Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing News Recommendation with Hierarchical LLM Prompting

Hai-Dang Kieu, Delvin Ce Zhang, Minh Duc Nguyen, Min Xu, Qiang Wu, Dung D. Le

TL;DR

The paper tackles sparsity in short news titles by generating enriched titles ${T'}_i$ and entities ${E'}_i$ via LLMs. It introduces PNR-LLM, which combines hierarchical prompting (Direct Prompting, Exploration, Hierarchical Prompting) with an attention-based News Encoder and a User Representation module to produce unified embeddings. Experiments on the MIND dataset show state-of-the-art results for PNR-LLM and demonstrate that the data enrichment module is model-agnostic and boosts baselines when applied to them. This work demonstrates the value of incorporating LLM-generated semantic and entity information to mitigate data sparsity in personalized news recommendations and suggests future work integrating generated entities with knowledge graphs.

Abstract

Personalized news recommendation systems often struggle to effectively capture the complexity of user preferences, as they rely heavily on shallow representations, such as article titles and abstracts. To address this problem, we introduce a novel method, namely PNR-LLM, for Large Language Models for Personalized News Recommendation. Specifically, PNR-LLM harnesses the generation capabilities of LLMs to enrich news titles and abstracts, and consequently improves recommendation quality. PNR-LLM contains a novel module, News Enrichment via LLMs, which generates deeper semantic information and relevant entities from articles, transforming shallow contents into richer representations. We further propose an attention mechanism to aggregate enriched semantic- and entity-level data, forming unified user and news embeddings that reveal a more accurate user-news match. Extensive experiments on MIND datasets show that PNR-LLM outperforms state-of-the-art baselines. Moreover, the proposed data enrichment module is model-agnostic, and we empirically show that applying our proposed module to multiple existing models can further improve their performance, verifying the advantage of our design.

Enhancing News Recommendation with Hierarchical LLM Prompting

TL;DR

The paper tackles sparsity in short news titles by generating enriched titles and entities via LLMs. It introduces PNR-LLM, which combines hierarchical prompting (Direct Prompting, Exploration, Hierarchical Prompting) with an attention-based News Encoder and a User Representation module to produce unified embeddings. Experiments on the MIND dataset show state-of-the-art results for PNR-LLM and demonstrate that the data enrichment module is model-agnostic and boosts baselines when applied to them. This work demonstrates the value of incorporating LLM-generated semantic and entity information to mitigate data sparsity in personalized news recommendations and suggests future work integrating generated entities with knowledge graphs.

Abstract

Personalized news recommendation systems often struggle to effectively capture the complexity of user preferences, as they rely heavily on shallow representations, such as article titles and abstracts. To address this problem, we introduce a novel method, namely PNR-LLM, for Large Language Models for Personalized News Recommendation. Specifically, PNR-LLM harnesses the generation capabilities of LLMs to enrich news titles and abstracts, and consequently improves recommendation quality. PNR-LLM contains a novel module, News Enrichment via LLMs, which generates deeper semantic information and relevant entities from articles, transforming shallow contents into richer representations. We further propose an attention mechanism to aggregate enriched semantic- and entity-level data, forming unified user and news embeddings that reveal a more accurate user-news match. Extensive experiments on MIND datasets show that PNR-LLM outperforms state-of-the-art baselines. Moreover, the proposed data enrichment module is model-agnostic, and we empirically show that applying our proposed module to multiple existing models can further improve their performance, verifying the advantage of our design.
Paper Structure (13 sections, 2 equations, 4 figures, 4 tables)

This paper contains 13 sections, 2 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. METHODOLOGY
  3. Problem Formulation
  4. LLM-based News Enrichment
  5. News Encoder
  6. User Representation
  7. Experiments
  8. Empirical Evaluation
  9. Ablation Study
  10. Effect of Pre-trained Language Models for Text Encoding.
  11. Performance of Baseline Models Using Our Enriched Title.
  12. Other ablation study.
  13. Conclusion

Figures (4)

  • Figure 1: Illustration of news connected to a relevant news through entities explored by LLM.
  • Figure 2: Illustration of our PNR-LLM.
  • Figure 3: Effect of different (a) promptings and (b) LLMs.
  • Figure 4: Performance comparison of PNR-LLM with entities provided by MIND-SMALL and our method.