Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MODABS: Multi-Objective Learning for Dynamic Aspect-Based Summarization

Xiaobo Guo, Soroush Vosoughi

TL;DR

MODABS addresses dynamic aspect-based summarization by jointly discovering and summarizing multiple aspects without predefined counts. It extends Longformer-Encoder-Decoder with a multi-objective fine-tuning regime that predicts the number of aspects, enforces distinct, high-quality per-aspect summaries, and promotes diversity across aspects via KL-divergence constraints. Empirical results on Disordered-DABS and OASUM show consistent improvements over strong baselines and LLM prompting, with notably accurate aspect-count predictions and enhanced multi-aspect quality in human judgments. This approach offers a scalable, encoder-agnostic strategy for robust, domain-agnostic dynamic AB summarization in long-form content.

Abstract

The rapid proliferation of online content necessitates effective summarization methods, among which dynamic aspect-based summarization stands out. Unlike its traditional counterpart, which assumes a fixed set of known aspects, this approach adapts to the varied aspects of the input text. We introduce a novel multi-objective learning framework employing a Longformer-Encoder-Decoder for this task. The framework optimizes aspect number prediction, minimizes disparity between generated and reference summaries for each aspect, and maximizes dissimilarity across aspect-specific summaries. Extensive experiments show our method significantly outperforms baselines on three diverse datasets, largely due to the effective alignment of generated and reference aspect counts without sacrificing single-aspect summarization quality.

MODABS: Multi-Objective Learning for Dynamic Aspect-Based Summarization

TL;DR

MODABS addresses dynamic aspect-based summarization by jointly discovering and summarizing multiple aspects without predefined counts. It extends Longformer-Encoder-Decoder with a multi-objective fine-tuning regime that predicts the number of aspects, enforces distinct, high-quality per-aspect summaries, and promotes diversity across aspects via KL-divergence constraints. Empirical results on Disordered-DABS and OASUM show consistent improvements over strong baselines and LLM prompting, with notably accurate aspect-count predictions and enhanced multi-aspect quality in human judgments. This approach offers a scalable, encoder-agnostic strategy for robust, domain-agnostic dynamic AB summarization in long-form content.

Abstract

The rapid proliferation of online content necessitates effective summarization methods, among which dynamic aspect-based summarization stands out. Unlike its traditional counterpart, which assumes a fixed set of known aspects, this approach adapts to the varied aspects of the input text. We introduce a novel multi-objective learning framework employing a Longformer-Encoder-Decoder for this task. The framework optimizes aspect number prediction, minimizes disparity between generated and reference summaries for each aspect, and maximizes dissimilarity across aspect-specific summaries. Extensive experiments show our method significantly outperforms baselines on three diverse datasets, largely due to the effective alignment of generated and reference aspect counts without sacrificing single-aspect summarization quality.
Paper Structure (27 sections, 1 equation, 4 figures, 6 tables)

This paper contains 27 sections, 1 equation, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Datasets for Aspect-Based Summarization
  4. Methods in Dynamic Aspect-Based Summarization
  5. Datasets
  6. Methodology
  7. Decoder Embeddings
  8. Multi-Objective Learning
  9. Experiments
  10. Baselines and Metrics
  11. Automatic Evaluation Results
  12. Human Evaluations
  13. Ablation Analysis
  14. Examples of Summaries
  15. Influence of the Multiple Objectives
  16. ...and 12 more sections

Figures (4)

  • Figure 1: Diagram illustrating aspect-based summarization, with distinct colors representing different aspects. "[SEN $i$]" indicates the $i$-th sentence in the source article.
  • Figure 2: Diagram of our framework. Colored sentences represent different aspects. "[SEN i]" indicates the i-th sentence in the input. Aspects, tokens, and their generated summaries are denoted as $Asp_i$, $T_i$, and $\widehat{Sum_i}$, respectively. The predicted number of aspects is $\widehat{\#Asp}$, while ground-truth summaries and aspect numbers are $Sum_i$ and $\#Asp$, respectively. Cross-entropy loss and KL divergence loss are indicated by "CE" and "KLD". Weights for different losses are $\lambda_{1/2/3}$.
  • Figure 3: An example of the source article, reference, and the generated Summaries. Empty quotes (" ") indicate that no generated summaries correspond to this reference summary.
  • Figure 4: Distribution of aspect number differences between reference and generated summaries for all three datasets.