def transform_senseval_contexts(inf: IO,
transform_tokens: Callable[[List[str]],
List[str]],
outf: IO) -> None:
def transform_context(context: etree.ElementBase) -> etree.ElementBase:
sent: List[str] = []
before = context.text
head_tag = context[0]
head = head_tag.text
after = head_tag.tail
before_tok = space_tokenize(before)
head_tok = space_tokenize(head)
after_tok = space_tokenize(after)
sent = before_tok + head_tok + after_tok
new_sent = transform_tokens(sent)
new_before = new_sent[:len(before_tok)]
new_head = new_sent[len(before_tok):len(before_tok) + len(head_tok)]
new_after = new_sent[len(before_tok) + len(head_tok):]
context.text = "\n" + "".join(tok + " " for tok in new_before)
head_tag.text = " ".join(tok for tok in new_head)
head_tag.tail = "".join(" " + tok for tok in new_after) + "\n"
return context
transform_blocks(eq_matcher("context"), inf, transform_context, outf)
def eurosense_reanchor(inf: IO, outf: IO):
"""
Reanchors Eurosense lemmas which are actually forms including some "light"
word like ei and olla by removing said unneccesary word.
"""
EXTRA_BITS = {"ei", "olla"}
fi2en, en2fi = get_en_fi_maps()
def ann_reanchor(ann):
all_lemma_names = []
for _, synset in iter_synsets(ann.text):
for lemma in synset.lemmas():
all_lemma_names.append(lemma.name())
if " " not in ann.attrib["lemma"]:
return
lem_begin, lem_rest = ann.attrib["lemma"].split(" ", 1)
if lem_begin not in EXTRA_BITS:
return
anchor_begin = ann.attrib["anchor"].split(" ", 1)[0]
for lemma_name in all_lemma_names:
if lemma_name.split("_", 1)[0] in (anchor_begin, lem_begin):
return
ann.attrib["lemma"] = lem_rest
ann.attrib["anchor"] = ann.attrib["anchor"].split(" ", 1)[1]
transform_blocks(eq_matcher("annotation"), inf, ann_reanchor, outf)
def lemma_to_synset(inf: IO, outf: IO):
from stiff.munge.utils import synset_id_of_ann
def l2ss(ann):
ann.text = pre_id_to_post(synset_id_of_ann(ann))
transform_blocks(eq_matcher("annotation"), inf, l2ss, outf)
def senseval_filter_lemma(lemmas, inf, outf, filter_key_out=None):
lemma_poses = pickle.load(lemmas)
filter_keys = set()
def filter_lexelt(lexelt):
if (str(lexelt.attrib["item"]),
lexelt.attrib["pos"]) not in lemma_poses:
return BYPASS
elif filter_key_out:
for instance in lexelt:
filter_keys.add(instance.attrib["id"])
transform_blocks(eq_matcher("lexelt"), inf, filter_lexelt, outf)
if filter_key_out:
pickle.dump(filter_keys, filter_key_out)
def senseval_rm_lemma(inf, outf, rm_key_out=None, lemmas=None):
lemmas = lemmas.split(",") if lemmas else []
rm_keys = set()
def filter_lexelt(lexelt):
if str(lexelt.attrib["item"]) in lemmas:
if rm_key_out:
for instance in lexelt:
rm_keys.add(instance.attrib["id"])
return BYPASS
transform_blocks(eq_matcher("lexelt"), inf, filter_lexelt, outf)
if rm_key_out:
pickle.dump(rm_keys, rm_key_out)
def stiff_select_wn(inf: IO, outf: IO, wn):
from stiff.munge.utils import langs_of_wns
selected_wns = set(wn)
selected_langs = langs_of_wns(selected_wns)
def filter_wns(wns):
return [wn for wn in wns if wn in selected_wns]
def select_wn(ann):
# annotation[wordnets]
ann_wns = ann.attrib["wordnets"].split()
common_wns = filter_wns(ann_wns)
if not len(common_wns):
return BYPASS
ann.attrib["wordnets"] = " ".join(common_wns)
# annotation[wnlemma]
wnlemma_bits = ann.attrib["wnlemma"].split(" ")
new_wmlemmas_bits = []
for wnlemma in wnlemma_bits:
wnlemma_dict = parse_qs_single(wnlemma)
wnlemma_wns = wnlemma_dict["wn"].split(",")
common_wns = filter_wns(wnlemma_wns)
if not common_wns:
continue
wnlemma_dict["wn"] = ",".join(common_wns)
new_wmlemmas_bits.append(urlencode(wnlemma_dict))
ann.attrib["wnlemma"] = " ".join(new_wmlemmas_bits)
# annotation > #text
ann_langs = langs_of_wns(ann_wns)
if len(ann_langs) <= len(selected_langs):
return
lemmas_str = ann.text
bits = lemmas_str.split(" ")
assert len(bits) <= 2
if len(bits) <= 1:
return
if "eng" in selected_langs:
ann.text = bits[0]
else:
ann.text = bits[1]
transform_blocks(eq_matcher("annotation"), inf, select_wn, outf)
def babelnet_lookup(inf: IO, map_bn2wn: IO, outf: IO):
"""
This stage converts BabelNet ids to WordNet ids.
"""
bn2wn_map: Dict[str, Set[str]] = {}
for line in map_bn2wn:
bn, wn_full = line[:-1].split("\t")
wn_off = wn_full.split(":", 1)[1]
bn2wn_map.setdefault(bn, set()).add(wn_off)
def ann_bn2wn(ann):
if ann.text not in bn2wn_map:
return BYPASS
wn_ids = bn2wn_map[ann.text]
bits = []
for wn_id in wn_ids:
off, pos = wn_id[:-1], wn_id[-1]
bits.append("{}-{}".format(off, pos))
ann.text = " ".join(bits)
transform_blocks(eq_matcher("annotation"), inf, ann_bn2wn, outf)
def senseval_gather(indir: str, outf: IO, keyout: IO):
"""
Gather individual per-word SenseEval files into one big file, usable by
ItMakesSense and Context2Vec.
"""
with lexical_sample(outf):
for word_dir in listdir(indir):
train_fn = pjoin(indir, word_dir, "train.xml")
key_fn = pjoin(indir, word_dir, "train.key")
with open(train_fn, "rb") as train_f:
stream = etree.iterparse(train_f, events=("start", "end"))
def cb(lexelt):
if not len(lexelt):
return
outf.write(etree.tostring(lexelt, encoding="unicode"))
chunk_cb(stream, eq_matcher("lexelt"), cb)
with open(key_fn) as key_f:
keyout.write(key_f.read())
def senseval_select_lemma(inf, keyin, outf, keyout, lemma_pos):
if "." in lemma_pos:
lemma, pos = lemma_pos.rsplit(".", 1)
else:
lemma = lemma_pos
pos = None
keys = set()
def filter_lexelt(lexelt):
if lexelt.attrib["item"] != lemma:
return BYPASS
if pos and lexelt.attrib["pos"] != pos:
return BYPASS
for instance in lexelt:
keys.add(instance.attrib["id"])
transform_blocks(eq_matcher("lexelt"), inf, filter_lexelt, outf)
for line in keyin:
if line.split(" ", 1)[0] not in keys:
continue
keyout.write(line)
def eurosense_fix_lemmas(inf: IO, outf: IO, keep_unknown: bool, quiet: bool):
"""
Eurosense contains many lemmas which are not in the set of lemmas for the
synset in FinnWordNet. There are two reasons this might occur.
Scenario A) Bad lemmatisation by Babelfy. In this case we can try and
recover the correct lemma by lemmatising ourself and combining with
information from WordNet
Scenario B) Extra lemmas have been associated with the WordNet synset in
BabelNet. In this case there's nothing to do, and we should usually just
drop the annotation.
"""
fi2en, en2fi = get_en_fi_maps()
def ann_fix_lemmas(ann):
# 1) check if their lemmatisation matches something in FiWN as is
orig_lemma_str = ann.attrib["lemma"]
orig_lemma_str = orig_lemma_str.replace("#", "").replace(" ", "_")
def mk_lemma_synset_map(lower=False):
lemma_synset_map = {}
for synset_id, synset in iter_synsets(ann.text):
for lemma in synset.lemmas():
lemma_str = lemma.name()
if lower:
lemma_str = lemma_str.lower()
lemma_synset_map.setdefault(lemma_str,
set()).add(synset_id)
return lemma_synset_map
lemma_synset_map = mk_lemma_synset_map()
if orig_lemma_str in lemma_synset_map:
ann.text = " ".join(lemma_synset_map[orig_lemma_str])
ann.attrib["lemma"] = orig_lemma_str
return
# 2) Try and just use the surface as is as the lemma
lemmatised_anchor = ann.attrib["anchor"].replace(" ", "_")
lemma_synset_map_lower = mk_lemma_synset_map(lower=True)
if lemmatised_anchor.lower() in lemma_synset_map_lower:
ann.text = " ".join(
lemma_synset_map_lower[lemmatised_anchor.lower()])
# XXX: Should be lemma in original case rather than anchor in original case
ann.attrib["lemma"] = lemmatised_anchor
return
# 3) Re-lemmatise the surface using OMorFi and try and match with FiWN
anchor_bits = ann.attrib["anchor"].split(" ")
matches = {}
for lemma_str, synset_id in lemma_synset_map.items():
lemma_bits = lemma_str.split("_")
common = lemma_intersect(anchor_bits, lemma_bits)
if common is not None:
matches.setdefault(lemma_str, set()).update(synset_id)
if len(matches) == 1:
lemma, synsets = next(iter(matches.items()))
ann.attrib["lemma"] = lemma
ann.text = " ".join(synsets)
return
elif len(matches) > 1:
if not quiet:
sys.stderr.write(
"Multiple lemmas found found for {}: {}\n".format(
ann.attrib["anchor"], matches))
# If nothing has worked, it's probably scenario B as above
elif len(matches) == 0:
if not quiet:
sys.stderr.write("No lemma found for {} {} {}\n".format(
ann.text, orig_lemma_str, lemmatised_anchor))
if keep_unknown:
ann.attrib["lemma"] = orig_lemma_str
else:
return BYPASS
transform_blocks(eq_matcher("annotation"), inf, ann_fix_lemmas, outf)
