요약

요약 (Summarization)

요약은 문서나 기사에서 중요한 정보를 모두 담고 있는 더 짧은 버전을 만드는 작업이에요. 번역과 함께 시퀀스-투-시퀀스(sequence-to-sequence) 작업으로 다룰 수 있는 또 다른 예시죠. 요약에는 두 가지 방식이 있어요:

출처: 문서

본문

추출적(Extractive): 문서에서 가장 관련성 높은 정보를 추출.
생성적(Abstractive): 가장 관련성 높은 정보를 담은 새 텍스트를 생성.

이 가이드에서는 다음을 배워요:

BillSum 데이터셋의 캘리포니아 주 법안 하위 집합에서 T5를 파인튜닝해서 생성적 요약 수행하기.
파인튜닝한 모델을 추론(inference)에 사용하기.

이 작업과 호환되는 모든 아키텍처와 체크포인트를 보려면 task-page를 확인하는 걸 추천해요.

시작하기 전에 필요한 라이브러리를 모두 설치했는지 확인해요:

pip install transformers datasets evaluate rouge_score

Hugging Face 계정에 로그인해서 모델을 커뮤니티에 업로드하고 공유하는 걸 권장해요. 로그인하라는 메시지가 나오면 토큰을 입력해서 로그인하세요:

>>> from huggingface_hub import notebook_login

>>> notebook_login()

BillSum 데이터셋 로드

먼저 🤗 Datasets 라이브러리에서 BillSum 데이터셋의 작은 캘리포니아 주 법안 하위 집합을 로드해요:

>>> from datasets import load_dataset

>>> billsum = load_dataset("billsum", split="ca_test")

train_test_split 메서드로 데이터셋을 훈련용과 테스트용으로 나눠요:

>>> billsum = billsum.train_test_split(test_size=0.2)

그다음 예시 하나를 살펴봐요:

>>> billsum["train"][0]
{'summary': 'Existing law authorizes state agencies to enter into contracts for the acquisition of goods or services upon approval by the Department of General Services. Existing law sets forth various requirements and prohibitions for those contracts, including, but not limited to, a prohibition on entering into contracts for the acquisition of goods or services of $100,000 or more with a contractor that discriminates between spouses and domestic partners or same-sex and different-sex couples in the provision of benefits. Existing law provides that a contract entered into in violation of those requirements and prohibitions is void and authorizes the state or any person acting on behalf of the state to bring a civil action seeking a determination that a contract is in violation and therefore void. Under existing law, a willful violation of those requirements and prohibitions is a misdemeanor.\nThis bill would also prohibit a state agency from entering into contracts for the acquisition of goods or services of $100,000 or more with a contractor that discriminates between employees on the basis of gender identity in the provision of benefits, as specified. By expanding the scope of a crime, this bill would impose a state-mandated local program.\nThe California Constitution requires the state to reimburse local agencies and school districts for certain costs mandated by the state. Statutory provisions establish procedures for making that reimbursement.\nThis bill would provide that no reimbursement is required by this act for a specified reason.',
 'text': 'The people of the State of California do enact as follows:\n\n\nSECTION 1.\nSection 10295.35 is added to the Public Contract Code, to read:\n10295.35.\n(a) (1) Notwithstanding any other law, a state agency shall not enter into any contract for the acquisition of goods or services in the amount of one hundred thousand dollars ($100,000) or more with a contractor that, in the provision of benefits, discriminates between employees on the basis of an employee’s or dependent’s actual or perceived gender identity, including, but not limited to, the employee’s or dependent’s identification as transgender.\n(2) For purposes of this section, “contract” includes contracts with a cumulative amount of one hundred thousand dollars ($100,000) or more per contractor in each fiscal year.\n(3) For purposes of this section, an employee health plan is discriminatory if the plan is not consistent with Section 1365.5 of the Health and Safety Code and Section 10140 of the Insurance Code.\n(4) The requirements of this section shall apply only to those portions of a contractor’s operations that occur under any of the following conditions:\n(A) Within the state.\n(B) On real property outside the state if the property is owned by the state or if the state has a right to occupy the property, and if the contractor’s presence at that location is connected to a contract with the state.\n(C) Elsewhere in the United States where work related to a state contract is being performed.\n(b) Contractors shall treat as confidential, to the maximum extent allowed by law or by the requirement of the contractor’s insurance provider, any request by an employee or applicant for employment benefits or any documentation of eligibility for benefits submitted by an employee or applicant for employment.\n(c) After taking all reasonable measures to find a contractor that complies with this section, as determined by the state agency, the requirements of this section may be waived under any of the following circumstances:\n(1) There is only one prospective contractor willing to enter into a specific contract with the state agency.\n(2) The contract is necessary to respond to an emergency, as determined by the state agency, that endangers the public health, welfare, or safety, or the contract is necessary for the provision of essential services, and no entity that complies with the requirements of this section capable of responding to the emergency is immediately available.\n(3) The requirements of this section violate, or are inconsistent with, the terms or conditions of a grant, subvention, or agreement, if the agency has made a good faith attempt to change the terms or conditions of any grant, subvention, or agreement to authorize application of this section.\n(4) The contractor is providing wholesale or bulk water, power, or natural gas, the conveyance or transmission of the same, or ancillary services, as required for ensuring reliable services in accordance with good utility practice, if the purchase of the same cannot practically be accomplished through the standard competitive bidding procedures and the contractor is not providing direct retail services to end users.\n(d) (1) A contractor shall not be deemed to discriminate in the provision of benefits if the contractor, in providing the benefits, pays the actual costs incurred in obtaining the benefit.\n(2) If a contractor is unable to provide a certain benefit, despite taking reasonable measures to do so, the contractor shall not be deemed to discriminate in the provision of benefits.\n(e) (1) Every contract subject to this chapter shall contain a statement by which the contractor certifies that the contractor is in compliance with this section.\n(2) The department or other contracting agency shall enforce this section pursuant to its existing enforcement powers.\n(3) (A) If a contractor falsely certifies that it is in compliance with this section, the contract with that contractor shall be subject to Article 9 (commencing with Section 10420), unless, within a time period specified by the department or other contracting agency, the contractor provides to the department or agency proof that it has complied, or is in the process of complying, with this section.\n(B) The application of the remedies or penalties contained in Article 9 (commencing with Section 10420) to a contract subject to this chapter shall not preclude the application of any existing remedies otherwise available to the department or other contracting agency under its existing enforcement powers.\n(f) Nothing in this section is intended to regulate the contracting practices of any local jurisdiction.\n(g) This section shall be construed so as not to conflict with applicable federal laws, rules, or regulations. In the event that a court or agency of competent jurisdiction holds that federal law, rule, or regulation invalidates any clause, sentence, paragraph, or section of this code or the application thereof to any person or circumstances, it is the intent of the state that the court or agency sever that clause, sentence, paragraph, or section so that the remainder of this section shall remain in effect.\nSEC. 2.\nSection 10295.35 of the Public Contract Code shall not be construed to create any new enforcement authority or responsibility in the Department of General Services or any other contracting agency.\nSEC. 3.\nNo reimbursement is required by this act pursuant to Section 6 of Article XIII B of the California Constitution because the only costs that may be incurred by a local agency or school district will be incurred because this act creates a new crime or infraction, eliminates a crime or infraction, or changes the penalty for a crime or infraction, within the meaning of Section 17556 of the Government Code, or changes the definition of a crime within the meaning of Section 6 of Article XIII B of the California Constitution.',
 'title': 'An act to add Section 10295.35 to the Public Contract Code, relating to public contracts.'}

사용할 두 가지 필드가 있어요:

text: 법안 본문으로, 모델의 입력이 돼요.
summary: text를 간결하게 압축한 버전으로, 모델의 목표(target)가 돼요.

전처리

다음 단계는 text와 summary를 처리할 T5 토크나이저를 로드하는 거예요:

>>> from transformers import AutoTokenizer

>>> checkpoint = "google-t5/t5-small"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)

만들려는 전처리 함수는 다음을 해야 해요:

입력 앞에 프롬프트를 붙여서 T5가 이게 요약 작업임을 알게 해요. 여러 NLP 작업을 처리할 수 있는 일부 모델은 특정 작업을 위해 프롬프트가 필요해요.
라벨을 토크나이즈할 때 text_target 인자를 사용해요.
시퀀스가 max_length 파라미터로 설정한 최대 길이를 넘지 않도록 자르는 것.

>>> prefix = "summarize: "

>>> def preprocess_function(examples):
...     inputs = [prefix + doc for doc in examples["text"]]
...     model_inputs = tokenizer(inputs, max_length=1024, truncation=True)

...     labels = tokenizer(text_target=examples["summary"], max_length=128, truncation=True)

...     model_inputs["labels"] = labels["input_ids"]
...     return model_inputs

전처리 함수를 전체 데이터셋에 적용하려면 🤗 Datasets의 map 메서드를 사용해요. batched=True를 설정하면 데이터셋의 여러 요소를 한 번에 처리할 수 있어서 map 함수를 빠르게 할 수 있어요:

>>> tokenized_billsum = billsum.map(preprocess_function, batched=True)

이제 DataCollatorForSeq2Seq를 사용해 예시 배치를 만들어요. 콜레이션(collation) 과정에서 데이터셋 전체를 최대 길이로 패딩하는 것보다 배치 안의 문장들을 가장 긴 길이에 맞춰 동적으로 패딩하는 편이 더 효율적이에요.

>>> from transformers import DataCollatorForSeq2Seq

>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint)

평가

학습 중에 메트릭을 포함하면 모델의 성능을 평가하는 데 종종 도움이 돼요. 🤗 Evaluate 라이브러리로 평가 방법을 빠르게 로드할 수 있어요. 이 작업에서는 ROUGE 메트릭을 로드해요 (메트릭을 로드하고 계산하는 방법은 🤗 Evaluate quick tour를 참고하세요):

>>> import evaluate

>>> rouge = evaluate.load("rouge")

그다음 예측값과 라벨을 compute에 넘겨 ROUGE 메트릭을 계산하는 함수를 만들어요:

>>> import numpy as np

>>> def compute_metrics(eval_pred):
...     predictions, labels = eval_pred
...     decoded_preds = tokenizer.batch_decode(predictions, skip_special_tokens=True)
...     labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
...     decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

...     result = rouge.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)

...     prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in predictions]
...     result["gen_len"] = np.mean(prediction_lens)

...     return {k: round(v, 4) for k, v in result.items()}

이제 compute_metrics 함수를 쓸 준비가 됐어요. 학습을 설정할 때 다시 사용하게 될 거예요.

학습

Trainer로 모델을 파인튜닝하는 방법에 익숙하지 않다면 여기의 기본 튜토리얼을 확인해 보세요!

이제 모델 학습을 시작할 준비가 됐어요! T5를 AutoModelForSeq2SeqLM으로 로드해요:

>>> from transformers import AutoModelForSeq2SeqLM, Seq2SeqTrainingArguments, Seq2SeqTrainer

>>> model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)

이 시점에서 남은 단계는 세 가지뿐이에요:

Seq2SeqTrainingArguments에서 학습 하이퍼파라미터를 정의해요. 유일하게 필수인 파라미터는 모델을 저장할 위치를 지정하는 output_dir이에요. push_to_hub=True로 설정하면 모델을 Hub에 푸시할 수 있어요 (모델을 업로드하려면 Hugging Face에 로그인해야 해요). 각 에폭이 끝날 때마다 Trainer가 ROUGE 메트릭을 평가하고 학습 체크포인트를 저장해요.
학습 인수를 Seq2SeqTrainer에 모델, 데이터셋, 토크나이저, 데이터 콜레이터, compute_metrics 함수와 함께 전달해요.
train()을 호출해서 모델을 파인튜닝해요.

>>> training_args = Seq2SeqTrainingArguments(
...     output_dir="my_awesome_billsum_model",
...     eval_strategy="epoch",
...     learning_rate=2e-5,
...     per_device_train_batch_size=16,
...     per_device_eval_batch_size=16,
...     weight_decay=0.01,
...     save_total_limit=3,
...     num_train_epochs=4,
...     predict_with_generate=True,
...     fp16=True, #change to bf16=True for XPU
...     push_to_hub=True,
... )

>>> trainer = Seq2SeqTrainer(
...     model=model,
...     args=training_args,
...     train_dataset=tokenized_billsum["train"],
...     eval_dataset=tokenized_billsum["test"],
...     processing_class=tokenizer,
...     data_collator=data_collator,
...     compute_metrics=compute_metrics,
... )

>>> trainer.train()

학습이 완료되면 push_to_hub() 메서드로 모델을 Hub에 공유해서 모두가 쓸 수 있게 해요:

>>> trainer.push_to_hub()

요약용으로 모델을 파인튜닝하는 더 심층적인 예시는 해당 PyTorch notebook을 참고하세요.

추론 (Inference)

좋아요, 이제 모델을 파인튜닝했으니 추론에 활용할 수 있어요!

요약하고 싶은 텍스트를 하나 떠올려요. T5의 경우 작업에 따라 입력에 프롬프트를 붙여야 해요. 요약 작업에서는 아래처럼 입력에 프롬프트를 붙이세요:

>>> text = "summarize: The Inflation Reduction Act lowers prescription drug costs, health care costs, and energy costs. It's the most aggressive action on tackling the climate crisis in American history, which will lift up American workers and create good-paying, union jobs across the country. It'll lower the deficit and ask the ultra-wealthy and corporations to pay their fair share. And no one making under $400,000 per year will pay a penny more in taxes."

텍스트를 토크나이즈하고 input_ids를 PyTorch 텐서로 반환해요:

>>> from transformers import AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_billsum_model")
>>> inputs = tokenizer(text, return_tensors="pt").input_ids

generate() 메서드를 사용해 요약을 만드세요. 다양한 텍스트 생성 전략과 생성을 제어하는 파라미터에 대한 자세한 내용은 Text Generation API를 확인해 보세요.

>>> from transformers import AutoModelForSeq2SeqLM

>>> model = AutoModelForSeq2SeqLM.from_pretrained("username/my_awesome_billsum_model")
>>> outputs = model.generate(inputs, max_new_tokens=100, do_sample=False)

생성된 토큰 id를 다시 텍스트로 디코딩해요:

>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'the inflation reduction act lowers prescription drug costs, health care costs, and energy costs. it's the most aggressive action on tackling the climate crisis in american history. it will ask the ultra-wealthy and corporations to pay their fair share.'