def rank_by_semantic_shift_degree(model1, model2):
wv1 = model1.wv.vocab
wv2 = model2.wv.vocab
top_n = 1000
print(list(wv1.keys())[:10])
print("Computing intersections of model1 and model2")
intersection = list(
set(list(wv1.keys())[:top_n]) & set(list(wv2.keys())[:top_n]))
print("Computing the scores of semantic shifts")
intersection_scores = []
for word in intersection:
if "_NOUN" in word:
#rint("Procrustes aligner score: {} (from -1 to 1)".format(
# ProcrustesAligner(model1, model2).get_score(word)))
score = Jaccard(model1, model2, 50).get_score(word)
if score > 0.0:
intersection_scores.append(tuple((word, score)))
intersection_scores = sorted(intersection_scores, key=lambda tup: tup[1])
print("All done")
return intersection_scores
def get_algo_by_kind_and_two_models(kind: str, model1: KeyedVectors,
model2: KeyedVectors):
if kind.lower() == "global_anchors":
return GlobalAnchors(model1, model2)
elif kind.lower() == "jaccard":
return Jaccard(model1, model2, top_n_neighbors=50)
elif kind.lower() == "kendall_tau":
return KendallTau(model1, model2, top_n_neighbors=50)
def main():
parser = ArgumentParser()
parser.add_argument('--word',
'-w',
required=True,
help='word to be scored')
parser.add_argument('--model1',
'-m1',
required=True,
help='Path to the word embeddings model')
parser.add_argument('--model2',
'-m2',
required=True,
help='Path to the second word embedding model')
parser.add_argument(
'--top-n-neighbors',
'-n',
dest='topn',
default=50,
help='number of word neighbors to analyze '
'(optional, used in Kendall tau and Jaccard algo, default=50)')
args = parser.parse_args()
if not os.path.isfile(args.model1):
raise FileNotFoundError("File {} not found".format(args.model1))
if not os.path.isfile(args.model2):
raise FileNotFoundError("File {} not found".format(args.model2))
model1 = load_model(args.model1)
model2 = load_model(args.model2)
if args.word not in model1.vocab:
raise ValueError("Word {} not in {}".format(args.word, args.model1))
if args.word not in model2.vocab:
raise ValueError("Word {} not in {}".format(args.word, args.model2))
print("KendallTau score: {} (from -1 to 1)".format(
KendallTau(model1, model2,
top_n_neighbors=args.topn).get_score(args.word)))
print("Jaccard score: {} (from 0 to 1)".format(
Jaccard(model1, model2, args.topn).get_score(args.word)))
print("Global Anchors score: {} (from -1 to 1)".format(
GlobalAnchors(model1, model2).get_score(args.word)))
print("Procrustes aligner score: {} (from -1 to 1)".format(
ProcrustesAligner(model1, model2).get_score(args.word)))
ranks = rank_by_semantic_shift_degree(model1, model2)
list_of_pairs = [rank[0] + ':' + str(rank[1]) for rank in ranks[:100]]
print("\n".join(list_of_pairs))
def get_mean_dist_jaccard(wordlist, modellist, top_n_neighbors):
mean_scores = {}
for word in wordlist:
scores = 0
for i in range(len(modellist) - 1):
score = Jaccard(w2v1=modellist[i],
w2v2=modellist[i + 1],
top_n_neighbors=top_n_neighbors).get_score(word)
scores += (1 - score)
mean_scores[word] = scores / (len(modellist) - 1)
return mean_scores
def get_move_from_initial_jaccard(wordlist, modellist, top_n_neighbors):
move_from_init = {}
for word in wordlist:
deltas = 0
previous = Jaccard(modellist[0],
modellist[1],
top_n_neighbors=top_n_neighbors).get_score(word)
for i in range(2, len(modellist)):
similarity = Jaccard(
modellist[0], modellist[i],
top_n_neighbors=top_n_neighbors).get_score(word)
delta = similarity - previous
if delta > 0:
deltas -= 1
elif delta < 0:
deltas += 1
previous = similarity
move_from_init[word] = deltas / (len(modellist) - 2)
return move_from_init