def run_oie(lines, batch_size=1, debug=False):
"""
Run the OIE model and process the output.
"""
if debug:
logging.basicConfig(level = logging.DEBUG)
else:
logging.basicConfig(level = logging.INFO)
# Init OIE
#model = open_information_extraction_stanovsky_2018()
model = PretrainedModel('model_final.tar.gz',
'open-information-extraction')
model = model.predictor() # type: ignore
# process sentences
logging.info("Processing sentences")
oie_lines = []
for chunk in tqdm(chunks(lines, batch_size)):
oie_inputs = []
for sent in chunk:
oie_inputs.extend(create_instances(model, sent))
if not oie_inputs:
# No predicates in this sentence
continue
# Run oie on sents
sent_preds = model.predict_batch_json(oie_inputs)
# Collect outputs in batches
predictions_by_sent = defaultdict(list)
for outputs in sent_preds:
sent_tokens = outputs["words"]
tags = outputs["tags"]
sent_str = " ".join(sent_tokens)
assert(len(sent_tokens) == len(tags))
predictions_by_sent[sent_str].append((outputs["tags"], outputs["class_probabilities"]))
# Create extractions by sentence
for sent_tokens, predictions_for_sent in predictions_by_sent.items():
raw_tags = list(map(itemgetter(0), predictions_for_sent))
class_probs = list(map(itemgetter(1), predictions_for_sent))
# Compute confidence per extraction
confs = [get_confidence(model, tag_per_token, class_prob)
for tag_per_token, class_prob in zip(raw_tags, class_probs)]
extractions, tags = format_extractions([Mock_token(tok) for tok in sent_tokens.split(" ")], raw_tags)
oie_lines.extend([extraction + f"\t{conf}" for extraction, conf in zip(extractions, confs)])
logging.info("DONE")
return oie_lines
sent_str = " ".join(sent_tokens)
assert (len(sent_tokens) == len(tags))
predictions_by_sent[sent_str].append(
(outputs["tags"], outputs["class_probabilities"]))
# Create extractions by sentence
for sent_tokens, predictions_for_sent in predictions_by_sent.items():
raw_tags = list(map(itemgetter(0), predictions_for_sent))
class_probs = list(map(itemgetter(1), predictions_for_sent))
# Compute confidence per extraction
confs = [
get_confidence(model, tag_per_token, class_prob)
for tag_per_token, class_prob in zip(raw_tags, class_probs)
]
extractions, tags = format_extractions(
[Mock_token(tok) for tok in sent_tokens.split(" ")], raw_tags)
oie_lines.extend([
extraction + f"\t{conf}"
for extraction, conf in zip(extractions, confs)
])
t2 = time.perf_counter()
print("E2E time: ", t2 - t1)
# Write to file
logging.info(f"Writing output to {out_fn}")
with open(out_fn, "w", encoding="utf8") as fout:
fout.write("\n".join(oie_lines))
logging.info("DONE")
