[Paper Review] Genetic Algorithms For Extractive Summarization

William Chen, Kensal Ramos+2 more|arXiv (Cornell University)|May 5, 2021

Topic Modeling11 references4 citations

TL;DR

This paper proposes a genetic algorithm (GA)-based approach for extractive text summarization that learns optimal word weights to identify important sentences, avoiding deep learning's data and compute demands. The GA evolves sentence importance scores by optimizing word weights in a vocabulary, achieving competitive ROUGE scores (up to 28.6 on training set) with minimal hyperparameter tuning, demonstrating generalization even with large, noisy vocabularies.

ABSTRACT

Most current work in NLP utilizes deep learning, which requires a lot of training data and computational power. This paper investigates the strengths of Genetic Algorithms (GAs) for extractive summarization, as we hypothesized that GAs could construct more efficient solutions for the summarization task due to their relative customizability relative to deep learning models. This is done by building a vocabulary set, the words of which are represented as an array of weights, and optimizing those set of weights with the GA. These weights can be used to build an overall weighting of a sentence, which can then be passed to some threshold for extraction. Our results showed that the GA was able to learn a weight representation that could filter out excessive vocabulary and thus dictate sentence importance based on common English words.

Motivation & Objective

• To explore genetic algorithms as a lightweight, customizable alternative to deep learning for extractive text summarization.
• To investigate whether GAs can learn generalizable sentence importance features without relying on sentence-level heuristics.
• To evaluate the impact of training set size and vocabulary size on summarization performance using ROUGE metrics.
• To determine if GAs can effectively filter noise in large vocabularies and converge on meaningful word weight representations.
• To assess the feasibility of using GAs for low-power, efficient summarization systems.

Proposed method

• The method represents each word in the vocabulary as a gene with a weight between 0 and 1, forming a chromosome for each individual in the GA population.
• Sentence importance is computed as the average weight of all words in the sentence, with unknown words defaulting to weight 0.
• A fixed threshold (0.6) is applied to select sentences whose average weight exceeds it, forming the summary.
• The GA uses standard operations: selection, crossover, and mutation (gene deletion by setting weight to 0) over multiple generations to optimize fitness.
• Fitness is evaluated using the ROUGE-1 F1 score against reference summaries on training and test sets.
• Vocabulary is built from training data, with experiments varying training set size (50, 100) and vocabulary size (50, 1,000, 90,000).

Experimental results

Research questions

• RQ1 Can a genetic algorithm learn meaningful word weights that effectively identify important sentences in extractive summarization?
• RQ2 How does the size of the training set and vocabulary affect the performance of the GA-based summarization model?
• RQ3 Does the GA generalize well across different data distributions, especially when vocabulary size exceeds training data size?
• RQ4 Can the GA effectively filter out irrelevant or noisy words in a large vocabulary to focus on semantically important terms?
• RQ5 To what extent can the GA reduce reliance on handcrafted sentence-level heuristics compared to prior work?

Key findings

• The best-performing model achieved a ROUGE-1 score of 28.6 on the training set when trained on 100 samples with a full vocabulary of 90,000 words.
• Models trained on 50 samples with a vocabulary built from those same 50 samples achieved a ROUGE score of 26.26 on the test set, outperforming models trained on 100 samples with larger vocabularies.
• The GA converged quickly, reaching a local optimum within 35 generations, even with a population size of 100 and a vocabulary of 325,000 tokens.
• Despite a large solution space, the GA effectively filtered noise and learned to prioritize common, high-impact English words for summarization.
• ROUGE scores showed minimal improvement with increased vocabulary size, suggesting the GA learns to focus on the most salient words regardless of vocabulary scale.
• The model trained on 100 samples with a 90,000-word vocabulary achieved a test ROUGE score of 23.59, indicating strong generalization from limited data.