📄 Frame Semantic-Enhanced Sentence Modeling for Sentence-level Extractive Text Summarization

Summary

The paper addresses sentence-level extractive summarization, where a system picks the document sentences that best convey its main content. The authors argue that sentences are usually modeled without explicit semantic structure, so the model can miss links between sentences that describe the same situation in different words. Their solution draws on frame semantics, the FrameNet-style idea that words evoke frames (e.g., Causation, Kinship, Taking_time) whose frame elements (such as Actor, Effect, Time) describe the participants. Sentences are first encoded with BERT, using a CLS token per sentence. The model also encodes frames using their textual definitions. All of this is placed in a graph module with three node types: sentences, frames and frame elements. Attention within a sentence (intra) and across sentences (inter) spreads frame information between related sentences before each sentence is scored for selection. Experiments use news data, including CNN/Daily Mail. The work matters because it shows that structured lexical-semantic resources can strengthen pretrained encoders for summarization.

Key points

• Venue: EMNLP 2021 Proceedings, pages 4045–4052; first author Yong Guan.
• Graph nodes are sentence nodes (V_sen), frame nodes (V_f) and frame element nodes (V_fe). Sentence representations come from a BERT document encoder.
• Frame representations are built from their FrameNet definitions. For example, Causation is defined roughly as "a cause causes an effect."
• Intra-sentence and inter-sentence attention model frame-to-frame relations. These relations connect, for instance, an article's lead sentences with a later sentence that restates the same event.
• The motivating example comes from CNN/Daily Mail. The extracted text available here did not include exact ROUGE scores or the full dataset list, so specific numbers are omitted.