[Paper Review] Abstract Meaning Representation for Multi-Document Summarization

Kexin Liao, Logan Lebanoff+1 more|arXiv (Cornell University)|Jun 14, 2018

Topic Modeling54 references78 citations

TL;DR

The paper investigates using Abstract Meaning Representation (AMR) as a semantic content representation to generate abstractive multi-document summaries, presenting a full AMR-based pipeline with source sentence selection, content planning, and surface realization.

ABSTRACT

Generating an abstract from a collection of documents is a desirable capability for many real-world applications. However, abstractive approaches to multi-document summarization have not been thoroughly investigated. This paper studies the feasibility of using Abstract Meaning Representation (AMR), a semantic representation of natural language grounded in linguistic theory, as a form of content representation. Our approach condenses source documents to a set of summary graphs following the AMR formalism. The summary graphs are then transformed to a set of summary sentences in a surface realization step. The framework is fully data-driven and flexible. Each component can be optimized independently using small-scale, in-domain training data. We perform experiments on benchmark summarization datasets and report promising results. We also describe opportunities and challenges for advancing this line of research.

Motivation & Objective

• Assess feasibility of AMR as a content representation for multi-document summarization.
• Develop a data-driven pipeline to condense multiple sources into summary AMR graphs and realize them as text.
• Evaluate AMR-based summarization against state-of-the-art baselines on standard datasets.
• Analyze the impact of AMR parsers and source sentence selection strategies on summarization performance.

Proposed method

• Three-component pipeline: source sentence selection, content planning, surface realization.
• Convert selected sentences to AMR graphs using JAMR or CAMR parsers.
• Merge graphs with coreference resolution into a connected source graph.
• Extract a summary graph via a trainable structured prediction framework using ILP decoding and structured ramp loss.
• Transform the summary AMR graph to PENMAN format and generate text with JAMR AMR-to-text generator.

Experimental results

Research questions

• RQ1 Can AMR serve as an effective content representation for abstractive multi-document summarization?
• RQ2 How do AMR parsers and source sentence selection strategies impact summary quality?
• RQ3 To what extent can a structured prediction approach produce salient summary graphs from a cluster of source AMR graphs?
• RQ4 How does AMR-based summarization compare to neural encoder-decoder baselines on standard benchmarks?

Key findings

• AMR-based summarization is competitive with state-of-the-art abstractive baselines on benchmark datasets.
• Using concept-based source sentence selection (Concept Cov) yields stronger node preservation in summary graphs.
• VSM-based edge prediction best preserves summary relations; oracle decoding boosts performance further.
• AMR parser quality impacts summarization, with JAMR outperforming CAMR by a small margin in this setup.
• Summaries produced by AMR-based methods are more abstractive than many extractive baselines and have lower n-gram overlap with source documents compared to pointer-generator variants.