Research map / Large language models / Language, safety and interpretability

Coherency Improved Explainable Recommendation via Large Language Model

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CIER, which fine-tunes LLaMA2-7B with LoRA to predict a rating, embeds the soft rating into the LLM word-embedding space, and generates a rating-aware explanation, and evaluates rating-explanation coherence with GPT-4o and a BERT sentiment model.

Problem

Explainable recommendation systems often produce explanations that are incoherent with the predicted ratings, since rating prediction and explanation generation are trained as separate heads in a multi-task setup; existing coherence evaluation also relies on costly annotations.

Approach

Use an LLM to predict a rating first, convert it into a soft-rating word embedding (SR2WE), and condition explanation generation on that embedding in a pipeline fashion, combined with rating smoothing, curriculum learning between keyword and explanation generation, and an annotation-free coherence metric built from GPT-4o and a sentiment classifier.

Key results

• On Yelp, CIER (curriculum learning) reaches FCR 53.84 and USR 32.63 versus 21.04 and 20.32 for the best prior baseline PEPLER.
• On Amazon, CIER improves FMR to 12.45 and BLEU-4 to 1.15 relative to the strongest baseline values in the table.
• Coherence on GPT-4o evaluation rises to 90.2/89.8/91.6 (Yelp/Amazon/TripAdvisor) versus 88.8/80.0/90.4 for CER; human annotation gives 73.5/74.0/87.0 for CIER.

Measured effects reported in the paper

Surprising findings

• Two-stage learning fails to improve explainability on Yelp, which the authors attribute to knowledge forgetting on the large dataset.
• Traditional NRT outperforms CIER on RMSE/MAE for TripAdvisor, the smaller and sparser dataset.

Limitations

• Automatic coherence metric depends on a sentiment classifier and GPT-4o judgments rather than ground-truth coherence labels.
• Curriculum learning does not beat baselines on FMR for Amazon, attributed to many single-occurrence keywords.
• Human annotation covers only 100 samples per dataset from two annotators.

Experimental setup

Datasets: Yelp Amazon Movies & TV TripAdvisor

Benchmarks: Yelp Amazon Movies & TV TripAdvisor

Metrics: RMSE MAE BLEU-1 BLEU-4 ROUGE-1 ROUGE-2 ROUGE-L FMR FCR DIV

Baselines: NRT Att2Seq PETER CER PEPLER ERRA SVD++ DeepCoNN NARRE

Builds on: CER PETER PEPLER LoRA

Closest papers by meaning

• Fine-Tuning Large Language Model Based Explainable Recommendation with Explainable Quality Reward (AAAI 2024 · similarity 0.88)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.87)
• Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.87)
• Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.86)
• LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.86)
• CoT4Rec: Revealing User Preferences Through Chain of Thought for Recommender Systems (AAAI 2025 · similarity 0.85)