Scholarlink

<p>Personalized summarization seeks to tailor the summary content to an individual user’s goals, preferences, or learning needs, offering a more adaptive alternative to generic, one-size-fits-all summaries. However, current summarization systems largely neglect explicit modeling of user intent, resulting in limited alignment between generated summaries and actual user expectations. This issue is especially critical in educational contexts, where learners may differ significantly in learning objectives, prior knowledge, or preferred granularity of information. In this paper, we propose a novel framework for intent-aware personalized summarization, designed for bilingual educational settings (English and Chinese), powered by moderately sized Large Language Models (LLMs) in the 7–9 billion parameter range. We formulate personalized summarization as the task of generating a factually consistent and concise summary of an input document, guided by an explicit, structured user intent representation. To address this challenge, we introduce a hybrid architecture that integrates a user intent encoder with a pointer-generator Transformer built atop LLM backbones. Our system features a two-stage generation strategy: (1) a content selector module retrieves segments aligned with user intent; (2) a guided decoder generates coherent summaries conditioned on both content and intent embeddings. We evaluate our method on two publicly available educational datasets (EduLect-Sum for English, EduCN-Sum for Chinese), and compare against 10+ strong baselines, including both traditional (e.g., T5, BART, Pegasus) and modern LLM-based summarizers. Empirical results show that our approach consistently outperforms baselines across four standard metrics (ROUGE, BLEU, BERTScore, FactCC) and human evaluations, achieving 6–8 point ROUGE gains, 5+% BERTScore and factuality improvements, and higher user satisfaction. Further ablation studies and cross-lingual generalization experiments validate the efficacy of explicit intent modeling. Our findings highlight the value of user-intent conditioning in educational summarization and demonstrate the feasibility of building compact, bilingual, and user-centric summarization systems with LLMs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Intent-aware personalized summarization for educational texts with large language models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月11日

查看全文
  • Wenyu Ning,
  • Shaotao Shi,
  • Yuchen Yin,
  • Zixuan Guo,
  • Jinrong Xie

摘要

Personalized summarization seeks to tailor the summary content to an individual user’s goals, preferences, or learning needs, offering a more adaptive alternative to generic, one-size-fits-all summaries. However, current summarization systems largely neglect explicit modeling of user intent, resulting in limited alignment between generated summaries and actual user expectations. This issue is especially critical in educational contexts, where learners may differ significantly in learning objectives, prior knowledge, or preferred granularity of information. In this paper, we propose a novel framework for intent-aware personalized summarization, designed for bilingual educational settings (English and Chinese), powered by moderately sized Large Language Models (LLMs) in the 7–9 billion parameter range. We formulate personalized summarization as the task of generating a factually consistent and concise summary of an input document, guided by an explicit, structured user intent representation. To address this challenge, we introduce a hybrid architecture that integrates a user intent encoder with a pointer-generator Transformer built atop LLM backbones. Our system features a two-stage generation strategy: (1) a content selector module retrieves segments aligned with user intent; (2) a guided decoder generates coherent summaries conditioned on both content and intent embeddings. We evaluate our method on two publicly available educational datasets (EduLect-Sum for English, EduCN-Sum for Chinese), and compare against 10+ strong baselines, including both traditional (e.g., T5, BART, Pegasus) and modern LLM-based summarizers. Empirical results show that our approach consistently outperforms baselines across four standard metrics (ROUGE, BLEU, BERTScore, FactCC) and human evaluations, achieving 6–8 point ROUGE gains, 5+% BERTScore and factuality improvements, and higher user satisfaction. Further ablation studies and cross-lingual generalization experiments validate the efficacy of explicit intent modeling. Our findings highlight the value of user-intent conditioning in educational summarization and demonstrate the feasibility of building compact, bilingual, and user-centric summarization systems with LLMs.