Research map / Large language models / Retrieval and knowledge
ESPRESSO: An Effective Approach to Passage Retrieval for High-Quality Conversational Recommender Systems
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 ESPRESSO, a passage retrieval module for conversational recommender systems that uses adaptive item selection and relevance-based groupwise learning to improve factuality of REC-responses.
Problem
Conversational recommender systems generate factually incorrect explanations (hallucinations), and existing passage retrieval methods are not designed to retrieve passages aligned with user-preferred items and lack supervision.
Approach
Adaptively selects a variable number of items from the recommender based on cumulative confidence, and trains the retriever with groupwise contrastive loss over subgroups of pseudo-relevant passages (ranked by relevance) to be robust to mislabeled passages.
Key results
- OURS+CoT-MAE reaches Hit@3 of 0.685 on DuRecDial2.0 vs 0.504 for the best competitor CoT-MAE, a gain of up to 35.91%.
- BART(ESPRESSO) improves BLEU4 over RAG by 29.17% on DuRecDial2.0 and 32.35% on KoRecDial (BLEU4 gains reported as 29.17% and 32.35% on the respective datasets).
- Adaptive item selection yields overall Hit@3 of 0.675 vs 0.648 for top-1 selection (4.2% gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive item selection instead of only the top-1 item (ablation) | ESPRESSO passage retrieval · DuRecDial2.0, Overall | Hit@3: 0.648 -> 0.675; gain of 4.2% | Helped |
| Use adaptive item selection instead of a fixed top-N item set (ablation) | ESPRESSO passage retrieval · DuRecDial2.0, Overall | Hit@3: 0.628 -> 0.675; gain of 7.5% | Helped |
| Use only the top-1 item rather than adaptive selection (ablation) | ESPRESSO passage retrieval · DuRecDial2.0, G=top-1 | Hit@3: 0.810 -> 0.815 | Helped |
| Use a fixed top-N item set rather than adaptive selection (ablation) | ESPRESSO passage retrieval · DuRecDial2.0, G=top-1 | Hit@3: 0.749 -> 0.810 | Hurt |
| Use a fixed top-N item set instead of only the top-1 item (ablation) | ESPRESSO passage retrieval · DuRecDial2.0, G̸=top-1 | Hit@3: 0.003 -> 0.160 | Helped |
| Replace individual-passage contrastive learning with grouped learning (ablation) | ESPRESSO passage retrieval · DuRecDial2.0 | Hit@3 | Helped |
| Group lower-ranked pseudo-relevant passages only with higher-ranked passages (ablation) | ESPRESSO passage retrieval · DuRecDial2.0 | Hit@3 | Helped |
| Add ESPRESSO to CoT-MAE passage retrieval | CoT-MAE · DuRecDial2.0 | Hit@3: 0.504 -> 0.685; gain of 35.91% | Helped |
| Add ESPRESSO to CoT-MAE passage retrieval | CoT-MAE · KoRecDial | Hit@3: 0.283 -> 0.315; gain of 11.31% | Helped |
| Use ESPRESSO-retrieved passages instead of CoT-MAE passages for BART response generation | BART · DuRecDial2.0 | BLEU4: 0.080 -> 0.093 | Helped |
| Use ESPRESSO-retrieved passages for BART response generation | BART · DuRecDial2.0 | BLEU4: 0.072 -> 0.093; gains of 29.17% | Helped |
| Use ESPRESSO-retrieved passages for BART response generation | BART · KoRecDial | BLEU4: 0.034 -> 0.045; gains of 32.35% | Helped |
Surprising findings
- CoT-MAE without ESPRESSO outperforms RankGPT, which uses GPT-4o reranking, by up to 19.76% on DuRecDial2.0, showing supervision matters more than LLM reranking here.
Limitations
- Recommendation module is out of scope and uses a fixed well-established approach rather than being evaluated.
- Relies on GPT-4o reranking for pseudo-label generation, which is costly and API-dependent.
- Generation module may not robustly handle noisy retrieved passages (acknowledged as future work).
Experimental setup
Datasets: DuRecDial2.0 KoRecDial
Benchmarks: DuRecDial2.0 KoRecDial
Metrics: Hit@1 Hit@3 Hit@5 BLEU2 BLEU3 BLEU4 GPTEval informativeness GPTEval relevance GPTEval fluency
Baselines: BM25 DPR RAG KERS DSI Contriever RankGPT CoT-MAE GPT2-base GPT2-large BART-base BART-large
Builds on: DPR Contriever CoT-MAE KERS RAG
Closest papers by meaning
- ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.85)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.85)
- FacetCRS: Multi-Faceted Preference Learning for Pricking Filter Bubbles in Conversational Recommender System (AAAI 2024 · similarity 0.85)
- Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.84)
- TRACE: Tourism Recommendation with Accountable Citation Evidence (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.