Hybrid Graph Attention with Adaptive Hierarchical Control for Scalable Bilingual Text Summarization

Abstract

For a text summarization model, there is a tradeoff among three aspects: factual accuracy, fluency, and scalability over large and diverse texts (including multilingual). This paper presents HGAHHS, a Hybrid Graph-Attention with Adaptive Hierarchical Hybrid Summarization. HGAHHS models sentences and their semantic inter-relationships into a heterogeneous graph, uses language-guided multi-head graph attention to learn a global document representation, performs hierarchical scoring of sentences and paragraphs, and incorporates an adaptive controller to dynamically select between extractive, hybrid, or abstractive summarization strategies according to document features (length, complexity, bilingual ratio, and graph density). It also has confidenceweighted fusion and explicit cross-lingual alignment mechanisms for robust bilingual and code-mixed summarization. We evaluate HGAHHS on standard English benchmarks (CNN/DailyMail, XSum), the public XL-Sum Hindi benchmark, and the newly constructed Hindi-XL cross-lingual/code-mixed corpus. HGAHHS outperforms strong extractive (TextRank), transformer-based (BART, LED), and multilingual (mBART-50, IndicBART) baselines on ROUGE-1/2/L, BERTScore, and human evaluations for fluency, factual consistency, and scalability in bilingual settings.