📄 Abstract Generation Based on Rhetorical Structure Extraction

Summary

The authors address domain-independent automatic summarization of Japanese expository texts. Frequency-based extractors ignore discourse structure and produce incoherent output, while knowledge- or script-based systems only work in narrow domains. Their system first builds a rhetorical structure for each section: a binary tree of relations between sentences, and then between paragraphs. It finds these relations from surface clues such as connectives, sentence predicates, anaphoric and idiomatic expressions. When several trees are possible, local preference rules assign penalty scores and the lowest-scoring tree is chosen. Each relation is labeled as favoring its right node, its left node or both. Penalties are passed from the root down to the leaves, and the most penalized sentences are pruned until the summary reaches the requested length. Because whole rhetorical substructures are kept or dropped together, the abstracts avoid isolated sentences that cannot be understood alone. The system was tested on newspaper editorials and technical papers and covered up to 74% of the most important sentences. It is used as a text browser in a prototype document retrieval system.

Key points

• The system recognizes 34 categories of connective expressions, for example serial, summarization, example, reason, parallel and extension. A sentence with no explicit connective defaults to the extension relation.
• Extraction runs in six steps: sentence analysis, relation extraction, segmentation constraints, candidate tree generation, and preference judgement by penalty scores. Intra-paragraph analysis comes first, then inter-paragraph analysis.
• The evaluation data were 30 editorials from "Asahi Shinbun" and 42 papers from "Toshiba Review". Three human subjects marked key sentences and the single most important sentence in each text.
• On Toshiba Review, abstracts were about 24% of the original length. They covered 51% of key sentences and 74% of most important sentences. On editorials, abstracts were about 30% of the original length, with 41% and 60% coverage. The authors explain that editorials contain fewer explicit rhetorical clues.
• The earlier rhetorical structure analyzer [Sumita 92] reached 74% accuracy. The authors list as future work phrase-level selection, anaphora resolution and topic supplementation.