def main() -> None:
args = parse_arguments(subtype="evaluate_shallow_metrics")
# get verbosity
if args.verbosity == 1:
logger = logging.getLogger('base')
else:
logger = logging.getLogger('root')
# define json glob
json_glob = args.json_glob
# define search space
files = glob(json_glob)
for input_file in files:
# log information
logger.info("Computing bleu and chrf scores: %s", input_file)
# load single dictionary and compute surface similarity scores
with open(input_file, "r") as f:
store = json.load(f)
for key in tqdm(store.keys()):
source_orig_de = store[key]["sentence_original"]["source"]
source_para_de = store[key]["sentence_paraphrase"]["source"]
target_orig_en = store[key]["sentence_original"]["target"]
target_para_en = store[key]["sentence_paraphrase"]["target"]
chrf_bar_source = (sacrebleu.sentence_chrf(
source_orig_de, [source_para_de]).score +
sacrebleu.sentence_chrf(
source_para_de, [source_orig_de]).score) / 2
chrf_bar_target = (sacrebleu.sentence_chrf(
target_orig_en, [target_para_en]).score +
sacrebleu.sentence_chrf(
target_para_en, [target_orig_en]).score) / 2
bleu_bar_source = (sacrebleu.sentence_bleu(
source_orig_de, [source_para_de]).score +
sacrebleu.sentence_bleu(
source_para_de, [source_orig_de]).score) / 2
bleu_bar_target = (sacrebleu.sentence_bleu(
target_orig_en, [target_para_en]).score +
sacrebleu.sentence_bleu(
target_para_en, [target_orig_en]).score) / 2
store[key]["chrf_bar_source"] = chrf_bar_source
store[key]["chrf_bar_target"] = chrf_bar_target
store[key]["bleu_bar_source"] = bleu_bar_source
store[key]["bleu_bar_target"] = bleu_bar_target
store[key]["chrf_bar_mean"] = (chrf_bar_source +
chrf_bar_target) / 2
store[key]["bleu_bar_mean"] = (bleu_bar_source +
bleu_bar_target) / 2
# write back json to disk
with open(input_file, "w") as f:
store = json.dump(store, f, ensure_ascii=False)
def evaluate_example(self, summary, reference):
score = sacrebleu.sentence_chrf(summary,
reference,
order=self.ncorder,
beta=self.beta)
score_dict = {"chrf": score.score}
return score_dict
def evaluate_example(self, summary, reference):
if not isinstance(reference, list):
reference = [reference]
score = sacrebleu.sentence_chrf(summary,
reference,
order=self.ncorder,
beta=self.beta)
score_dict = {"chrf": score.score}
return score_dict
def run_sentence_chrf(candidates: list, references: list) -> list:
""" Runs sentence chrF from Sacrebleu. """
assert len(candidates) == len(references)
chrf_scores = []
for i in tqdm(range(len(candidates)), desc="Running chrF..."):
chrf_scores.append(
sentence_chrf(hypothesis=candidates[i],
reference=references[i]).score)
return chrf_scores
def score_sentence(self, hyp, ref, lang=None):
return sacrebleu.sentence_chrf(hyp, ref)
def _analyze(
self,
inputs_list: List[lit_types.JsonDict]) -> List[lit_types.JsonDict]:
features_list: List[str] = list(
map(lambda input: encode_sp(self.src_spp, input["src_text"]),
inputs_list))
infer_config: dict = self.config["infer"]
dataset = make_inference_dataset(
self.model,
features_list,
infer_config["batch_size"],
length_bucket_width=infer_config["length_bucket_width"],
prefetch_buffer_size=infer_config.get("prefetch_buffer_size"),
)
if self._analyze_fn is None:
self._analyze_fn = tf.function(
self.model.analyze, input_signature=(dataset.element_spec, ))
if not tf.config.functions_run_eagerly():
tf.get_logger().info(
"Tracing and optimizing the analyze graph...")
self._analyze_fn.get_concrete_function(
) # Trace the function now.
results: List[lit_types.JsonDict] = [None] * len(features_list)
for features in dataset:
predictions = self._analyze_fn(features)
top_k_probs, top_k_ids = tf.nn.top_k(tf.nn.softmax(
predictions["logits"]),
k=10)
del predictions["logits"]
predictions["top_k_probs"] = top_k_probs
predictions["top_k_ids"] = top_k_ids
masks = tf.sequence_mask(features["length"],
maxlen=tf.shape(features["ids"])[1])
predictions["encoder_final_embedding"] = masked_token_mean(
predictions["encoder_outputs"], masks)
del predictions["encoder_outputs"]
predictions = tf.nest.map_structure(lambda t: t.numpy(),
predictions)
for prediction in extract_batches(predictions):
index: int = prediction["index"]
target_length = prediction["length"]
trg_tokens = prediction["tokens"][:target_length]
tok_trg_text = self.model.labels_inputter.tokenizer.detokenize(
trg_tokens)
trg_text = decode_sp(tok_trg_text)
attention = prediction["alignment"][:target_length]
probs = prediction["top_k_probs"]
ids = prediction["top_k_ids"]
pred_tokens = list(
self._convert_top_k(ids, probs, target_length))
encoder_final_embedding = prediction["encoder_final_embedding"]
ref_text = inputs_list[index]["ref_text"]
tok_ref_text = encode_sp(self.trg_spp, ref_text)
ter_score = sacrebleu.sentence_ter(tok_trg_text,
[tok_ref_text])
chrf_score = sacrebleu.sentence_chrf(trg_text, [ref_text],
order=3)
results[index] = {
"trg_tokens": [t.decode("utf-8") for t in trg_tokens],
"trg_text": trg_text,
"attention": np.expand_dims(attention, axis=0),
"src_tokens": features_list[index].split(),
"pred_tokens": pred_tokens,
"encoder_final_embedding": encoder_final_embedding,
"ter": ter_score.score,
"chrf3": chrf_score.score,
}
return results
prev = [' '.join(o) for o in pred_sys_stream]
# choose one (gold or pred) and postprocess
sys_stream = pred_sys_stream
sys_stream = [
pp.post_process(o, abstract[i], graph[i])
for i, o in enumerate(sys_stream)
]
bleu = sacrebleu.corpus_bleu(sys_stream,
ref_streams,
force=True,
lowercase=True,
tokenize='none').score
chrf = sacrebleu.corpus_chrf(sys_stream, ref_stream)
all_sent_chrf = [
sacrebleu.sentence_chrf(x, y) for x, y in zip(sys_stream, ref_stream)
]
avg_sent_chrf = sum(all_sent_chrf) / len(all_sent_chrf)
if args.output:
with open(args.pred_file + '.final', 'w') as fo:
for x in sys_stream:
fo.write(x + '\n')
with open(args.pred_file + '.ref', 'w') as fo:
for x in ref_stream:
fo.write(x + '\n')
print(avg_sent_chrf)
print(bleu, chrf)
def _get_sent_chrf(hypothesis: List[str],
references: List[List[str]],
extra_args: Optional[Dict[str, str]] = None):
return [
sb.sentence_chrf(h, r).score for h, r in zip(hypothesis, references[0])
]
def compute_chrf(references, translation):
hypo = ' '.join(translation)
refs = [' '.join(r) for r in references][0]
return sacrebleu.sentence_chrf(hypo, refs).score
def chrf(self):
"""
ChRF of the hypothesis
"""
return sacrebleu.sentence_chrf(self.new_hyp.replace(' ', '').replace('</s>', '').replace('▁', ' '), [self.cur_ref]).score
def test_chrf_sentence_level(hypothesis, references, expected_score):
score = sacrebleu.sentence_chrf(hypothesis, references,
eps_smoothing=True).score
assert abs(score - expected_score) < EPSILON
def sentchrf(y_hat, y):
return sentence_chrf(y_hat, y) * 100
if '<generate>' in hyp2[i]:
hyp2[i] = hyp2[i].split('<generate>')[-1]
hyp2[i] = tokenize_sentence(hyp2[i].lower())
# Run evaluation
print("BLEU hyp1", round(raw_corpus_bleu(hyp1, [ref]), 2))
print("chrF++ hyp1", round(raw_corpus_chrf(hyp1, ref).score * 100, 2))
print("BLEU hyp2", round(raw_corpus_bleu(hyp2, [ref]), 2))
print("chrF++ hyp2", round(raw_corpus_chrf(hyp2, ref).score * 100, 2))
h1_sent_scores = list()
h2_sent_scores = list()
for r, h1, h2 in zip(ref, hyp1, hyp2):
h1_sent_scores.append(sacrebleu.sentence_bleu(h1, r).score)
h2_sent_scores.append(sacrebleu.sentence_bleu(h2, r).score)
t = ttest_ind(h1_sent_scores, h2_sent_scores)
print("BLEU P value:", "{:.20f}".format(t[1]))
h1_sent_scores = list()
h2_sent_scores = list()
for r, h1, h2 in zip(ref, hyp1, hyp2):
h1_sent_scores.append(sacrebleu.sentence_chrf(h1, r).score)
h2_sent_scores.append(sacrebleu.sentence_chrf(h2, r).score)
t = ttest_ind(h1_sent_scores, h2_sent_scores)
print("ChrF++ P value:", "{:.20f}".format(t[1]))
dataframe = pd.DataFrame(columns=columns)
for i in range(len(source)):
BLEU_mmt_b = float("{:.3f}".format(
sentence_bleu(base_MMT[i], [ref[i]], smooth_method='exp').score))
BLEU_mmt_c_I = float("{:.3f}".format(
sentence_bleu(mmt_c_I[i], [ref[i]], smooth_method='exp').score))
BLEU_mmt_c_II = float("{:.3f}".format(
sentence_bleu(mmt_c_II[i], [ref[i]], smooth_method='exp').score))
diff_BLEU_I = float(
"{:.3f}".format(float(BLEU_mmt_c_I) - float(BLEU_mmt_b)))
diff_BLEU_II = float(
"{:.3f}".format(float(BLEU_mmt_c_II) - float(BLEU_mmt_b)))
diff_BLEU_I_II = float(
"{:.3f}".format(float(BLEU_mmt_c_II) - float(BLEU_mmt_c_I)))
chrF3_mmt_b = float("{:.3f}".format(
sentence_chrf(base_MMT[i], [ref[i]]).score))
chrF3_mmt_c_I = float("{:.3f}".format(
sentence_chrf(mmt_c_I[i], [ref[i]], beta=3).score))
chrF3_mmt_c_II = float("{:.3f}".format(
sentence_chrf(mmt_c_II[i], [ref[i]], beta=3).score))
diff_chrF3_I = float("{:.3f}".format(chrF3_mmt_c_I - chrF3_mmt_b))
diff_chrF3_II = float("{:.3f}".format(chrF3_mmt_c_II - chrF3_mmt_b))
diff_chrF3_I_II = float("{:.3f}".format(chrF3_mmt_c_II - chrF3_mmt_c_I))
hlepor_mmt_b = float("{:.3f}".format(
single_hlepor_score(ref[i], base_MMT[i])))
hlepor_mmt_c_I = float("{:.3f}".format(
single_hlepor_score(ref[i], mmt_c_I[i])))
hlepor_mmt_c_II = float("{:.3f}".format(
single_hlepor_score(ref[i], mmt_c_II[i])))
diff_hlepor_I = float("{:.3f}".format(hlepor_mmt_c_I - hlepor_mmt_b))
diff_hlepor_II = float("{:.3f}".format(hlepor_mmt_c_II - hlepor_mmt_b))
