Pith. sign in

REVIEW 3 major objections 4 minor 58 references

A Framework for Generating Conversational Recommendation Datasets from Behavioral Interactions

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A framework that simulates realistic multi-turn conversational recommendation datasets from timestamped interactions and reviews, plus a cross-attention model that fuses both signals and beats either alone.

desk verdict Solid dataset-generation framework, but the headline fusion results are undermined by target leakage built into the dialog plan. read the letter →

arxiv 2506.17285 v1 pith:TWAZBVNX submitted 2025-06-14 cs.IR cs.LG

classification cs.IRcs.LG
keywords conversationalrecommendationdatasetgenerationLLMsimulationtemporalprofilingsemanticdialogplanningcross-attentionfusionbehavioralgrounding
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

The paper tries to close a gap: no large-scale dataset pairs users' long-term interaction histories with multi-turn conversational recommendation dialogs. It proposes ConvRecStudio, a three-stage LLM pipeline that builds temporal user and item profiles from reviews, plans a semantic dialog using a DAG of flexible intent nodes, and instantiates the plan with paired user and system agents. Applied to MobileRec, Yelp, and Amazon Electronics, it produces over 12,000 dialogs per domain that human raters and an LLM judge score high on naturalness, coherence, and grounding. To show the data matters, the paper trains a GPT-2-based cross-attention model that jointly encodes history and dialog and reports consistent Hit@K and NDCG@K gains, including a 10.9% Hit@1 improvement on Yelp over the strongest baseline.

What carries the argument

The load-bearing machinery is the three-stage pipeline: (1) Temporal Profiling uses contrastive spherical k-means with an InfoNCE objective on sentence embeddings to induce 20 aspect clusters, exponential time-decay user profiles, and sentiment-weighted global item profiles; (2) Semantic Dialog Planning represents each conversation as a DAG of super-nodes (Greet, AspectEx, Recommend, Close) whose vertices are (role, act, aspect) triples; (3) Multi-Turn Simulation instantiates the plan with two role-specialized LLM agents that must stay within frontier intents, with fallback resampling and post-hoc alignment checks. The evaluation model is a GPT-2 backbone with cross-attention between the dialog and history encoders at every layer.

What would settle it

A direct check is to regenerate dialogs with the ground-truth item removed from the user agent's instructions and with no guaranteed final acceptance, then retrain the cross-attention model: if Hit@1 gains over the Dialog-Only baseline largely disappear, the reported improvement was carried by target leakage built into the plan rather than by genuine history-dialog fusion.

Watch

Extended reading notes

Core claim

The central claim is that historical user-item interactions, not just conversation text, can anchor synthetic dialogs: given only timestamped reviews, the framework induces an aspect vocabulary, builds user and community item profiles, and uses a plan-constrained simulation so the system elicits preferences and recommends the ground-truth item without being told it. The paper further claims that a model which fuses long-term history with dialog context through cross-attention outperforms models using either signal alone or a naive concatenation, demonstrating that the two paradigms carry complementary information.

Load-bearing premise

The plan guarantees the system recommends the ground-truth item and the user eventually accepts it, so everything depends on these orchestrated conversations still representing how real users and systems would actually talk; if real interactions do not follow that shape, both the dataset's realism and the downstream model's gains are in question.

Editorial extensions

If this is right

  • If the dialogs are faithful to real behavior, ConvRecStudio can generate conversational recommendation data cheaply for any domain that has timestamped reviews, removing a bottleneck that manual dataset construction cannot scale past.
  • A unified recommender that cross-attends history and dialog should beat both a pure collaborative filter and a pure conversational model, because the two signals are complementary rather than redundant.
  • The generated datasets come with fine-grained aspect labels and negative recommendations, enabling evaluation of preference elicitation and contrastive reasoning, not just final item prediction.
  • The planned structure with flexible local ordering is meant to keep conversations coherent while preserving linguistic diversity, so the framework can produce varied dialogs from the same semantic scaffold.

Reading between the lines

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

  • The reported gains are measured against baselines trained on the same synthetic dialogs; the harder test is whether the fusion advantage transfers to real user-system dialogs, which the paper does not run.
  • Because the plan guarantees the target item is eventually recommended and accepted, a reader should check that downstream improvement is not partly leakage; a natural ablation would withhold the target from the plan and measure the drop.
  • A cheap extension would apply the same pipeline to implicit-feedback domains (clicks, watch time) with no review text, testing whether review-derived aspect induction is essential or can be replaced by item categories.
  • The aspect sets are induced once per corpus; for a new domain, a human check would be needed to confirm the induced aspects match how users actually talk about those items.
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

3 major / 4 minor

Summary. The paper presents ConvRecStudio, a pipeline that converts timestamped user-item interaction data into synthetic multi-turn conversational recommendation dialogs. Stage 1 induces aspect clusters from review sentences and builds time-decayed user profiles and community-level item sentiment profiles. Stage 2 samples informative interactions and constructs a DAG of dialog-act super-nodes (Greet, AspectEx, Recommend, Close) whose vertices include the ground-truth item as the recommendation target. Stage 3 instantiates the plan with paired user/system LLMs under plan-constrained utterance generation. The authors generate datasets for MobileRec, Yelp, and Amazon Electronics, report human and LLM-judge quality scores, and train a GPT-2-based cross-attention model that fuses user history with dialog context; they report gains in Hit@K, NDCG@K, and response-generation metrics, including a 10.9% Yelp Hit@1 improvement over the strongest baseline.

Significance. The intended contribution is timely: a scalable, behaviorally grounded source of conversational recommendation dialogs with long-term history would fill a real gap, and the pipeline contains well-designed components (unsupervised aspect/sentiment profiling, plan-constrained generation with resampling and fidelity checks, cross-domain application). If the realism and evaluation issues were resolved, the released datasets and model could be useful to the community. As it stands, however, the central claims are not established: the system agent is given the target item through the plan, so the dialogs do not demonstrate emergent recommendation inference, and the downstream evaluation is confounded by the structural guarantee that every dialog recommends and accepts the target.

major comments (3)
  1. [§3.1, Stage 2 and Stage 3] The Recommend super-node is defined with the concrete vertex (System, Recommend, v_k), and Stage 3 states that at each turn the speaking agent receives the eligible frontier act. The System LLM therefore receives the ground-truth item identity through the plan, which directly contradicts the Stage 3 claim that 'the true item v_k is not revealed' and the stated behavioral property that recommendations 'emerge from multi-turn elicitation and clarification, not from privileged access to the ground-truth item.' Because the system agent is the component that should be inferring the target from the conversation, this contradiction undermines the dataset-realism claim that the paper's title, abstract, and Section 3 present as the main contribution.
  2. [§4 and Tables 3–5] The experimental protocol does not specify whether the dialog context used for candidate ranking or recommendation generation includes the system's own recommendation of the target item. Since every generated dialog is guaranteed by the plan to recommend and accept v_k, any evaluation performed after the recommendation turn reduces to reading the target from the transcript; in that case the reported gains (e.g., the 10.9% Yelp Hit@1 improvement in Table 4) cannot support the claim that cross-attention fusion of history and dialogue is responsible. Even if the prediction point precedes the target mention, the oracle-informed system turns in the training data make the dialogue artificially informative, so the experiments as described cannot separate genuine fusion from exploitation of the plan-induced dialogue structure.
  3. [§3.3] The human and LLM-judge evaluations rate naturalness, coherence, and groundedness of the generated utterances, but they do not evaluate whether the system's recommendation behavior is plausible for an agent that must infer the target without privileged access. The high groundedness scores (4.13–4.24) are therefore not evidence against the oracle concern; they only confirm that the generated text is fluent and consistent with the user profile, which is expected given that the system was given the answer.
minor comments (4)
  1. [§4 and §5.1] The model size is given as 117M in Section 4 and 124M in Section 5.1; these numbers should be reconciled.
  2. [Table 4] The task name 'Candidate Apps Ranking' is used for Yelp and Amazon Electronics, which are not app domains; rename to 'Candidate Item Ranking' or similar.
  3. [§5.2] The construction of the candidate set for the ranking experiment is not described; without knowing the number and source of distractors, the Hit@K and NDCG@K numbers are hard to interpret.
  4. [Tables 4–5] No confidence intervals or significance tests are reported for the ranking comparisons, despite the text calling the gains 'significant'.

Circularity Check

2 steps flagged · score 8.0 of 10

Target item is baked into the semantic plan (System, Recommend, v_k) and into every dialog's final acceptance, so the System agent has privileged access and downstream Hit@K gains reduce to reading a plan-inserted answer.

  1. self definitional [Section 3.1, Stage 2 (Semantic Dialog Planning, Recommend super-node) and Stage 3 (Multi-Turn Dialog Simulation)]
    "The core of this phase is (System, Recommend, 𝑣𝑘), followed by a user (Accept/Reject, 𝑣𝑘) decision that flips to Reject with probability 𝑞=0.5."

    The plan phi_{i,k} is the conditioning structure for both agents, and at each turn the speaking agent receives the next eligible frontier vertex. Since the Recommend super-node's core vertex is literally labeled with the target v_k, the System LLM is given the ground-truth item through the plan. This contradicts the Stage 3 assertion that 'the true item v_k is not revealed' and that recommendations 'emerge from multi-turn elicitation and clarification, not from privileged access to the ground-truth item.' The claimed emergent behavior is therefore true by construction of the plan, not by simulation result.

  2. self definitional [Section 3.1 (formal dataset definition) and Section 4/Tables 4-5 (joint model evaluation)]
    "Each pair (ℎ(𝑖,𝑘) 𝑡, 𝑠(𝑖,𝑘) 𝑡) represents a user and system’s utterances where the dialog concludes by recommending the ground-truth item 𝑣𝑘 to user 𝑢𝑖."

    The dataset is defined so that every dialog terminates with a recommendation and acceptance of the target item. The downstream cross-attention model is trained and evaluated on these dialogs, with the same v_k as the ranking ground truth; Table 4 reports a 10.9% Hit@1 gain on Yelp. Because the dialog text was engineered to state and accept the target, Hit@K and NDCG@K can be inflated by exploiting the plan-inserted outcome rather than by genuinely fusing long-term history with conversational cues.

full rationale

Stage 1 (temporal profiling) is not circular: aspect induction and sentiment-based profiles are computed from review text independently of the target recommendation. The circularity enters in the simulation-evaluation loop. Stage 2 places the concrete vertex (System, Recommend, v_k) in the plan, and Stage 3 hands the speaking agent the eligible frontier vertex, so the System LLM is conditioned on the target that the paper says is hidden. The formal dataset definition then guarantees every dialog concludes by recommending and accepting v_k. When the cross-attention model is scored on Hit@K/NDCG@K using these same generated dialogs, the predicted item is the item the generator was forced to output, so the headline 10.9% Yelp Hit@1 improvement is at least partly an artifact of construction rather than evidence of genuine multi-turn elicitation or fusion. All naturalness/groundedness and ranking evaluations are internal to the generated data, with no external human-human CRS benchmark to validate transfer. This is a definitional leakage of the target, not a self-citation problem, so the appropriate score is 8.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The framework depends on a large number of hand-set hyperparameters (K, tau, gamma, q, eta, etc.) that shape the generated data distribution. It also assumes that unsupervised aspect clusters are meaningful, that LLM agents simulate real user behavior, and that LLM-judge evaluations are valid. No invented physical entities are introduced.

free parameters (9)
  • K (number of aspect clusters) = 20
    Number of aspect clusters induced from review sentences via contrastive spherical k-means; chosen because larger K creates redundant clusters and overloads the dialog planner (Section 3.1, Stage 1).
  • tau (neutral cluster distance margin) = 0.35
    Sentence-centroid cosine distance threshold above which sentences are assigned to a neutral cluster and excluded; validated to retain >95% of aspect-bearing sentences (Section 3.1, Stage 1).
  • gamma (time decay rate) = 0.015
    Exponential decay parameter in user profiling; set so an interaction one year old gets half the weight of a current one (Section 3.1, Stage 1).
  • epsilon (aspect exploration) = 1e-4
    Additive smoothing constant in aspect sampling probability to ensure all aspects can be selected (Section 3.1, Stage 2).
  • rho (Clarify probability) = 0.15
    Probability of adding a Clarify intent when community confidence mass for an aspect is below threshold (Section 3.1, Stage 2).
  • q (initial rejection probability) = 0.5
    Probability that the simulated user rejects the first recommendation, creating a fallback scenario; the third recommendation is always accepted (Section 3.1, Stage 2).
  • eta (RequestInfo probability) = 0.3
    Probability of inserting up to two user follow-up questions after a recommendation (Section 3.1, Stage 2).
  • delta (max consecutive plan violations) = 4
    If the LLM utterance cannot be parsed to a frontier act for 4 consecutive tries, the dialog is discarded (Section 3.1, Stage 3).
  • max turns budget = 30
    Dialogs that do not reach the terminal super-node within 30 turns are discarded (Section 3.1, Stage 3).
assumptions (5)
  • domain assumption Aspect clusters induced by contrastive spherical k-means on Flan-T5-XL sentence embeddings correspond to meaningful product aspects.
    The entire profile construction and dialog planning assume the unsupervised clusters are semantically coherent; if clusters are noisy, user and item profiles are mis-grounded. Invoked in Stage 1 of Section 3.1.
  • domain assumption GPT-4o role-specialized agents produce utterances that approximate how real users and systems converse about preferences.
    The dataset generation relies on LLMs simulating both roles; human ratings of fluency and coherence do not establish fidelity to real user behavior. Invoked in Stage 3 of Section 3.1.
  • domain assumption The system agent can infer and recommend the ground-truth item from history and item profiles without privileged access.
    The plan guarantees the ground-truth item is recommended and accepted, so this assumption is not actually tested; the dialog structure forces the outcome. Invoked in Stage 2/3 of Section 3.1.
  • domain assumption LLM judge scores (Claude 3.7 Sonnet) are a valid proxy for dialog quality.
    Automatic evaluation uses an LLM judge; correlation with human ratings (rho=0.70-0.78) supports but does not prove validity. Invoked in Section 3.3, Automatic Evaluation.
  • standard math Temporal splitting of interactions into train/test reflects real deployment and avoids leakage.
    Standard evaluation protocol; the generated dialogs condition only on information available up to the target interaction time, so the split is reasonable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Framework for Generating Conversational Recommendation Datasets from Behavioral Interactions." pith.science (2026). https://pith.science/paper/TWAZBVNX

@misc{pith2026250617285,
  author       = {Pith},
  title        = {Pith review of: A Framework for Generating Conversational Recommendation Datasets from Behavioral Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TWAZBVNX}},
  note         = {Machine review of arXiv:2506.17285}
}
read the original abstract

Modern recommendation systems typically follow two complementary paradigms: collaborative filtering, which models long-term user preferences from historical interactions, and conversational recommendation systems (CRS), which interact with users in natural language to uncover immediate needs. Each captures a different dimension of user intent. While CRS models lack collaborative signals, leading to generic or poorly personalized suggestions, traditional recommenders lack mechanisms to interactively elicit immediate needs. Unifying these paradigms promises richer personalization but remains challenging due to the lack of large-scale conversational datasets grounded in real user behavior. We present ConvRecStudio, a framework that uses large language models (LLMs) to simulate realistic, multi-turn dialogs grounded in timestamped user-item interactions and reviews. ConvRecStudio follows a three-stage pipeline: (1) Temporal Profiling, which constructs user profiles and community-level item sentiment trajectories over fine-grained aspects; (2) Semantic Dialog Planning, which generates a structured plan using a DAG of flexible super-nodes; and (3) Multi-Turn Simulation, which instantiates the plan using paired LLM agents for the user and system, constrained by executional and behavioral fidelity checks. We apply ConvRecStudio to three domains -- MobileRec, Yelp, and Amazon Electronics -- producing over 12K multi-turn dialogs per dataset. Human and automatic evaluations confirm the naturalness, coherence, and behavioral grounding of the generated conversations. To demonstrate utility, we build a cross-attention transformer model that jointly encodes user history and dialog context, achieving gains in Hit@K and NDCG@K over baselines using either signal alone or naive fusion. Notably, our model achieves a 10.9% improvement in Hit@1 on Yelp over the strongest baseline.

Figures

Figures reproduced from arXiv: 2506.17285 by the authors.

Figure 1
Figure 1. First, we develop an automatic conversational [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the ConvRecStudio framework. generation framework, generated datasets, and our proposed model to promote future research in conversational recommendations. Specifically, this work makes the following contributions: • We propose ConvRecStudio, a three-stage framework for synthesizing realistic, multi-turn conversational recommen￾dation dialogs grounded in timestamped user-item interac￾tions and user revie… view at source ↗
Figure 3
Figure 3. Distribution of human ratings on a sample of 3,000 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Distribution of LLM-predicted ratings for all 38K di [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 19 canonical work pages

  1. [1]

    [n. d.]. Goodreads Book Graph Datasets. https://mengtingwan.github.io/data/ goodreads.html. Accessed: 2024-11-01

  2. [2]

    Yelp Open Dataset

    2022. Yelp Open Dataset. https://www.yelp.com/dataset. Accessed: 2023-18-02

  3. [3]

    2023. Amazon. https://amazon-reviews-2023.github.io/. Accessed: 2024-11-01

  4. [4]

    Gediminas Adomavicius and Alexander Tuzhilin. 2005. Toward the next gen- eration of recommender systems: A survey of the state-of-the-art and possible extensions.IEEE transactions on knowledge and data engineering17, 6 (2005), 734–749

  5. [5]

    Diego Antognini and Boi Faltings. 2020. Hotelrec: a novel very large-scale hotel recommendation dataset.arXiv preprint arXiv:2002.06854(2020)

  6. [6]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Dem- szky, Chris Donahue, Moussa Doumbouya, Esin Durmus, St...

  7. [7]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

  8. [8]

    Chong Chen, Min Zhang, Yiqun Liu, and Shaoping Ma. 2018. Neural Attentional Rating Regression with Review-level Explanations. InProceedings of the 2018 World Wide Web Conference(Lyon, France)(WWW ’18). International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, CHE, 1583–1592. doi:10.1145/3178876.3186070

Show all 58 references
  1. [9]

    Bosheng Ding, Chengwei Qin, Ruochen Zhao, Tianze Luo, Xinze Li, Guizhen Chen, Wenhan Xia, Junjie Hu, Luu Anh Tuan, and Shafiq Joty. 2024. Data augmentation using llms: Data perspectives, learning paradigms and challenges. InFindings of the Association for Computational Linguis...

  2. [10]

    Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexan- der Miller, Arthur Szlam, and Jason Weston. 2015. Evaluating prerequisite qual- ities for learning end-to-end dialog systems.arXiv preprint arXiv:1511.06931 (2015)

  3. [11]

    Zuohui Fu, Yikun Xian, Yaxin Zhu, Shuyuan Xu, Zelong Li, Gerard De Melo, and Yongfeng Zhang. 2021. Hoops: Human-in-the-loop graph reasoning for conversational recommendation. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Informatio...

  4. [12]

    Chongming Gao, Wenqiang Lei, Xiangnan He, Maarten de Rijke, and Tat-Seng Chua. 2021. Advances and challenges in conversational recommender systems: A survey.AI Open2 (2021), 100–126. doi:10.1016/j.aiopen.2021.06.002

  5. [13]

    Shirley Anugrah Hayati, Dongyeop Kang, Qingxiaoyang Zhu, Weiyan Shi, and Zhou Yu. 2020. Inspired: Toward sociable recommendation dialog systems.arXiv preprint arXiv:2009.14306(2020)

  6. [14]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th international conference on world wide web. 173–182

  7. [15]

    Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian McAuley. 2023. Large language models as zero-shot conversational recommenders. InProceedings of the 32nd ACM international conference on informatio...

  8. [16]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  9. [17]

    Myeongjun Jang, Antonios Georgiadis, Yiyun Zhao, and Fran Silavong. 2024. DriftWatch: A Tool that Automatically Detects Data Drift and Extracts Repre- sentative Examples Affected by Drift. InProceedings of the 2024 Conference of the North American Chapter of the Association fo...

  10. [18]

    Dietmar Jannach, Ahtsham Manzoor, Wanling Cai, and Li Chen. 2021. A survey on conversational recommender systems.ACM Computing Surveys (CSUR)54, 5 (2021), 1–36

  11. [19]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques.ACM Transactions on Information Systems (TOIS)20, 4 (2002), 422–446

  12. [20]

    Meihuizi Jia, Ruixue Liu, Peiying Wang, Yang Song, Zexi Xi, Haobin Li, Xin Shen, Meng Chen, Jinhui Pang, and Xiaodong He. 2022. E-ConvRec: a large- scale conversational recommendation dataset for E-commerce customer service. InProceedings of the Thirteenth Language Resources a...

  13. [21]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

  14. [22]

    Tom Kocmi, Eleftherios Avramidis, Rachel Bawden, Ondřej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, et al. 2024. Findings of the WMT24 general machine trans- lation shared task: the LLM era is here but mt is not ...

  15. [23]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems.Computer42, 8 (2009), 30–37

  16. [24]

    Levenshtein

    Vladimir I. Levenshtein. 1965. Binary codes capable of correcting deletions, insertions, and reversals.Soviet physics. Doklady10 (1965), 707–710. https: //api.semanticscholar.org/CorpusID:60827152

  17. [25]

    Chuang Li, Hengchang Hu, Yan Zhang, Min-Yen Kan, and Haizhou Li. 2023. A Conversation is Worth A Thousand Recommendations: A Survey of Holistic Conversational Recommender Systems. arXiv:2309.07682 [cs.CL] https://arxiv. org/abs/2309.07682

  18. [26]

    Dong Li, Ruoming Jin, Jing Gao, and Zhi Liu. 2020. On sampling top-k rec- ommendation evaluation. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2114–2124

  19. [27]

    Raymond Li, Samira Ebrahimi Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, and Chris Pal. 2018. Towards deep conversational recommendations. Advances in neural information processing systems31 (2018)

  20. [28]

    Lizi Liao, Le Hong Long, Zheng Zhang, Minlie Huang, and Tat-Seng Chua. 2021. MMConv: an environment for multimodal conversational search across multiple domains. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 675–684

  21. [29]

    Allen Lin, Ziwei Zhu, Jianling Wang, and James Caverlee. 2023. Enhancing User Personalization in Conversational Recommenders. InProceedings of the ACM Web Conference 2023(Austin, TX, USA)(WWW ’23). Association for Computing Machinery, New York, NY, USA, 770–778. doi:10.1145/35...

  22. [30]

    Yuanxing Liu, Weinan Zhang, Baohua Dong, Yan Fan, Hang Wang, Fan Feng, Yifan Chen, Ziyu Zhuang, Hengbin Cui, Yongbin Li, et al. 2023. U-NEED: A Fine-grained Dataset for User Needs-Centric E-commerce Conversational Recommendation. arXiv preprint arXiv:2305.04774(2023)

  23. [31]

    Zeming Liu, Haifeng Wang, Zheng-Yu Niu, Hua Wu, Wanxiang Che, and Ting Liu. 2020. Towards conversational recommendation over multi-type dialogs. arXiv preprint arXiv:2005.03954(2020). A Framework for Generating Conversational Recommendation Datasets from Behavioral Interaction...

  24. [32]

    M. H. Maqbool, Umar Farooq, Adib Mosharrof, A. B. Siddique, and Hassan Foroosh. 2023. MobileRec: A Large Scale Dataset for Mobile Apps Recommenda- tion. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval(<conf-loc>...

  25. [33]

    Joao Monteiro, Pierre-Andre Noel, Etienne Marcotte, Sai Rajeswar Mudumba, Valentina Zantedeschi, David Vazquez, Nicolas Chapados, Chris Pal, and Perouz Taslakian. 2024. RepLiQA: A question-answering dataset for benchmarking LLMs on unseen reference content.Advances in Neural I...

  26. [34]

    Seungwhan Moon, Pararth Shah, Anuj Kumar, and Rajen Subba. 2019. Open- dialkg: Explainable conversational reasoning with attention-based walks over knowledge graphs. InProceedings of the 57th annual meeting of the association for computational linguistics. 845–854

  27. [35]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

  28. [36]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a Method for Automatic Evaluation of Machine Translation. InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics, Pierre Isabelle, Eugene Charniak, and Dekang Lin (Eds.)...

  29. [37]

    Xiao Pu, Mingqi Gao, and Xiaojun Wan. 2023. Summarization is (almost) dead. arXiv preprint arXiv:2309.09558(2023)

  30. [38]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. https: //api.semanticscholar.org/CorpusID:160025533

  31. [39]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the lim- its of transfer learning with a unified text-to-text transformer.arXiv preprint arXiv:1910.10683(2019)

  32. [40]

    Vipula Rawte, Amit Sheth, and Amitava Das. 2023. A survey of hallucination in large foundation models.arXiv preprint arXiv:2309.05922(2023)

  33. [41]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International conference on data mining. IEEE, 995–1000

  34. [42]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  35. [43]

    Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining(Marina Del Rey, CA, USA)(WSDM ’18). Association for Computing Machinery, Ne...

  36. [44]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucu- rull, David Esiobu, Jude Fernandes, Jeremy...

  37. [45]

    Pengkun Wang, Zhe Zhao, HaiBin Wen, Fanfu Wang, Binwu Wang, Qingfu Zhang, and Yang Wang. 2024. Llm-autoda: Large language model-driven automatic data augmentation for long-tailed problems.Advances in Neural Information Processing Systems37 (2024), 64915–64941

  38. [46]

    Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models.arXiv preprint arXiv:2309.11674(2023)

  39. [47]

    Hu Xu, Seungwhan Moon, Honglei Liu, Bing Liu, Pararth Shah, and Philip S Yu. 2020. User memory reasoning for conversational recommendation.arXiv preprint arXiv:2006.00184(2020)

  40. [48]

    Haopeng Zhang, Philip S Yu, and Jiawei Zhang. 2024. A systematic survey of text summarization: From statistical methods to large language models.Comput. Surveys(2024)

  41. [49]

    Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep learning based recom- mender system: A survey and new perspectives.ACM computing surveys (CSUR) 52, 1 (2019), 1–38

  42. [50]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi

  43. [51]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Com...

  44. [52]

    Lei Zheng, Vahid Noroozi, and Philip S Yu. 2017. Joint deep modeling of users and items using reviews for recommendation. InProceedings of the tenth ACM international conference on web search and data mining. 425–434

  45. [53]

    arXiv:1904.09675 http://arxiv.org/abs/1904.09675

    BERTScore: Evaluating Text Generation with BERT.CoRRabs/1904.09675 (2019). arXiv:1904.09675 http://arxiv.org/abs/1904.09675

  46. [54]

    Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. 2023. Toolqa: A dataset for llm question answering with external tools.Advances in Neural Information Processing Systems36 (2023), 50117–50143

  47. [56]

    Kun Zhou, Yuanhang Zhou, Wayne Xin Zhao, Xiaoke Wang, and Ji-Rong Wen

  48. [2015]

    Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939(2015)

  49. [2019]

    InProceedings of the 28th ACM international conference on information and knowledge management

    BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowledge management. 1441–1450

  50. [2020]

    Towards topic-guided conversational recommender system.arXiv preprint arXiv:2010.04125(2020)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.