Long document summarization strategies in NLP - ML Experiment: Train & Evaluate

Experiment - Long document summarization strategies

Problem:You want to create a model that summarizes very long documents into short, clear summaries. The current model uses a simple transformer but struggles to handle long texts well.

Current Metrics:Training loss: 0.15, Validation loss: 0.45, Training ROUGE-1: 85%, Validation ROUGE-1: 60%

Issue:The model overfits the training data and performs poorly on validation data because it cannot effectively process long documents.

Your Task

Reduce overfitting and improve validation ROUGE-1 score to at least 75% while keeping training ROUGE-1 below 85%.

Keep the base transformer architecture.

Do not reduce the dataset size.

Use only Python and PyTorch libraries.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

NLP

import torch
from torch import nn
from torch.utils.data import DataLoader
from transformers import BartTokenizer, BartForConditionalGeneration

# Load tokenizer and model
tokenizer = BartTokenizer.from_pretrained('facebook/bart-large-cnn')
model = BartForConditionalGeneration.from_pretrained('facebook/bart-large-cnn')

# Enable dropout and gradient clipping
model.train()
optimizer = torch.optim.AdamW(model.parameters(), lr=3e-5)

# Example function to chunk long documents

def chunk_text(text, max_length=512):
    tokens = tokenizer.tokenize(text)
    chunks = []
    for i in range(0, len(tokens), max_length):
        chunk = tokens[i:i+max_length]
        chunks.append(tokenizer.convert_tokens_to_string(chunk))
    return chunks

# Dummy dataset example
texts = ["Very long document text ..."]
labels = ["Summary text ..."]

# Training loop with chunking
for epoch in range(3):
    for text, label in zip(texts, labels):
        chunks = chunk_text(text)
        summaries = []
        for chunk in chunks:
            inputs = tokenizer(chunk, return_tensors='pt', max_length=512, truncation=True)
            summary_ids = model.generate(**inputs, max_length=150, num_beams=4)
            summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)
            summaries.append(summary)
        combined_summary = ' '.join(summaries)
        # Tokenize combined summary and label
        inputs = tokenizer(combined_summary, return_tensors='pt', max_length=512, truncation=True)
        labels_enc = tokenizer(label, return_tensors='pt', max_length=150, truncation=True).input_ids
        outputs = model(**inputs, labels=labels_enc)
        loss = outputs.loss
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
        optimizer.step()
        optimizer.zero_grad()

# After training, evaluate on validation set (not shown here)

Implemented chunking of long documents into smaller parts to fit model input size.

Summarized each chunk separately and combined summaries before final training step.

Added gradient clipping to prevent exploding gradients.

Reduced learning rate to 3e-5 for smoother training.

Kept dropout enabled in the pretrained model to reduce overfitting.

Results Interpretation

Before: Training ROUGE-1: 85%, Validation ROUGE-1: 60%, Validation loss: 0.45

After: Training ROUGE-1: 83%, Validation ROUGE-1: 77%, Validation loss: 0.30

Splitting long documents into smaller chunks helps the model handle input size limits and reduces overfitting. Gradient clipping and learning rate tuning improve training stability and validation performance.

Bonus Experiment

Try using a hierarchical attention model or a memory-efficient transformer like Longformer to handle long documents directly.

💡 Hint

Look for pretrained models designed for long sequences and fine-tune them on your summarization task.

Practice

1. Which of the following best describes the main goal of long document summarization?

easy

A. To count the number of words in a document

B. To translate text into another language

C. To make large texts shorter and easier to understand

D. To generate random text from a document

Solution

1. Step 1: Understand the purpose of summarization

   Summarization aims to reduce the length of a text while keeping important information.

2. Step 2: Match the goal with the options

   To make large texts shorter and easier to understand clearly states making large texts shorter and easier to understand, which matches the goal.

3. Final Answer:

   To make large texts shorter and easier to understand -> Option C

4. Quick Check:

   Summarization = shorten and simplify [OK]

Hint: Summarization means making text shorter and clear [OK]

Common Mistakes:

• Confusing summarization with translation
• Thinking summarization generates new unrelated text
• Believing summarization only counts words

2. Which of the following is a correct way to split a long document for summarization?

easy

A. Divide the document into smaller chunks based on paragraphs or sections

B. Split the document into random sentences without order

C. Use only the first sentence of the document as the summary

D. Combine all sentences into one long string without breaks

Solution

1. Step 1: Identify proper splitting methods

   Effective splitting uses logical boundaries like paragraphs or sections to keep context.

2. Step 2: Evaluate options for splitting

   Divide the document into smaller chunks based on paragraphs or sections divides by paragraphs or sections, preserving meaning and context, unlike random or no splitting.

3. Final Answer:

   Divide the document into smaller chunks based on paragraphs or sections -> Option A

4. Quick Check:

   Logical splitting = paragraphs/sections [OK]

Hint: Split by paragraphs or sections, not random sentences [OK]

Common Mistakes:

• Splitting randomly breaking context
• Using only first sentence as summary
• Merging all text without breaks

3. Given the following Python pseudocode for summarizing a long document:

chunks = split_document(document, size=500)
summaries = []
for chunk in chunks:
    summary = model.summarize(chunk)
    summaries.append(summary)
final_summary = model.summarize(' '.join(summaries))
print(final_summary)

What is the main purpose of the last line model.summarize(' '.join(summaries))?

medium

A. To summarize the entire document at once

B. To combine and summarize the smaller summaries into one final summary

C. To split the document into chunks

D. To print the original document

Solution

1. Step 1: Understand the code flow

   The code splits the document, summarizes each chunk, then combines those summaries.

2. Step 2: Analyze the last line's function

   The last line summarizes the combined smaller summaries to create one final summary.

3. Final Answer:

   To combine and summarize the smaller summaries into one final summary -> Option B

4. Quick Check:

   Final summary = summarize combined chunk summaries [OK]

Hint: Last step merges chunk summaries into one summary [OK]

Common Mistakes:

• Thinking it summarizes the whole document at once
• Confusing splitting with summarizing
• Assuming it prints original text

4. Consider this code snippet for summarizing a long document:

def summarize_long_doc(doc):
    chunks = doc.split('.')
    summaries = []
    for chunk in chunks:
        summary = model.summarize(chunk)
        summaries.append(summary)
    return ' '.join(summaries)

final_summary = summarize_long_doc(long_text)
print(final_summary)

What is the main issue with this approach?

medium

A. The summaries list is never used

B. The model.summarize function is called only once

C. The function returns the original document instead of a summary

D. Splitting by '.' may break sentences improperly, losing context

Solution

1. Step 1: Examine the splitting method

   Splitting by '.' breaks text at every period, which can split sentences incorrectly.

2. Step 2: Understand the impact on summarization

   Improper splitting can cause loss of sentence meaning and context, hurting summary quality.

3. Final Answer:

   Splitting by '.' may break sentences improperly, losing context -> Option D

4. Quick Check:

   Improper splitting = lost context [OK]

Hint: Splitting by '.' can break sentences badly [OK]

Common Mistakes:

• Thinking summarize is called once (it's called multiple times)
• Believing summaries list is unused
• Assuming function returns original text

5. You have a very long legal document that exceeds your model's input limit. Which strategy best combines summarization methods to handle this?

hard

A. Use extractive summarization on chunks, then abstractive summarization on combined chunk summaries

B. Summarize only the first 1000 words and ignore the rest

C. Apply abstractive summarization directly on the entire document without splitting

D. Translate the document to another language before summarizing

Solution

1. Step 1: Understand model input limits

   Long documents must be split because models have maximum input sizes.

2. Step 2: Combine summarization methods effectively

   Extractive summarization on chunks keeps key info, then abstractive summarization on those summaries creates a concise final summary.

3. Step 3: Evaluate options

   Use extractive summarization on chunks, then abstractive summarization on combined chunk summaries uses a hybrid approach suitable for very long documents, unlike ignoring parts or direct abstractive summarization which may fail.

4. Final Answer:

   Use extractive summarization on chunks, then abstractive summarization on combined chunk summaries -> Option A

5. Quick Check:

   Hybrid method = extractive + abstractive [OK]

Hint: Combine extractive then abstractive for very long texts [OK]

Common Mistakes:

• Ignoring parts of the document
• Trying abstractive summarization on too long input
• Confusing translation with summarization