Pith. sign in

REVIEW 4 major objections 9 minor 1 cited by

Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation

T0 review · 4 major / 9 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Hierarchical aggregation of all reviews into a compact profile, plus fast pseudo-document retrieval, improves LLM-generated recommendation explanations by up to 12.6% while cutting inference retrieval to under one second.

desk verdict A sensible profiling-and-retrieval framework for LLM-based explainable recommendation, but the evaluation likely leaks the ground-truth review into profiles and retrieval, so the headline gains are suspect. read the letter →

arxiv 2507.09188 v1 pith:YHZCAM3R submitted 2025-07-12 cs.IR

Bangcheng Sun , Yazhe Chen , Jilin Yang , Xiaodong Li , Hui Li This is my paper
classification cs.IR
keywords explainable-recommendationlarge-language-modelshierarchical-aggregationretrieval-augmented-generationpseudo-documentsrecommendation-explanationretrieval-efficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes REXHA, a framework for generating recommendation explanations with large language models. It claims that a hierarchical aggregation module, which recursively summarizes all of a user's and an item's reviews into a compact profile, removes the profile deviation caused by random review sampling in prior systems such as XRec and G-Refer. On top of that profile, two pseudo-document queries—a latent representation query and a profile query—retrieve relevant reviews to ground the generated explanation. Across Yelp, Amazon-books, and Google-reviews, the authors report that REXHA improves explanation quality by up to 12.6% over baselines and keeps inference-time retrieval under one second, compared with over four minutes for G-Refer. If the claim holds, it means that holistic review compression plus targeted retrieval is a practical recipe for LLM explainability without sacrificing speed.

What carries the argument

The central objects are the hierarchical aggregation tree and the pseudo-document query pair. In the tree, each raw review is a leaf; the LLM merges two adjacent leaves into a 50-word summary, and this merge repeats level by level to a root profile for the user and for the item, so all reviews contribute within context limits. The retrieval module then builds two queries: a latent representation query, the averaged embedding of all summarized reviews of the user and item, which retrieves diverse opinions; and a profile query, the profile text itself, aligned to review embeddings through contrastive fine-tuning, which retrieves semantically concentrated opinions. A GCN (LightGCN) supplies user and item embeddings that are projected into the LLM's input space, and the final prompt concatenates profiles, embeddings, and retrieved reviews.

What would settle it

Run the same LLaMA-2-7B generator on a long-context model such as Qwen2.5-7B-1M and compare explanations produced from (a) the hierarchical profile, (b) all raw reviews concatenated directly, and (c) a random sample with the same token budget as the profile. If (b) or (c) matches or beats (a) on BERTScore and human judgment, the hierarchical compression is not the source of the improvement.

Watch

Extended reading notes

Core claim

REXHA's central discovery, on the paper's own terms, is that holistic profiles built by recursive LLM summarization plus retrieved review evidence generate recommendation explanations that beat random-sampling and graph-retrieval baselines on the datasets studied. The authors report the largest gain of 12.6% in BERTPrecision for the latent-query variant on Amazon-books, and under-one-second inference retrieval across all three datasets. The hierarchy compresses every available review into a single root profile, so the final prompt carries information no prior method feeds the LLM, while the retrieval module supplies supporting opinions from similar users and items.

Load-bearing premise

The recursive 50-word summarization at every level preserves all user-preference and item-characteristic details needed for the final explanation; if an intermediate summary drops a key preference, the profile will mislead the generator and the reported gains over random sampling could disappear.

Editorial extensions

If this is right

  • On the three public datasets, the latent-query variant lifts BERTPrecision over XRec by 12.6%, 26.8%, and 7.46%, and both REXHA variants beat G-Refer on BERTF1.
  • Inference-time retrieval runs in under one second per interaction, versus over four minutes for G-Refer, because REXHA uses cosine search over a precomputed vector store instead of CPU-bound path retrieval.
  • The ablation shows that neither module alone gives the full benefit: retrieval without hierarchical profiles hurts quality, while profiles without retrieval help only slightly; the combined system yields the best BERTF1.
  • Compared with the latent representation query, the profile query retrieves reviews that are more semantically similar to one another, aligning with long-term user and item preferences, while the latent query yields more diverse evidence.
  • The hierarchical preprocessing takes up to 20 hours per dataset, but runs offline and stays below G-Refer's more than 40 hours of training-time retrieval.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Applied more generally, the same recursive merge scheme is a recipe for distilling any review or document collection that exceeds an LLM's context window into a single faithful summary, and the contrastive query-tuning step shows how to make such summaries usable as search queries.
  • A testable extension is to expose intermediate-level summaries to the generator or to re-rank retrieved reviews by their agreement with the profile; the paper names these as future work but reports no experiments for them.
  • The efficiency claim hinges on profile construction being offline; a live recommender with continuously arriving reviews would need incremental updates to the hierarchy, which the paper does not address.
  • Replacing the LLM summarizer with a cheaper extractive method at the lower levels would isolate whether the quality gain comes from LLM semantics or simply from the hierarchical structure, and it could cut the 20-hour preprocessing cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. This manuscript introduces REXHA, a retrieval-augmented LLM framework for generating recommendation explanations. REXHA builds user and item profiles by hierarchically summarizing all reviews with an LLM (Sec. 2.2), retrieves supporting review summaries using latent and profile pseudo-document queries (Sec. 2.3), and feeds the profiles, retrieved reviews, and LightGCN collaborative embeddings to a frozen LLaMA-2-7B generator (Secs. 2.1 and 2.4). Experiments on Amazon-books, Yelp, and Google-reviews compare against NRT, Att2Seq, PETER, PEPLER, XRec, and G-Refer using GPTscore and BERT-family metrics, with an additional efficiency comparison. The paper claims up to 12.6% improvement in explanation quality and sub-second inference retrieval.

Significance. If the reported results survive a leakage-free evaluation, the hierarchical aggregation module is a sensible response to the profile-deviation problem in random-sampling baselines, and pseudo-document queries offer a practical alternative to G-Refer's graph-path retrieval. The paper contains extensive experiments, ablations, and an explicit efficiency analysis. However, the central empirical claim is currently jeopardized by the apparent inclusion of the target ground-truth review in both the profiles and the retrieval pool; this must be resolved before the numbers can be interpreted as evidence of improved explanation generation. The paper also does not release code, which limits reproducibility.

major comments (4)
  1. [Sec. 2.2.1, Sec. 2.3.2, and Fig. 8] The target ground-truth review is not excluded from the inputs, creating a test-set leakage risk that is load-bearing for the central claim. For a target user-item pair (u, i), the ground-truth explanation is u's review of i. When constructing the user profile, Sec. 2.2.1 states that 'the raw review comes from one item interacted by that user,' which includes the target item i; when constructing the item profile, 'we use this item's reviews as raw reviews,' which includes u's review. In Sec. 2.3.2, Eqs. (4)-(8) build the latent query by averaging embeddings over all reviews of u and i, so the gold review also shapes the retrieval query, and the candidate pool from Sec. 2.2.1 includes its summary. The paper neither states a masking rule nor uses a temporal split that would remove the target review. Consequently, REXHA may be given a compressed version of the label, and the reported advantage over XRec and G-Refer, which randomly sample only a few reviews, may reflect more complete access to the gold explanation rather than a genuine improvement in explanation generation. The authors must re-run the experiments with the target review removed from profile construction, retrieval candidate pools, and query construction, and report whether the gains persist.
  2. [Abstract and Table 1] The claim that 'our method outperforms existing approaches by up to 12.6% w.r.t. the explanation quality' is overstated. On Google-reviews, both REXHA variants have lower GPTscore than G-Refer (70.35 and 69.91 vs. 71.47), and on Amazon-books, REXHA-L's GPTscore of 81.44 is below both XRec (82.57) and G-Refer (82.70). The 12.6% figure is specific to BERTPrecision on Amazon-books; even the 26.8% improvement mentioned in Sec. 3.2 is for BERTPrecision on Yelp only. The abstract should be qualified to state that gains are metric- and dataset-specific, or the reported results should be summarized with the caveat that no single method dominates on all metrics.
  3. [Appendix D and Sec. 3.1.4] The validation of the hierarchical aggregation (HA) module is performed only with Qwen2.5-7B and Qwen2.5-7B-1M in Appendix D, while the main generation experiments use LLaMA-2-7B as the base model. The implementation details do not specify which LLM is used for the hierarchical summarization steps in the main experiments. If the summarizer in the main experiments is LLaMA-2-7B but the HA validation is only with Qwen2.5-7B, then the claimed benefit of HA over direct summarization is not established for the actual configuration tested. The authors should either run the HA validation with the same LLM used in the main experiments or clearly state that the summarizer is Qwen2.5-7B in both settings and justify the transfer.
  4. [Table 1 and Sec. 3.2] The paper reports only point estimates and standard deviations, with no significance tests or confidence intervals. Many of the claimed improvements are small relative to the reported standard deviations; for example, on Google-reviews the BERTF1 improvement over G-Refer is 0.15% (0.4573 vs. 0.4566), and several GPTscore differences are comparable to the reported std. Without paired significance tests or error bars across repeated runs, the phrase 'outperforms' is not supported for these close comparisons. The authors should add statistical testing or at least bootstrap confidence intervals for the key comparisons.
minor comments (9)
  1. [Sec. 2.2.2] The symbol p is used both for the group size in hierarchical aggregation ('concatenate every p summaries') and for the number of retrieved reviews in Sec. 2.3.1 and Sec. 3.4. This is confusing; a different symbol, such as k, should be used for one of the two quantities.
  2. [Eq. (2)] The second equation uses K in the summation upper limit while the first uses L; presumably both should be L.
  3. [Sec. 3.4] The sentence 'we can observe that on both Yelp, BERTprecision score consistently increases as p grows' is ungrammatical and ambiguous; it should say 'on Yelp' or 'on both Yelp and Google-reviews' depending on the intended scope.
  4. [Figures 2 and 3] The figure captions include corrupted escape sequences such as '/uni00000015/uni00000017...' that must be cleaned before publication.
  5. [Table 2] The ablation study reports only BERTPrecision and BERTF1, omitting GPTscore and BERTRecall, which makes it hard to assess the contribution of each module on the headline metric. The statement that the worst results are 'in RED' is also not visible in the rendered table.
  6. [Appendix B] The metric name 'BAR Tscore' should be 'BERTscore', and the paragraph describing it should use consistent terminology with Sec. 3.1.2.
  7. [Sec. 3.1.4] The model name 'LigntGCN' is a typo for 'LightGCN'.
  8. [Reference [23]] Reference [23] is cited for the llm-embedder model, but the title 'Interpretable unified language checking' does not match the expected llm-embedder paper; the reference should be checked and corrected.
  9. [Appendix F, Fig. 8] The generation prompt in Fig. 8 reads 'Explain why the user would buy with the book within 50 words,' which is awkward and inconsistent with the book/item terminology used for the Yelp and Amazon datasets; this should be edited for clarity.

Circularity Check

1 steps flagged · score 6.0 of 10

The gold explanation (the target user's review of the target item) is included in both the hierarchical profiles and the retrieval query by construction, so the reported gains may reflect direct label access rather than an independent prediction.

  1. self definitional [Sec. 2.2.1, Sec. 2.2.2, Sec. 2.3.2 (Eq. 4-8), and Fig. 8 prompt]
    "At the bottom of the hierarchical aggregation tree, each leaf contains one raw item review and each review is randomly placed in a leaf node. When constructing a user profile, the raw review comes from one item interacted by that user. When constructing an item profile, we use this item's reviews as raw reviews. ... REXHA encodes all raw reviews of the target user and item using the embedding model f (semantic encoder) and then aggregates all the encoded representations as the latent query."

    The benchmark's ground-truth explanation for a target pair (u, i) is u's own review of i. Under the quoted construction, that review is simultaneously 'one item interacted by that user' and one of 'this item's reviews', so it is summarized into the user and item profiles in Sec. 2.2.2. It also contributes its embedding to q_latent through Eq. (4)-(8), since those equations average over all raw reviews of the target user and item. Retrieval by cosine similarity over all raw review summaries (Sec. 2.3.1) can then return the gold review as a top-q 'relevant review', and the Fig. 8 prompt feeds these profiles and retrieved summaries directly into the generator. The paper states no masking rule, no exclusion of the target review, and no temporal split that would remove it.

full rationale

The paper's own derivation chain is self-contained with respect to its equations: hierarchical aggregation summarizes raw reviews, the latent query is a mean of review embeddings, retrieval uses cosine similarity, and the generator conditions on profiles plus retrieved summaries. No fitted parameter is renamed as a prediction, and no self-citation or imported uniqueness theorem carries the argument. However, the central evaluation does exhibit a construction-level circularity risk: the target's ground-truth explanation is one of the raw reviews that the pipeline explicitly summarizes and retrieves. Because Sec. 2.2.1 defines user-profile raw reviews as 'one item interacted by that user' and item-profile raw reviews as 'this item's reviews', and Sec. 2.3.2 averages over 'all raw reviews of the target user and item', the gold review is included in both profiles and the query unless an unstated filter removes it. The prompt in Fig. 8 then supplies those profiles and the retrieved summaries to the LLM. The paper never states that the target review is excluded, so the reported 12.6% improvement over XRec and G-Refer could be driven by the model having access to a compressed version of the answer. This is a single, specific, and load-bearing reduction of the prediction to its input, warranting a score of 6 rather than a higher score because the generation step still involves nontrivial LLM text production and the paper is not otherwise circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No invented entities. Free parameters are routine hyperparameters rather than quantities fitted to force the result. The axioms are domain assumptions about summarization faithfulness and embedding alignment, which are plausible but not proven for the main model.

free parameters (4)
  • p (number of retrieved reviews) = 2 to 15 in sensitivity analysis (Fig. 2), best varies by dataset
    Chosen by hand; affects retrieval augmentation quality.
  • Summary word limit = 50 words
    Set in the prompt templates (Appendix F); hand-chosen limit for LLM summarization.
  • Hierarchical group size = 4 reviews per summarization set
    Stated in Sec. 3.1.4; hand-chosen.
  • Contrastive temperature tau = not reported
    Used in Eq. 9; no value or tuning details are given.
assumptions (4)
  • domain assumption LLM summarization at each hierarchy level preserves key preference information
    Load-bearing for HA; assumed in Sec. 2.2.2, tested only indirectly in Appendix D with Qwen.
  • domain assumption The llm-embedder model represents review summaries in a space aligned with user/item profiles
    Relied upon in Sec. 2.3.1; fine-tuned with contrastive loss, but transfer to LLaMA-2-7B generation is assumed.
  • domain assumption LightGCN collaborative embeddings help the LLM generate better explanations
    Used as in XRec; Sec. 2.1, no ablation isolating this component.
  • standard math Cosine similarity is a valid relevance measure for review retrieval
    Used in Eq. 9 and retrieval; standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation." pith.science (2026). https://pith.science/paper/YHZCAM3R

@misc{pith2026250709188,
  author       = {Pith},
  title        = {Pith review of: Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YHZCAM3R}},
  note         = {Machine review of arXiv:2507.09188}
}
read the original abstract

Explainable Recommender System (ExRec) provides transparency to the recommendation process, increasing users' trust and boosting the operation of online services. With the rise of large language models (LLMs), whose extensive world knowledge and nuanced language understanding enable the generation of human-like, contextually grounded explanations, LLM-powered ExRec has gained great momentum. However, existing LLM-based ExRec models suffer from profile deviation and high retrieval overhead, hindering their deployment. To address these issues, we propose Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation (REXHA). Specifically, we design a hierarchical aggregation based profiling module that comprehensively considers user and item review information, hierarchically summarizing and constructing holistic profiles. Furthermore, we introduce an efficient retrieval module using two types of pseudo-document queries to retrieve relevant reviews to enhance the generation of recommendation explanations, effectively reducing retrieval latency and improving the recall of relevant reviews. Extensive experiments demonstrate that our method outperforms existing approaches by up to 12.6% w.r.t. the explanation quality while achieving high retrieval efficiency.

Figures

Figures reproduced from arXiv: 2507.09188 by the authors.

Figure 1
Figure 1. Overview of REXHA. It contains three key components:(1) Collaborative Signal Extrac￾tion provides collaborative filtering information to LLMs. (2) Hierarchical Aggregation compresses and summarizes reviews layer by layer, finally construct textual profiles for user/item. (3) Review Retrieval module retrieves relative reviews to enhance LLM generating explanations. complementing later LLM-based explanation generation… view at source ↗
Figure 3
Figure 3. Comparisons of efficiency between REXHA and G-Refer. observe that REXHA-P with profile query achieves higher GPT scores compared to REXHA-L across all three datasets, with improvements of 2.26%, 2.60%, and 0.64% on Amazon-books, Yelp, and Google-reviews, respectively. However, it performs slightly worse in terms of BERT scores. 3.3 Ablation Study We provide the ablation study in Tab. 2. It is observed that the hiera… view at source ↗
Figure 5
Figure 5. Similarity of each reviews retrieved by profile query. F Prompt Template [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: demonstrates how LLMs generate item profiles on Yelp by combining item metadata and user reviews. This enables the model to identify key user characteristics associated with the item, offering deeper insights into user preferences and improving the understanding of use…
Figure 7
Figure 7. Figure 7: Prompt for User Profile System prompt: Explain why the user would buy with the book within 50 words. User prompt: Here are some comments from similar users and items, you can use them to help you write the review. 1. The user would enjoy the business because of the ...…
Figure 8
Figure 8. Figure 8: Prompt for Reviews Summarization. G Case Study We present two cases in Tab. 5 and Tab. 6 to demonstrate the effectiveness of our method and illustrate how hierarchical aggregation based profile and retrieved reviews benefit the generated explanations. The table only sh…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

    cs.IR 2026-07 accept novelty 5.0 of 10

    Agentic recommender systems are organized by agent role (assisted, as-recommender, as-simulator) crossed with autonomy levels L2–L5, yielding a roadmap of architectures, evaluation limits, and open challenges.

Reference graph

Works this paper leans on

42 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    Herlocker, and John Riedl

    Al Borchers, Jonathan L. Herlocker, and John Riedl. Ganging up on information overload. Computer, 31(4):106–108, 1998

    work page 1998

  2. [2]

    A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation.IEEE Trans

    Le Wu, Xiangnan He, Xiang Wang, Kun Zhang, and Meng Wang. A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation.IEEE Trans. Knowl. Data Eng., 35(5):4425–4445, 2023

    work page 2023

  3. [3]

    Recommender Systems Handbook

    Francesco Ricci, Lior Rokach, and Bracha Shapira, editors. Recommender Systems Handbook. Springer US, 2022

    work page 2022

  4. [4]

    Explainable recommendation: A survey and new perspectives

    Yongfeng Zhang and Xu Chen. Explainable recommendation: A survey and new perspectives. F ound. Trends Inf. Retr ., 14(1):1–101, 2020

    work page 2020

  5. [5]

    A comparative analysis of text-based explainable recommender systems

    Alejandro Ariza-Casabona, Ludovico Boratto, and Maria Salamó. A comparative analysis of text-based explainable recommender systems. In RecSys, pages 105–115, 2024

    work page 2024

  6. [6]

    Neural rating regression with abstractive tips generation for recommendation

    Piji Li, Zihao Wang, Zhaochun Ren, Lidong Bing, and Wai Lam. Neural rating regression with abstractive tips generation for recommendation. In SIGIR, pages 345–354, 2017

    work page 2017

  7. [7]

    Graph-based extractive explainer for recommen- dations

    Peng Wang, Renqin Cai, and Hongning Wang. Graph-based extractive explainer for recommen- dations. In WWW, pages 2163–2171, 2022

    work page 2022

  8. [9]

    Unsupervised extractive summarization-based representations for accurate and explainable collaborative filtering

    Reinald Adrian Pugoy and Hung-Yu Kao. Unsupervised extractive summarization-based representations for accurate and explainable collaborative filtering. In ACL/IJCNLP, pages 2981–2990, 2021

    work page 2021

Show all 42 references
  1. [10]

    Recexplainer: Aligning large language models for explaining recommendation models

    Yuxuan Lei, Jianxun Lian, Jing Yao, Xu Huang, Defu Lian, and Xing Xie. Recexplainer: Aligning large language models for explaining recommendation models. In KDD, pages 1530–1541, 2024

    2024

  2. [11]

    Xrec: Large language models for explainable recommendation

    Qiyao Ma, Xubin Ren, and Chao Huang. Xrec: Large language models for explainable recommendation. In EMNLP (Findings), pages 391–402, 2024

    2024

  3. [12]

    G-refer: Graph retrieval-augmented large language model for explainable recommendation

    Yuhan Li, Xinni Zhang, Linhao Luo, Heng Chang, Yuxiang Ren, Irwin King, and Jia Li. G-refer: Graph retrieval-augmented large language model for explainable recommendation. In WWW, pages 240–251, 2025

    2025

  4. [13]

    Graph neural networks in recom- mender systems: A survey

    Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, and Bin Cui. Graph neural networks in recom- mender systems: A survey. ACM Comput. Surv., 55(5):97:1–97:37, 2023

    2023

  5. [14]

    Lightgcn: Simplifying and powering graph convolution network for recommendation

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. In SIGIR, pages 639–648, 2020

    2020

  6. [15]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Trans. Assoc. Comput. Linguistics, 12:157–173, 2024

    2024

  7. [16]

    Jiacheng Li, Jingbo Shang, and Julian J. McAuley. Uctopic: Unsupervised contrastive learning for phrase representations and topic mining. In ACL, pages 6159–6169, 2022

    2022

  8. [17]

    Personalized showcases: Generating multi-modal explanations for recommendations

    An Yan, Zhankui He, Jiacheng Li, Tianyang Zhang, and Julian John McAuley. Personalized showcases: Generating multi-modal explanations for recommendations. In SIGIR, pages 2251–2255, 2023

    2023

  9. [18]

    Is chatgpt a good NLG evaluator? A preliminary study

    Jiaan Wang, Yunlong Liang, Fandong Meng, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. Is chatgpt a good NLG evaluator? A preliminary study. arXiv Preprint, 2023. URL https://arxiv.org/abs/2303.04048. 10

    2023 arXiv

  10. [19]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with BERT. In ICLR, 2020

    2020

  11. [20]

    Learning to generate product reviews from attributes

    Li Dong, Shaohan Huang, Furu Wei, Mirella Lapata, Ming Zhou, and Ke Xu. Learning to generate product reviews from attributes. In EACL, pages 623–632, 2017

    2017

  12. [21]

    Personalized transformer for explainable recommenda- tion

    Lei Li, Yongfeng Zhang, and Li Chen. Personalized transformer for explainable recommenda- tion. In ACL/IJCNLP, pages 4947–4957, 2021

    2021

  13. [22]

    Personalized prompt learning for explainable recommen- dation

    Lei Li, Yongfeng Zhang, and Li Chen. Personalized prompt learning for explainable recommen- dation. ACM Trans. Inf. Syst., 41(4):103:1–103:26, 2023

    2023

  14. [23]

    Tianhua Zhang, Hongyin Luo, Yung-Sung Chuang, Wei Fang, Luc Gaitskell, Thomas Hartvigsen, Xixin Wu, Danny Fox, Helen Meng, and James R. Glass. Interpretable unified language checking. arXiv Preprint, 2023. URL https://arxiv.org/abs/2304.03728

    2023 arXiv

  15. [24]

    Towards self-explaining sequence-aware recommendation

    Alejandro Ariza-Casabona, Maria Salamó, Ludovico Boratto, and Gianni Fenu. Towards self-explaining sequence-aware recommendation. In RecSys, pages 904–911, 2023

    2023

  16. [25]

    Explainable recommendation with personalized review retrieval and aspect learning

    Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. Explainable recommendation with personalized review retrieval and aspect learning. In ACL, pages 51–64, 2023

    2023

  17. [26]

    Huijing Zhan, Ling Li, Shaohua Li, Weide Liu, Manas Gupta, and Alex C. Kot. Towards explainable recommendation via bert-guided explanation generator. In ICASSP, pages 1–5, 2023

    2023

  18. [27]

    Prompt distillation for efficient llm-based recommenda- tion

    Lei Li, Yongfeng Zhang, and Li Chen. Prompt distillation for efficient llm-based recommenda- tion. In CIKM, pages 1348–1357, 2023

    2023

  19. [28]

    In-context retrieval-augmented language models

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. Trans. Assoc. Comput. Linguistics, 11:1316–1331, 2023

    2023

  20. [29]

    Self-knowledge guided retrieval augmentation for large language models

    Yile Wang, Peng Li, Maosong Sun, and Yang Liu. Self-knowledge guided retrieval augmentation for large language models. In EMNLP (Findings), pages 10303–10315, 2023

    2023

  21. [30]

    Precise zero-shot dense retrieval without relevance labels

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. Precise zero-shot dense retrieval without relevance labels. In ACL, pages 1762–1777, 2023

    2023

  22. [31]

    Directly

    Shuting Wang, Xin Yu, Mang Wang, Weipeng Chen, Yutao Zhu, and Zhicheng Dou. Richrag: Crafting rich responses for multi-faceted queries in retrieval-augmented generation. InCOLING, pages 11317–11333, 2025. 11 A Datasets Table 3: Statistics of the experimental datasets Datasets ...

    2025

  23. [32]

    name": "the name of the business

    The basic information will be described in JSON format, with the following attributes: { "name": "the name of the business" }

  24. [33]

    the first review

    Reviews from users will be managed in the following List format: [ "the first review", "the second review", .... ]

  25. [34]

    USER REVIEWS: a List object containing some reviews from users about the business

    The information I will give you: BASIC INFORMATION: a JSON string describing the basic information about the business. USER REVIEWS: a List object containing some reviews from users about the business. Requirements:

  26. [35]

    summarization

    Please provide your answer in JSON format, following this structure: { "summarization": "A summarization of what types of users would enjoy this business." (if you are unable to summarize it, please set this value to "None") }

  27. [37]

    summarization

    Do not provide any other text outside the JSON string. Output: { "summarization": "Users who appreciate ... would enjoy this business." } Figure 6: Prompt for Item Profile Fig. 7 presents a method for generating user profiles on the Yelp dataset using LLMs. By combining variou...

  28. [38]

    title":

    Each interacted business will be described in DICTIONARY format, with the following attributes: { "title": "the name of the business", (if there is no business, I will set this value to "None") "description": "a description of what types of users will like this business", "rev...

  29. [39]

    Requirements:

    The information I will give you: PURCHASED BUSINESSES: a list of dictionaries describing the businesses that the user has interacted with. Requirements:

  30. [40]

    summarization

    Please provide your answer in JSON format, following this structure: { "summarization": "A summarization of what types of business this user is likely to enjoy." (if you are unable to summarize it, please set this value to "None") }

  31. [41]

    summarization

    Please ensure that the "summarization" is no longer than 50 words

  32. [42]

    summarization

    Do not provide any other text outside the JSON string. Output: { "summarization": "This user enjoys ... experiences, ... service, ... atmospheres." } Figure 7: Prompt for User Profile System prompt: Explain why the user would buy with the book within 50 words. User prompt: Her...

  33. [43]

    The user would enjoy the business because of the

  34. [44]

    high quality seafood and steak

    This business would be enjoyed by the user ... ... user record: <USER_EMBED> item record: <ITEM_EMBED> item name: … user profile: … item profile: … Output: The user would enjoy the business because of... Figure 8: Prompt for Reviews Summarization. G Case Study We present two c...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.