
if not hindi:
lang_postags = process_postags(gloss_snt, gloss_tokens)
gloss_postags = lang_postags
add_pos_tags(inst, wt.id, lang_postags, tag_method=INTENT_POS_MANUAL)
add_pos_tags(inst, gwt.id, gloss_postags, tag_method=INTENT_POS_MANUAL)
create_dt_tier(inst, assemble_ds(gloss_snt, gloss_indices), wt, INTENT_DS_MANUAL)
create_dt_tier(inst, assemble_ds(trans_snt, trans_indices), tt, INTENT_DS_MANUAL)
# -------------------------------------------
# Now, the word alignments.
# -------------------------------------------
a = Alignment()
for word_alignment in word_alignments:
a_ref = word_alignment.find('./a.rf').text.split('#')[1]
b_ref = word_alignment.find('./b.rf').text.split('#')[1]
a_word = a_snt.getid(a_ref)
b_word = b_snt.getid(b_ref)
if a_word is None or b_word is None:
continue
if not hindi:
a_idx = a_word.order
b_idx = b_word.order
else:
a_idx = a_snt.index(a_word)+1
b_idx = b_snt.index(b_word)+1
# Make sure the gloss is in the
if a_glossed:
trans_idx = b_idx
lang_idx = a_idx
else:
trans_idx = a_idx
lang_idx = b_idx
a.add((trans_idx, lang_idx))
set_bilingual_alignment(inst, trans(inst), lang(inst), a, INTENT_ALN_MANUAL)
set_bilingual_alignment(inst, trans(inst), gloss(inst), a, INTENT_ALN_MANUAL)
xc.append(inst)
with open(out_path, 'w', encoding='utf-8') as f:
xigtxml.dump(f, xc)
def test_line_lengths(self):
inst = self.xc[1]
self.assertEqual(5, len(gloss(inst)))
self.assertEqual(6, len(lang(inst)))