README.md

---
language:
- en
tags:
- summarization
- led
- summary
- longformer
- booksum
- long-document
- long-form
license:
- apache-2.0
- bsd-3-clause
datasets:
- kmfoda/booksum
metrics:
- rouge
widget:
- text: large earthquakes along a given fault segment do not occur at random intervals
    because it takes time to accumulate the strain energy for the rupture. The rates
    at which tectonic plates move and accumulate strain at their boundaries are approximately
    uniform. Therefore, in first approximation, one may expect that large ruptures
    of the same fault segment will occur at approximately constant time intervals.
    If subsequent main shocks have different amounts of slip across the fault, then
    the recurrence time may vary, and the basic idea of periodic mainshocks must be
    modified. For great plate boundary ruptures the length and slip often vary by
    a factor of 2. Along the southern segment of the San Andreas fault the recurrence
    interval is 145 years with variations of several decades. The smaller the standard
    deviation of the average recurrence interval, the more specific could be the long
    term prediction of a future mainshock.
  example_title: earthquakes
- text: " A typical feed-forward neural field algorithm. Spatiotemporal coordinates\
    \ are fed into a neural network that predicts values in the reconstructed domain.\
    \ Then, this domain is mapped to the sensor domain where sensor measurements are\
    \ available as supervision. Class and Section Problems Addressed Generalization\
    \ (Section 2) Inverse problems, ill-posed problems, editability; symmetries. Hybrid\
    \ Representations (Section 3) Computation & memory efficiency, representation\
    \ capacity, editability: Forward Maps (Section 4) Inverse problems Network Architecture\
    \ (Section 5) Spectral bias, integration & derivatives. Manipulating Neural Fields\
    \ (Section 6) Edit ability, constraints, regularization. Table 2: The five classes\
    \ of techniques in the neural field toolbox each addresses problems that arise\
    \ in learning, inference, and control. (Section 3). We can supervise reconstruction\
    \ via differentiable forward maps that transform Or project our domain (e.g, 3D\
    \ reconstruction via 2D images; Section 4) With appropriate network architecture\
    \ choices, we can overcome neural network spectral biases (blurriness) and efficiently\
    \ compute derivatives and integrals (Section 5). Finally, we can manipulate neural\
    \ fields to add constraints and regularizations, and to achieve editable representations\
    \ (Section 6). Collectively, these classes constitute a 'toolbox' of techniques\
    \ to help solve problems with neural fields There are three components in a conditional\
    \ neural field: (1) An encoder or inference function \u20AC that outputs the conditioning\
    \ latent variable 2 given an observation 0 E(0) =2. 2 is typically a low-dimensional\
    \ vector, and is often referred to aS a latent code Or feature code_ (2) A mapping\
    \ function 4 between Z and neural field parameters O: Y(z) = O; (3) The neural\
    \ field itself $. The encoder \u20AC finds the most probable z given the observations\
    \ O: argmaxz P(2/0). The decoder maximizes the inverse conditional probability\
    \ to find the most probable 0 given Z: arg- max P(Olz). We discuss different encoding\
    \ schemes with different optimality guarantees (Section 2.1.1), both global and\
    \ local conditioning (Section 2.1.2), and different mapping functions Y (Section\
    \ 2.1.3) 2. Generalization Suppose we wish to estimate a plausible 3D surface\
    \ shape given a partial or noisy point cloud. We need a suitable prior over the\
    \ sur- face in its reconstruction domain to generalize to the partial observations.\
    \ A neural network expresses a prior via the function space of its architecture\
    \ and parameters 0, and generalization is influenced by the inductive bias of\
    \ this function space (Section 5)."
  example_title: scientific paper
- text: ' the big variety of data coming from diverse sources is one of the key properties
    of the big data phenomenon. It is, therefore, beneficial to understand how data
    is generated in various environments and scenarios, before looking at what should
    be done with this data and how to design the best possible architecture to accomplish
    this The evolution of IT architectures, described in Chapter 2, means that the
    data is no longer processed by a few big monolith systems, but rather by a group
    of services In parallel to the processing layer, the underlying data storage has
    also changed and became more distributed This, in turn, required a significant
    paradigm shift as the traditional approach to transactions (ACID) could no longer
    be supported. On top of this, cloud computing is becoming a major approach with
    the benefits of reducing costs and providing on-demand scalability but at the
    same time introducing concerns about privacy, data ownership, etc In the meantime
    the Internet continues its exponential growth: Every day both structured and unstructured
    data is published and available for processing: To achieve competitive advantage
    companies have to relate their corporate resources to external services, e.g.
    financial markets, weather forecasts, social media, etc While several of the sites
    provide some sort of API to access the data in a more orderly fashion; countless
    sources require advanced web mining and Natural Language Processing (NLP) processing
    techniques: Advances in science push researchers to construct new instruments
    for observing the universe O conducting experiments to understand even better
    the laws of physics and other domains. Every year humans have at their disposal
    new telescopes, space probes, particle accelerators, etc These instruments generate
    huge streams of data, which need to be stored and analyzed. The constant drive
    for efficiency in the industry motivates the introduction of new automation techniques
    and process optimization: This could not be done without analyzing the precise
    data that describe these processes. As more and more human tasks are automated,
    machines provide rich data sets, which can be analyzed in real-time to drive efficiency
    to new levels. Finally, it is now evident that the growth of the Internet of Things
    is becoming a major source of data. More and more of the devices are equipped
    with significant computational power and can generate a continuous data stream
    from their sensors. In the subsequent sections of this chapter, we will look at
    the domains described above to see what they generate in terms of data sets. We
    will compare the volumes but will also look at what is characteristic and important
    from their respective points of view. 3.1 The Internet is undoubtedly the largest
    database ever created by humans. While several well described; cleaned, and structured
    data sets have been made available through this medium, most of the resources
    are of an ambiguous, unstructured, incomplete or even erroneous nature. Still,
    several examples in the areas such as opinion mining, social media analysis, e-governance,
    etc, clearly show the potential lying in these resources. Those who can successfully
    mine and interpret the Internet data can gain unique insight and competitive advantage
    in their business An important area of data analytics on the edge of corporate
    IT and the Internet is Web Analytics.'
  example_title: data science textbook
- text: "Transformer-based models have shown to be very useful for many NLP tasks.\
    \ However, a major limitation of transformers-based models is its O(n^2)O(n 2)\
    \ time & memory complexity (where nn is sequence length). Hence, it's computationally\
    \ very expensive to apply transformer-based models on long sequences n > 512n>512.\
    \ Several recent papers, e.g. Longformer, Performer, Reformer, Clustered attention\
    \ try to remedy this problem by approximating the full attention matrix. You can\
    \ checkout \U0001F917's recent blog post in case you are unfamiliar with these\
    \ models.\nBigBird (introduced in paper) is one of such recent models to address\
    \ this issue. BigBird relies on block sparse attention instead of normal attention\
    \ (i.e. BERT's attention) and can handle sequences up to a length of 4096 at a\
    \ much lower computational cost compared to BERT. It has achieved SOTA on various\
    \ tasks involving very long sequences such as long documents summarization, question-answering\
    \ with long contexts.\nBigBird RoBERTa-like model is now available in \U0001F917\
    Transformers. The goal of this post is to give the reader an in-depth understanding\
    \ of big bird implementation & ease one's life in using BigBird with \U0001F917\
    Transformers. But, before going into more depth, it is important to remember that\
    \ the BigBird's attention is an approximation of BERT's full attention and therefore\
    \ does not strive to be better than BERT's full attention, but rather to be more\
    \ efficient. It simply allows to apply transformer-based models to much longer\
    \ sequences since BERT's quadratic memory requirement quickly becomes unbearable.\
    \ Simply put, if we would have \u221E compute & \u221E time, BERT's attention\
    \ would be preferred over block sparse attention (which we are going to discuss\
    \ in this post).\nIf you wonder why we need more compute when working with longer\
    \ sequences, this blog post is just right for you!\nSome of the main questions\
    \ one might have when working with standard BERT-like attention include:\nDo all\
    \ tokens really have to attend to all other tokens? Why not compute attention\
    \ only over important tokens? How to decide what tokens are important? How to\
    \ attend to just a few tokens in a very efficient way? In this blog post, we will\
    \ try to answer those questions.\nWhat tokens should be attended to? We will give\
    \ a practical example of how attention works by considering the sentence 'BigBird\
    \ is now available in HuggingFace for extractive question answering'. In BERT-like\
    \ attention, every word would simply attend to all other tokens.\nLet's think\
    \ about a sensible choice of key tokens that a queried token actually only should\
    \ attend to by writing some pseudo-code. Will will assume that the token available\
    \ is queried and build a sensible list of key tokens to attend to.\n>>> # let's\
    \ consider following sentence as an example >>> example = ['BigBird', 'is', 'now',\
    \ 'available', 'in', 'HuggingFace', 'for', 'extractive', 'question', 'answering']\n\
    >>> # further let's assume, we're trying to understand the representation of 'available'\
    \ i.e. >>> query_token = 'available' >>> # We will initialize an empty `set` and\
    \ fill up the tokens of our interest as we proceed in this section. >>> key_tokens\
    \ = [] # => currently 'available' token doesn't have anything to attend Nearby\
    \ tokens should be important because, in a sentence (sequence of words), the current\
    \ word is highly dependent on neighboring past & future tokens. This intuition\
    \ is the idea behind the concept of sliding attention."
  example_title: bigbird blog intro
- text: 'The majority of available text summarization datasets include short-form
    source documents that lack long-range causal and temporal dependencies, and often
    contain strong layout and stylistic biases. While relevant, such datasets will
    offer limited challenges for future generations of text summarization systems.
    We address these issues by introducing BookSum, a collection of datasets for long-form
    narrative summarization. Our dataset covers source documents from the literature
    domain, such as novels, plays and stories, and includes highly abstractive, human
    written summaries on three levels of granularity of increasing difficulty: paragraph-,
    chapter-, and book-level. The domain and structure of our dataset poses a unique
    set of challenges for summarization systems, which include: processing very long
    documents, non-trivial causal and temporal dependencies, and rich discourse structures.
    To facilitate future work, we trained and evaluated multiple extractive and abstractive
    summarization models as baselines for our dataset.'
  example_title: BookSum Abstract
inference:
  parameters:
    max_length: 64
    min_length: 8
    no_repeat_ngram_size: 3
    early_stopping: true
    repetition_penalty: 3.5
    length_penalty: 0.3
    encoder_no_repeat_ngram_size: 3
    num_beams: 4
model-index:
- name: pszemraj/led-large-book-summary
  results:
  - task:
      type: summarization
      name: Summarization
    dataset:
      name: kmfoda/booksum
      type: kmfoda/booksum
      config: kmfoda--booksum
      split: test
    metrics:
    - name: ROUGE-1
      type: rouge
      value: 31.7308
      verified: true
    - name: ROUGE-2
      type: rouge
      value: 5.3311
      verified: true
    - name: ROUGE-L
      type: rouge
      value: 16.1465
      verified: true
    - name: ROUGE-LSUM
      type: rouge
      value: 29.0883
      verified: true
    - name: loss
      type: loss
      value: 4.815707206726074
      verified: true
    - name: gen_len
      type: gen_len
      value: 154.9036
      verified: true
  - task:
      type: summarization
      name: Summarization
    dataset:
      name: samsum
      type: samsum
      config: samsum
      split: test
    metrics:
    - name: ROUGE-1
      type: rouge
      value: 33.4484
      verified: true
    - name: ROUGE-2
      type: rouge
      value: 10.4249
      verified: true
    - name: ROUGE-L
      type: rouge
      value: 24.5802
      verified: true
    - name: ROUGE-LSUM
      type: rouge
      value: 29.8226
      verified: true
    - name: loss
      type: loss
      value: 4.176078796386719
      verified: true
    - name: gen_len
      type: gen_len
      value: 65.4005
      verified: true
  - task:
      type: summarization
      name: Summarization
    dataset:
      name: billsum
      type: billsum
      config: default
      split: test
    metrics:
    - name: ROUGE-1
      type: rouge
      value: 40.5843
      verified: true
    - name: ROUGE-2
      type: rouge
      value: 17.3401
      verified: true
    - name: ROUGE-L
      type: rouge
      value: 25.1256
      verified: true
    - name: ROUGE-LSUM
      type: rouge
      value: 34.6619
      verified: true
    - name: loss
      type: loss
      value: 4.792657375335693
      verified: true
    - name: gen_len
      type: gen_len
      value: 163.9394
      verified: true
  - task:
      type: summarization
      name: Summarization
    dataset:
      name: multi_news
      type: multi_news
      config: default
      split: test
    metrics:
    - name: ROUGE-1
      type: rouge
      value: 39.0834
      verified: true
    - name: ROUGE-2
      type: rouge
      value: 11.4043
      verified: true
    - name: ROUGE-L
      type: rouge
      value: 19.1813
      verified: true
    - name: ROUGE-LSUM
      type: rouge
      value: 35.1581
      verified: true
    - name: loss
      type: loss
      value: 4.654905319213867
      verified: true
    - name: gen_len
      type: gen_len
      value: 186.2494
      verified: true
---

# Longformer Encoder-Decoder (LED) for Narrative-Esque Long Text Summarization

<a href="https://colab.research.google.com/gist/pszemraj/3eba944ddc9fc9a4a1bfb21e83b57620/summarization-token-batching.ipynb">
  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
</a>

A fine-tuned version of [allenai/led-large-16384](https://huggingface.co/allenai/led-large-16384) on the `BookSum` dataset.

Goal: a model that can generalize well and is useful in summarizing long text in academic and daily usage. The result works well on lots of text and can handle 16384 tokens/batch (_if you have the GPU memory to handle that_)

 - See the Colab demo linked above or try the [demo on Spaces](https://huggingface.co/spaces/pszemraj/summarize-long-text)


> Note: the API is set to generate a max of 64 tokens for runtime reasons, so the summaries may be truncated (depending on the length of input text). For best results use python as below.

---

# Usage - Basic

- use `encoder_no_repeat_ngram_size=3` when calling the pipeline object to improve summary quality.
  - this forces the model to use new vocabulary and create an abstractive summary, otherwise it may compile the best _extractive_ summary from the input provided.

Load the model into a pipeline object:

```python
import torch
from transformers import pipeline

hf_name = 'pszemraj/led-large-book-summary'

summarizer = pipeline(
    "summarization",
    hf_name,
    device=0 if torch.cuda.is_available() else -1,
)
```

- put words into the pipeline object:

```python
wall_of_text = "your words here"

result = summarizer(
    wall_of_text,
    min_length=16,
    max_length=256,
    no_repeat_ngram_size=3,
    encoder_no_repeat_ngram_size=3,
    repetition_penalty=3.5,
    num_beams=4,
    early_stopping=True,
)
```


**Important:** To generate the best quality summaries, you should use the global attention mask when decoding, as demonstrated in [this community notebook here](https://colab.research.google.com/drive/12INTTR6n64TzS4RrXZxMSXfrOd9Xzamo?usp=sharing), see the definition of `generate_answer(batch)`.

If having computing constraints, try the base version [`pszemraj/led-base-book-summary`](https://huggingface.co/pszemraj/led-base-book-summary)
- all the parameters for generation on the API here are the same as [the base model](https://huggingface.co/pszemraj/led-base-book-summary) for easy comparison between versions.

## Training and evaluation data

- the [booksum](https://arxiv.org/abs/2105.08209) dataset (this is what adds the `bsd-3-clause` license)
- During training, the input text was the text of the `chapter`, and the output was `summary_text`
- Eval results can be found [here](https://huggingface.co/datasets/autoevaluate/autoeval-staging-eval-project-kmfoda__booksum-79c1c0d8-10905463) with metrics on the sidebar.

## Training procedure

- Training completed on the BookSum dataset for 13 total epochs
- **The final four epochs combined the training and validation sets as 'train' in an effort to increase generalization.**

### Training hyperparameters

#### Initial Three Epochs

The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 4
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 3

#### In-between Epochs

Unfortunately, don't have all records on-hand for middle epochs; the following should be representative:

- learning_rate: 4e-05
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 16
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.05
- num_epochs: 6 (in addition to prior model)

#### Final Two Epochs

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 16
- total_train_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.03
- num_epochs: 2 (in addition to prior model)


### Framework versions

- Transformers 4.19.2
- Pytorch 1.11.0+cu113
- Datasets 2.2.2
- Tokenizers 0.12.1