Generating weekly reports from multi-source news at operational scale is a high-throughput information processing task that requires scalable ingestion, evidence selection, redundancy control, temporal grounding, and auditable human revision. This paper presents a scalable parallel information system for evidence-grounded weekly report generation, using Association of Southeast Asian Nations (ASEAN)-related news as a case study. The system organizes web-based job orchestration, document-level parallel preprocessing, batched semantic encoding, constrained drafting, interactive revision, and version archiving into an evidence-first workflow. Its core algorithm layer consists of three coordinated modules: an entropy-adaptive relevance-diversity selector that combines SBERT semantic scores, TF–IDF lexical evidence, lexical uncertainty, and redundancy-aware Top-N filtering; a dual-evidence summarizer that fuses position-decayed TextRank, semantic centrality, and structured evidence cues; and a metadata–text temporal grounding module that separates current-week primary time from historical/background references. Experiments on 9989 one-week ASEAN news documents show that the proposed selector detects no Top-N near-duplicate pairs under <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\delta =0.92$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>δ</mi> <mo>=</mo> <mn>0.92</mn> </mrow> </math></EquationSource> </InlineEquation>, compared with 0.0068 for SBERT and 0.0222 for BM25. The fused summarizer maintains comparable-to-best coverage and achieves the highest recall for time, location, and actors, reaching 0.7393, 0.8034, and 0.7974, respectively. Under missing metadata, the temporal module maintains time completeness at 1.0000 and reduces historical leakage to 0.0000. A stage-level parallel runtime test further achieves a 3.6415<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> speedup with 8 workers for document-level preprocessing and structured exraction, demonstrating the system’s scalability, auditability, and suitability for continuous weekly news processing.

A scalable parallel information system for evidence-grounded weekly report generation from multi-source news

期刊论文
发表日期：

2026年7月3日

查看全文

Peng Gao,
Xi Qin,
Zhenrong Zhang

摘要

Generating weekly reports from multi-source news at operational scale is a high-throughput information processing task that requires scalable ingestion, evidence selection, redundancy control, temporal grounding, and auditable human revision. This paper presents a scalable parallel information system for evidence-grounded weekly report generation, using Association of Southeast Asian Nations (ASEAN)-related news as a case study. The system organizes web-based job orchestration, document-level parallel preprocessing, batched semantic encoding, constrained drafting, interactive revision, and version archiving into an evidence-first workflow. Its core algorithm layer consists of three coordinated modules: an entropy-adaptive relevance-diversity selector that combines SBERT semantic scores, TF–IDF lexical evidence, lexical uncertainty, and redundancy-aware Top-N filtering; a dual-evidence summarizer that fuses position-decayed TextRank, semantic centrality, and structured evidence cues; and a metadata–text temporal grounding module that separates current-week primary time from historical/background references. Experiments on 9989 one-week ASEAN news documents show that the proposed selector detects no Top-N near-duplicate pairs under $\delta =0.92$ $δ = 0.92$ , compared with 0.0068 for SBERT and 0.0222 for BM25. The fused summarizer maintains comparable-to-best coverage and achieves the highest recall for time, location, and actors, reaching 0.7393, 0.8034, and 0.7974, respectively. Under missing metadata, the temporal module maintains time completeness at 1.0000 and reduces historical leakage to 0.0000. A stage-level parallel runtime test further achieves a 3.6415 $\times $ $\times$ speedup with 8 workers for document-level preprocessing and structured exraction, demonstrating the system’s scalability, auditability, and suitability for continuous weekly news processing.