|
import os |
|
import json |
|
import pandas as pd |
|
from statistics import mean |
|
from datasets import load_dataset |
|
from relbert import RelBERT |
|
|
|
|
|
def cosine_similarity(a, b): |
|
norm_a = sum(map(lambda x: x * x, a)) ** 0.5 |
|
norm_b = sum(map(lambda x: x * x, b)) ** 0.5 |
|
return sum(map(lambda x: x[0] * x[1], zip(a, b))) / (norm_a * norm_b) |
|
|
|
|
|
|
|
data = load_dataset("cardiffnlp/relentless", split="test") |
|
full_result = [] |
|
|
|
for lm in ['base', 'large']: |
|
os.makedirs(f"results/relbert/relbert-roberta-{lm}", exist_ok=True) |
|
scorer = None |
|
for d in data: |
|
ppl_file = f"results/relbert/relbert-roberta-{lm}/ppl.{d['relation_type'].replace(' ', '_').replace('/', '__')}.jsonl" |
|
anchor_embeddings = [(a, b) for a, b in d['prototypical_examples']] |
|
option_embeddings = [(x, y) for x, y in d['pairs']] |
|
|
|
if not os.path.exists(ppl_file): |
|
|
|
if scorer is None: |
|
scorer = RelBERT(f"relbert/relbert-roberta-{lm}") |
|
anchor_embeddings = scorer.get_embedding(d['prototypical_examples']) |
|
option_embeddings = scorer.get_embedding(d['pairs'], batch_size=64) |
|
similarity = [[cosine_similarity(a, b) for b in anchor_embeddings] for a in option_embeddings] |
|
output = [{"similarity": s} for s in similarity] |
|
with open(ppl_file, "w") as f: |
|
f.write("\n".join([json.dumps(i) for i in output])) |
|
|
|
with open(ppl_file) as f: |
|
similarity = [json.loads(i)['similarity'] for i in f.read().split("\n") if len(i) > 0] |
|
|
|
true_rank = d['ranks'] |
|
assert len(true_rank) == len(similarity), f"Mismatch in number of examples: {len(true_rank)} vs {len(similarity)}" |
|
prediction = [max(s) for s in similarity] |
|
rank_map = {p: n for n, p in enumerate(sorted(prediction, reverse=True), 1)} |
|
prediction_max = [rank_map[p] for p in prediction] |
|
|
|
prediction = [min(s) for s in similarity] |
|
rank_map = {p: n for n, p in enumerate(sorted(prediction, reverse=True), 1)} |
|
prediction_min = [rank_map[p] for p in prediction] |
|
|
|
prediction = [mean(s) for s in similarity] |
|
rank_map = {p: n for n, p in enumerate(sorted(prediction, reverse=True), 1)} |
|
prediction_mean = [rank_map[p] for p in prediction] |
|
|
|
tmp = pd.DataFrame([true_rank, prediction_max, prediction_min, prediction_mean]).T |
|
cor_max = tmp.corr("spearman").values[0, 1] |
|
cor_min = tmp.corr("spearman").values[0, 2] |
|
cor_mean = tmp.corr("spearman").values[0, 3] |
|
full_result.append({"model": f"RelBERT\textsubscript{'{'}{lm.upper()}{'}'}", "relation_type": d['relation_type'], "correlation": cor_max}) |
|
|
|
df = pd.DataFrame(full_result) |
|
df = df.pivot(columns="relation_type", index="model", values="correlation") |
|
df['average'] = df.mean(1) |
|
df.to_csv("results/relbert/relbert.csv") |
|
df = (100 * df).round() |
|
print(df.to_markdown()) |
|
print(df.to_latex()) |