β–Ί Code examples / Natural Language Processing / Abstractive Summarization with Hugging Face Transformers

Abstractive Summarization with Hugging Face Transformers

Author: Sreyan Ghosh
Date created: 2022/07/04
Last modified: 2022/08/28
Description: Training T5 using Hugging Face Transformers for Abstractive Summarization.

β“˜ This example uses Keras 2

View in Colab β€’ GitHub source


Introduction

Automatic summarization is one of the central problems in Natural Language Processing (NLP). It poses several challenges relating to language understanding (e.g. identifying important content) and generation (e.g. aggregating and rewording the identified content into a summary).

In this tutorial, we tackle the single-document summarization task with an abstractive modeling approach. The primary idea here is to generate a short, single-sentence news summary answering the question β€œWhat is the news article about?”. This approach to summarization is also known as Abstractive Summarization and has seen growing interest among researchers in various disciplines.

Following prior work, we aim to tackle this problem using a sequence-to-sequence model. Text-to-Text Transfer Transformer (T5) is a Transformer-based model built on the encoder-decoder architecture, pretrained on a multi-task mixture of unsupervised and supervised tasks where each task is converted into a text-to-text format. T5 shows impressive results in a variety of sequence-to-sequence (sequence in this notebook refers to text) like summarization, translation, etc.

In this notebook, we will fine-tune the pretrained T5 on the Abstractive Summarization task using Hugging Face Transformers on the XSum dataset loaded from Hugging Face Datasets.


Setup

Installing the requirements

!pip install transformers==4.20.0
!pip install keras_hub==0.3.0
!pip install datasets
!pip install huggingface-hub
!pip install nltk
!pip install rouge-score

Importing the necessary libraries

import os
import logging

import nltk
import numpy as np
import tensorflow as tf
from tensorflow import keras

# Only log error messages
tf.get_logger().setLevel(logging.ERROR)

os.environ["TOKENIZERS_PARALLELISM"] = "false"

Define certain variables

# The percentage of the dataset you want to split as train and test
TRAIN_TEST_SPLIT = 0.1

MAX_INPUT_LENGTH = 1024  # Maximum length of the input to the model
MIN_TARGET_LENGTH = 5  # Minimum length of the output by the model
MAX_TARGET_LENGTH = 128  # Maximum length of the output by the model
BATCH_SIZE = 8  # Batch-size for training our model
LEARNING_RATE = 2e-5  # Learning-rate for training our model
MAX_EPOCHS = 1  # Maximum number of epochs we will train the model for

# This notebook is built on the t5-small checkpoint from the Hugging Face Model Hub
MODEL_CHECKPOINT = "t5-small"

Load the dataset

We will now download the Extreme Summarization (XSum). The dataset consists of BBC articles and accompanying single sentence summaries. Specifically, each article is prefaced with an introductory sentence (aka summary) which is professionally written, typically by the author of the article. That dataset has 226,711 articles divided into training (90%, 204,045), validation (5%, 11,332), and test (5%, 11,334) sets.

Following much of literature, we use the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) metric to evaluate our sequence-to-sequence abstrative summarization approach.

We will use the Hugging Face Datasets library to download the data we need to use for training and evaluation. This can be easily done with the load_dataset function.

from datasets import load_dataset

raw_datasets = load_dataset("xsum", split="train")

The dataset has the following fields:

  • document: the original BBC article to be summarized
  • summary: the single sentence summary of the BBC article
  • id: ID of the document-summary pair
print(raw_datasets)
Dataset({
    features: ['document', 'summary', 'id'],
    num_rows: 204045
})

We will now see how the data looks like:

print(raw_datasets[0])
{'document': 'The full cost of damage in Newton Stewart, one of the areas worst affected, is still being assessed.\nRepair work is ongoing in Hawick and many roads in Peeblesshire remain badly affected by standing water.\nTrains on the west coast mainline face disruption due to damage at the Lamington Viaduct.\nMany businesses and householders were affected by flooding in Newton Stewart after the River Cree overflowed into the town.\nFirst Minister Nicola Sturgeon visited the area to inspect the damage.\nThe waters breached a retaining wall, flooding many commercial properties on Victoria Street - the main shopping thoroughfare.\nJeanette Tate, who owns the Cinnamon Cafe which was badly affected, said she could not fault the multi-agency response once the flood hit.\nHowever, she said more preventative work could have been carried out to ensure the retaining wall did not fail.\n"It is difficult but I do think there is so much publicity for Dumfries and the Nith - and I totally appreciate that - but it is almost like we\'re neglected or forgotten," she said.\n"That may not be true but it is perhaps my perspective over the last few days.\n"Why were you not ready to help us a bit more when the warning and the alarm alerts had gone out?"\nMeanwhile, a flood alert remains in place across the Borders because of the constant rain.\nPeebles was badly hit by problems, sparking calls to introduce more defences in the area.\nScottish Borders Council has put a list on its website of the roads worst affected and drivers have been urged not to ignore closure signs.\nThe Labour Party\'s deputy Scottish leader Alex Rowley was in Hawick on Monday to see the situation first hand.\nHe said it was important to get the flood protection plan right but backed calls to speed up the process.\n"I was quite taken aback by the amount of damage that has been done," he said.\n"Obviously it is heart-breaking for people who have been forced out of their homes and the impact on businesses."\nHe said it was important that "immediate steps" were taken to protect the areas most vulnerable and a clear timetable put in place for flood prevention plans.\nHave you been affected by flooding in Dumfries and Galloway or the Borders? Tell us about your experience of the situation and how it was handled. Email us on selkirk.news@bbc.co.uk or dumfries@bbc.co.uk.', 'summary': 'Clean-up operations are continuing across the Scottish Borders and Dumfries and Galloway after flooding caused by Storm Frank.', 'id': '35232142'}

For the sake of demonstrating the workflow, in this notebook we will only take small stratified balanced splits (10%) of the train as our training and test sets. We can easily split the dataset using the train_test_split method which expects the split size and the name of the column relative to which you want to stratify.

raw_datasets = raw_datasets.train_test_split(
    train_size=TRAIN_TEST_SPLIT, test_size=TRAIN_TEST_SPLIT
)

Data Pre-processing

Before we can feed those texts to our model, we need to pre-process them and get them ready for the task. This is done by a Hugging Face Transformers Tokenizer which will tokenize the inputs (including converting the tokens to their corresponding IDs in the pretrained vocabulary) and put it in a format the model expects, as well as generate the other inputs that model requires.

The from_pretrained() method expects the name of a model from the Hugging Face Model Hub. This is exactly similar to MODEL_CHECKPOINT declared earlier and we will just pass that.

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(MODEL_CHECKPOINT)

If you are using one of the five T5 checkpoints we have to prefix the inputs with "summarize:" (the model can also translate and it needs the prefix to know which task it has to perform).

if MODEL_CHECKPOINT in ["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b"]:
    prefix = "summarize: "
else:
    prefix = ""

We will write a simple function that helps us in the pre-processing that is compatible with Hugging Face Datasets. To summarize, our pre-processing function should:

  • Tokenize the text dataset (input and targets) into it's corresponding token ids that will be used for embedding look-up in BERT
  • Add the prefix to the tokens
  • Create additional inputs for the model like token_type_ids, attention_mask, etc.
def preprocess_function(examples):
    inputs = [prefix + doc for doc in examples["document"]]
    model_inputs = tokenizer(inputs, max_length=MAX_INPUT_LENGTH, truncation=True)

    # Setup the tokenizer for targets
    with tokenizer.as_target_tokenizer():
        labels = tokenizer(
            examples["summary"], max_length=MAX_TARGET_LENGTH, truncation=True
        )

    model_inputs["labels"] = labels["input_ids"]

    return model_inputs

To apply this function on all the pairs of sentences in our dataset, we just use the map method of our dataset object we created earlier. This will apply the function on all the elements of all the splits in dataset, so our training and testing data will be preprocessed in one single command.

tokenized_datasets = raw_datasets.map(preprocess_function, batched=True)

Defining the model

Now we can download the pretrained model and fine-tune it. Since our task is sequence-to-sequence (both the input and output are text sequences), we use the TFAutoModelForSeq2SeqLM class from the Hugging Face Transformers library. Like with the tokenizer, the from_pretrained method will download and cache the model for us.

The from_pretrained() method expects the name of a model from the Hugging Face Model Hub. As mentioned earlier, we will use the t5-small model checkpoint.

from transformers import TFAutoModelForSeq2SeqLM, DataCollatorForSeq2Seq

model = TFAutoModelForSeq2SeqLM.from_pretrained(MODEL_CHECKPOINT)
Downloading:   0%|          | 0.00/231M [00:00<?, ?B/s]

All model checkpoint layers were used when initializing TFT5ForConditionalGeneration.
All the layers of TFT5ForConditionalGeneration were initialized from the model checkpoint at t5-small.
If your task is similar to the task the model of the checkpoint was trained on, you can already use TFT5ForConditionalGeneration for predictions without further training.

For training Sequence to Sequence models, we need a special kind of data collator, which will not only pad the inputs to the maximum length in the batch, but also the labels. Thus, we use the DataCollatorForSeq2Seq provided by the Hugging Face Transformers library on our dataset. The return_tensors='tf' ensures that we get tf.Tensor objects back.

from transformers import DataCollatorForSeq2Seq

data_collator = DataCollatorForSeq2Seq(tokenizer, model=model, return_tensors="tf")

Next we define our training and testing sets with which we will train our model. Again, Hugging Face Datasets provides us with the to_tf_dataset method which will help us integrate our dataset with the collator defined above. The method expects certain parameters:

  • columns: the columns which will serve as our independent variables
  • batch_size: our batch size for training
  • shuffle: whether we want to shuffle our dataset
  • collate_fn: our collator function

Additionally, we also define a relatively smaller generation_dataset to calculate ROUGE scores on the fly while training.

train_dataset = tokenized_datasets["train"].to_tf_dataset(
    batch_size=BATCH_SIZE,
    columns=["input_ids", "attention_mask", "labels"],
    shuffle=True,
    collate_fn=data_collator,
)
test_dataset = tokenized_datasets["test"].to_tf_dataset(
    batch_size=BATCH_SIZE,
    columns=["input_ids", "attention_mask", "labels"],
    shuffle=False,
    collate_fn=data_collator,
)
generation_dataset = (
    tokenized_datasets["test"]
    .shuffle()
    .select(list(range(200)))
    .to_tf_dataset(
        batch_size=BATCH_SIZE,
        columns=["input_ids", "attention_mask", "labels"],
        shuffle=False,
        collate_fn=data_collator,
    )
)

Building and Compiling the the model

Now we will define our optimizer and compile the model. The loss calculation is handled internally and so we need not worry about that!

optimizer = keras.optimizers.Adam(learning_rate=LEARNING_RATE)
model.compile(optimizer=optimizer)
No loss specified in compile() - the model's internal loss computation will be used as the loss. Don't panic - this is a common way to train TensorFlow models in Transformers! To disable this behaviour please pass a loss argument, or explicitly pass `loss=None` if you do not want your model to compute a loss.

Training and Evaluating the model

To evaluate our model on-the-fly while training, we will define metric_fn which will calculate the ROUGE score between the groud-truth and predictions.

import keras_hub

rouge_l = keras_hub.metrics.RougeL()


def metric_fn(eval_predictions):
    predictions, labels = eval_predictions
    decoded_predictions = tokenizer.batch_decode(predictions, skip_special_tokens=True)
    for label in labels:
        label[label < 0] = tokenizer.pad_token_id  # Replace masked label tokens
    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
    result = rouge_l(decoded_labels, decoded_predictions)
    # We will print only the F1 score, you can use other aggregation metrics as well
    result = {"RougeL": result["f1_score"]}

    return result

Now we can finally start training our model!

from transformers.keras_callbacks import KerasMetricCallback

metric_callback = KerasMetricCallback(
    metric_fn, eval_dataset=generation_dataset, predict_with_generate=True
)

callbacks = [metric_callback]

# For now we will use our test set as our validation_data
model.fit(
    train_dataset, validation_data=test_dataset, epochs=MAX_EPOCHS, callbacks=callbacks
)
WARNING:root:No label_cols specified for KerasMetricCallback, assuming you want the 'labels' key.

2551/2551 [==============================] - 652s 250ms/step - loss: 2.9159 - val_loss: 2.5875 - RougeL: 0.2065

<keras.callbacks.History at 0x7f1d002f9810>

For best results, we recommend training the model for atleast 5 epochs on the entire training dataset!


Inference

Now we will try to infer the model we trained on an arbitrary article. To do so, we will use the pipeline method from Hugging Face Transformers. Hugging Face Transformers provides us with a variety of pipelines to choose from. For our task, we use the summarization pipeline.

The pipeline method takes in the trained model and tokenizer as arguments. The framework="tf" argument ensures that you are passing a model that was trained with TF.

from transformers import pipeline

summarizer = pipeline("summarization", model=model, tokenizer=tokenizer, framework="tf")

summarizer(
    raw_datasets["test"][0]["document"],
    min_length=MIN_TARGET_LENGTH,
    max_length=MAX_TARGET_LENGTH,
)
Your max_length is set to 128, but you input_length is only 88. You might consider decreasing max_length manually, e.g. summarizer('...', max_length=44)

[{'summary_text': 'Boss Wagner says he is "a 100% professional and has a winning mentality to play on the pitch."'}]

Now you can push this model to Hugging Face Model Hub and also share it with with all your friends, family, favorite pets: they can all load it with the identifier "your-username/the-name-you-picked" so for instance:

model.push_to_hub("transformers-qa", organization="keras-io")
tokenizer.push_to_hub("transformers-qa", organization="keras-io")

And after you push your model this is how you can load it in the future!

from transformers import TFAutoModelForSeq2SeqLM

model = TFAutoModelForSeq2SeqLM.from_pretrained("your-username/my-awesome-model")