DSGSUM Document-Level Semantic Graph Abstractive Summarization
Architectures AIMech:DSGSUMDocumentLevelSemanticGraphAbstractiveSummarization DRAFT
Aliases: DSGSUM, Document-Level Semantic Graph Abstractive Summarization, DSGSUM Hybrid Extractive-Abstractive Summarizer
A hybrid extractive-abstractive neural text summarization architecture that represents a source document as a sentence-level semantic graph weighted by log-length-normalized lexical overlap, computes damped random-walk TextRank authority scores to select salient sentences before encoding, and feeds the filtered sentence sequence into a pre-trained BERT encoder and multi-layer Transformer decoder to generate abstractive summaries within fixed token-length limits.
Identity Notes & Concept Scope
No AIO class represents DSGSUM or document-level semantic graph TextRank-guided BERT summarization; ArtificialNeuralNetwork (AIO:ArtificialNeuralNetwork) is the closest broad match covering neural network architectures.
Claims & Grounded Evidence
4 Claims
Mechanism
#document-semantic-graph-textrank-centrality
DSGSUM constructs a sentence-level graph with edge weights proportional to shared word counts divided by the sum of log sentence lengths and computes the stationary distribution of a damped random walk to score global sentence salience.
Mechanism
#salient-sentence-filtering-before-bert-transformer-decoding
By selecting top-ranked sentences from the document-level semantic graph before applying BERT token/segment/position encoding and a 6-layer Transformer decoder, DSGSUM retains late-occurring salient content that standard 512-token truncation discards.
Application
#news-and-long-document-abstractive-summarization-application
DSGSUM is applied to single-document abstractive summarization on news corpora such as CNN/DailyMail and New York Times, improving ROUGE-1, ROUGE-2, and ROUGE-L over lead-truncated BERT baselines.
Limitation
#lexical-overlap-graph-synonym-insensitivity-limitation
Because the sentence similarity graph relies on exact surface word overlap normalized by log sentence length, sentences that express core document themes using synonyms without shared lexical tokens receive lower graph centrality scores.
⚠️ Curated Limitations & Negative Results
- Because the sentence similarity graph relies on exact surface word overlap normalized by log sentence length, sentences that express core document themes using synonyms without shared lexical tokens receive lower graph centrality scores.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
broad_match | DSGSUM is a neural text summarization architecture combining graph-based sentence centrality filtering with a BERT encoder and Transformer decoder. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r825-primary-pdf
|
Qiwei Bi, Haoyuan Li, Kun Lu, Hanfang Yang
|
primary paper
full_text
|
ID: https://arxiv.org/abs/2109.06046
sha256:4c51a2ada050…
|
r825-arxiv-abs
|
Qiwei Bi, Haoyuan Li, Kun Lu, Hanfang Yang
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/2109.06046
sha256:cc5c94fc9c39…
|
r825-metadata-cr
|
Qiwei Bi, Haoyuan Li, Kun Lu, Hanfang Yang
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/2109.06046
sha256:61e503dbc030…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T15:08:00Z
Updated: 2026-10-19T15:08:00Z
Source File:
data/curated/DSGSUMDocumentLevelSemanticGraphAbstractiveSummarization.yaml
· Schema:
curated_concept.yaml
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories