Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Conversational recommenders work better when they switch from attribute questions early to item choices later.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 21:45 UTC pith:5QFIQZIO

load-bearing objection Stage-aware elicitation (attributes early, items later) is a real, usable design rule for CRS, backed by a fresh user study, the InPE annotations, and a competent MoE baseline that wins offline; annotation agreement and single-domain offline eval are the soft spots, not the core claim. the 3 major comments →

arxiv 2607.06765 v1 pith:5QFIQZIO submitted 2026-07-07 cs.IR

When and How to Ask: Dynamic Preference Elicitation Strategies for Conversational Recommendation

Feng Xia , Shuo Zhang , Xi Wang This is my paper
classification cs.IR
keywords conversational recommender systemspreference elicitationstage-aware strategiesmixture of expertsproactive dialogueInPE datasetattribute-based vs item-based
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Conversational recommender systems ask users about their tastes, but most treat the way they ask as fixed for the whole dialogue. This paper shows that the right way to ask depends on how far the conversation has gone: broad attribute questions work early, when preferences are still vague, while concrete item comparisons work better once tastes have sharpened. The authors build InPE by adding turn-level labels for whether elicitation is needed and which strategy people prefer, then train COPE, a mixture-of-experts model that routes each turn to a specialized elicitation, recommendation, or chat expert. Offline tests show gains in both ranking the right item and producing responses people prefer. The same routing analysis recovers a consistent stage-wise pattern that matches the human study, suggesting common interaction rhythms in these systems.

Core claim

Optimal preference elicitation is stage-dependent and context-sensitive: attribute-based questions dominate early dialogue stages when preferences are abstract, while item-based strategies become preferred as preferences grow concrete; hybrid strategies stay useful throughout. Explicitly modeling this timing and choice with a routed mixture of experts improves both recommendation recall and pairwise response preference over strong baselines.

What carries the argument

COPE (Conversational Preference Elicitation via Mixture of Experts): a hierarchical router that first chooses among elicitation, recommendation, or general chat, then (if eliciting) chooses attribute, item, or hybrid strategy, each executed by a lightweight expert adapter on a shared frozen language-model backbone.

Load-bearing premise

That majority-vote crowd labels of preferred strategy and response quality on movie dialogues are a reliable enough signal of real multi-turn user utility.

What would settle it

A live multi-turn user study in which systems forced to follow the stage-dependent strategy schedule (attribute early, item later) fail to outperform fixed-strategy or random-routing controls on actual recommendation success or user satisfaction.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper argues that preference elicitation in conversational recommender systems (CRSs) is stage-dependent: attribute-based strategies work best early when preferences are abstract, while item-based (and hybrid) strategies become preferable as preferences refine. Supporting evidence comes from a preliminary crowdsourced user study (Figure 1) and from InPE, a re-annotation of the INSPIRED movie dialogues that labels elicitation necessity, strategy type (attribute/item/hybrid), and response preference. The authors then introduce COPE, a hierarchical Mixture-of-Experts architecture on a frozen LLM backbone with LoRA experts and a two-stage router (task then strategy). Offline experiments on InPE report gains in Recall@k, task/strategy accuracy, and pairwise response win-rate over neighborhood, KG, and LLM baselines, plus mechanistic analysis of router activations that recovers the same stage-wise pattern.

Significance. If the stage-dependent pattern and the value of explicit strategy modeling hold, the work supplies both a useful empirical regularity for CRS design and a concrete modeling recipe (structured action space + hard-routed MoE). The public InPE annotations and the released code are genuine community resources that enable strategy-aware training and evaluation, which prior CRS datasets largely lack. The offline gains of COPE over strong LLM baselines (Tables 6–7) and the router transition analysis (Figure 5) are concrete, reproducible contributions even if live-user transfer remains open.

major comments (3)
  1. Table 4 reports Krippendorff’s α = 0.244 for the binary elicitation-necessity decision (Q1) and only moderate α = 0.531 for strategy type (Q2). Because these majority-vote labels supervise both the hierarchical router (Eq. 7, Algorithm 1) and the pairwise preference judgments used for win-rate (Tables 3 and 7), the central claim that attribute-based strategies are optimal early and item-based later rests on a noisy ground-truth signal. Majority voting and expert adjudication mitigate but do not remove the risk that reported stage-wise patterns and COPE gains largely reflect annotator consensus rather than multi-turn user utility. A reliability analysis (e.g., sensitivity to low-agreement turns, or a second independent annotation round) is needed before the labels can be treated as reliable supervision.
  2. All quantitative claims (Tables 6–9, Figure 5) are offline and confined to the single-domain INSPIRED movie corpus. The paper never demonstrates that a model trained on InPE labels improves live recommendation success, conversation length, or user satisfaction in an interactive setting, nor that the stage-dependent pattern transfers outside movies. Given that the weakest modeling assumption is precisely the transfer of these offline labels to real multi-turn utility, at least a small-scale online or simulated interactive evaluation (or an explicit discussion of this limitation with a concrete transfer experiment) is required to support the claim that “context-aware preference elicitation strategies are beneficial for conversational recommendation.”
  3. Table 8 shows that the “w/o all” (single-adapter) variant obtains higher Task Accuracy, Strategy Accuracy, and Recall@10 than the full multi-expert COPE, while under-performing only on pairwise win-rate. This raises a load-bearing question about whether the hierarchical MoE specialization itself is necessary, or whether the performance edge of the full model is largely attributable to the strategy labels rather than to expert separation. A clearer analysis of when specialization helps versus when a unified adapter suffices would strengthen the architectural claim.
minor comments (5)
  1. Minor numerical inconsistencies appear between text and tables (e.g., 60.02 % vs. 60.07 % for elicitation-yes turns; hybrid proportions 46.97 % vs. 47.03 %). Please reconcile.
  2. Figure 3 and Figure 5 would benefit from absolute counts or confidence intervals alongside the normalized proportions, so readers can judge sample size per turn.
  3. The GitHub URL uses the handle “juanfacabian” while the author list is Xia/Zhang/Wang; a brief note on repository ownership would avoid confusion.
  4. Hyper-parameter choices (LoRA rank, loss weights in Table 5, InfoNCE temperature) are reported but not ablated; a short sensitivity paragraph would help reproducibility.
  5. In §5.1 the action triplet ⟨a_t, σ_t, ϕ_t⟩ is introduced, yet ϕ_t (recommendation objective) receives little subsequent modeling detail compared with a_t and σ_t.

Circularity Check

2 steps flagged

Mild circularity: router predictions recover annotation stage patterns by construction, and SFT on preferred responses makes pairwise likelihood win-rate largely forced.

specific steps
  1. fitted input called prediction [Abstract; §6.4 Mechanistic Analysis / Figure 5]
    "the analysis of the predicted strategies uncovers consistent stage-wise tendencies in dialogue progression, providing empirical evidence of common interaction patterns in conversational recommendation systems."

    Ground-truth strategy labels already exhibit the stage-dependent distribution (Figure 3). The router is trained by supervised cross-entropy on exactly those labels (Eq. 7; teacher-forced in Algorithm 1). Accurate prediction therefore recovers the identical stage-wise pattern by construction of the fit; presenting the recovered distribution as independent empirical evidence is circular.

  2. fitted input called prediction [§5.2.2 Multi-Task Learning (Lsft Eq. 8); §6.1 (iii) Response Quality / Table 7 / Eq. 10]
    "We evaluate response quality using human-annotated response pairs (𝑦𝑤, 𝑦𝑙 ), where𝑦𝑤 is the preferred response and 𝑦𝑙 is the less preferred one. We calculate (1)Pair-wise Win Rate by checking whether the model assigns a higher log-likelihood to the preferred response 𝑦𝑤 than to 𝑦𝑙 ; and (2)the Log-likelihood Margin"

    Task experts are updated via supervised fine-tuning Lsft (Eq. 8) whose targets are the annotator-selected preferred (strategy-aware) responses that define the positive side of the contrastive pairs. Measuring whether the resulting model assigns higher normalized log-likelihood to those same yw than to yl is statistically forced by the SFT objective rather than an independent test of preference.

full rationale

The paper's central stage-dependence claim has independent non-circular support from a separate preliminary user study (328 participants, Figure 1) showing attribute-based preference early and item-based later; this is not derived from the later model. InPE annotations (Figure 3, Table 1) and COPE training/eval are self-contained offline ML experiments against baselines, with no self-definitional equations, no uniqueness theorems imported via self-citation, and no ansatz smuggling. However, two analysis/eval steps partially reduce to fitted inputs: (1) the hierarchical router is teacher-forced on the exact InPE strategy labels (Eq. 7, Algorithm 1) that already encode the stage-wise distribution, so accurate predictions necessarily reproduce it—yet Abstract and §6.4 present the recovered pattern in Figure 5 as new 'empirical evidence of common interaction patterns'; (2) experts receive Lsft (Eq. 8) on the annotator-preferred (strategy-aware) responses that define the positive pairs, after which pairwise win-rate/margin (Table 7, Eq. 10) simply checks higher likelihood on those same yw vs yl. These do not collapse the modeling contribution or the Recall gains, but they mean the 'uncovering' and response-quality claims are not fully independent of the annotation inputs used for supervision. Score 3 reflects partial/mild circularity only on those secondary claims; primary empirical pattern and architecture remain non-circular.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on crowd-sourced preference labels and a standard LLM+LoRA training regime; the free parameters are ordinary ML hyperparameters rather than physical constants. No new physical entities are postulated. Domain assumptions about dialogue stages and the sufficiency of offline pairwise judgments are the main non-standard premises.

free parameters (4)
  • LoRA rank r and alpha
    Set to r=8, α=16; controls capacity of each expert adapter and therefore the specialization claimed by the MoE design.
  • Multi-task loss weights (w_sft, w_task, w_eli, w_rec)
    Hand-tuned values (0.5 / 1.0 / 1.0 / 1.0) that determine the relative importance of routing versus generation versus ranking objectives.
  • InfoNCE temperature τ
    Temperature in the contrastive recommendation loss; affects ranking sharpness but is not given an explicit numeric value in the text.
  • Learning rate and warmup ratio
    Peak LR 2e-5, warmup 0.03; standard but still free choices that affect final performance numbers.
axioms (4)
  • domain assumption User preferences evolve from abstract to concrete across dialogue stages, making attribute-based elicitation preferable early and item-based later.
    Stated as empirical finding from the user study (Fig. 1) and used as the organizing principle for both annotation and model design.
  • domain assumption Majority-vote crowd annotations of elicitation necessity and preferred strategy constitute a valid supervisory signal for real multi-turn utility.
    Core training and evaluation labels (Tables 1–4); low α on Q1 is acknowledged but majority vote is still treated as ground truth.
  • ad hoc to paper A frozen LLM backbone plus lightweight LoRA experts can specialize into distinct CRS behaviors without mutual interference under hard routing.
    Architectural premise of COPE (Eq. 3 and hierarchical router); not independently validated outside this work.
  • standard math Standard cross-entropy and InfoNCE losses are appropriate objectives for routing and item ranking respectively.
    Ordinary supervised and contrastive losses used throughout §5.2.
invented entities (2)
  • InPE dataset independent evidence
    purpose: Provide turn-level labels for elicitation necessity, strategy type, and preferred response on top of INSPIRED dialogues.
    New resource constructed by the authors; independent evidence is limited to the public GitHub release and the reported annotation statistics.
  • COPE hierarchical MoE architecture no independent evidence
    purpose: Explicitly model task-level and strategy-level decisions via separate experts and routers.
    Novel combination of known components (MoE, LoRA, CRS action space) introduced for this paper; no external validation yet.

pith-pipeline@v1.1.0-grok45 · 20190 in / 3040 out tokens · 31032 ms · 2026-07-10T21:45:13.080112+00:00 · methodology

0 comments
read the original abstract

Conversational Recommender Systems (CRSs) are interactive systems that use multi-turn natural language dialogue to understand evolving user preferences and provide personalized recommendations. To achieve this goal, CRSs rely on preference elicitation strategies to actively gather informative preference cues from users; however, the timing and selection of these strategies during a conversation remain largely unexplored. While many existing studies emphasize eliciting explicit item attributes and tend to adopt relatively static elicitation strategies, the use of item-based preference elicitation and how it varies across different dialogue stages remains less explored. In this work, we conduct a systematic investigation of preference elicitation strategies from a stage-aware perspective. We provide empirical evidence that optimal preference elicitation strategies are stage-dependent and context-sensitive: attribute-based inquiries are effective in early stages, while item-based strategies become superior as preferences refine. To support this paradigm, we introduce InPE, a dataset enriched with fine-grained annotations for elicitation necessity and strategy selection. With this dataset, we propose COPE (COnversational Preference Elicitation via Mixture of Experts), a novel architecture for strategy modeling. Extensive offline evaluation on our dataset indicates that context-aware preference elicitation strategies are beneficial for conversational recommendation. In addition, the analysis of the predicted strategies uncovers consistent stage-wise tendencies in dialogue progression, providing empirical evidence of common interaction patterns in conversational recommendation systems. Our dataset is available at https://github.com/juanfacabian/InPE.

Figures

Figures reproduced from arXiv: 2607.06765 by Feng Xia, Shuo Zhang, Xi Wang.

Figure 1
Figure 1. Figure 1: User selection ratios of preference elicitation strate [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the step-by-step annotation workflow. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Proportion of different elicitation strategies in an [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of our proposed conversational recom [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Quantitative analysis of the dynamic routing mechanism. (a) Task-level expert activation probabilities across dialogue [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 7 internal anchors

  1. [1]

    Laura Blédaité and Francesco Ricci. 2015. Pairwise preferences elicitation and exploitation for conversational collaborative filtering. InProceedings of the 26th ACM Conference on Hypertext and Social Media. 231–236

    work page 2015

  2. [2]

    Qibin Chen, Junyang Lin, Yichang Zhang, Ming Ding, Yukuo Cen, Hongxia Yang, and Jie Tang. 2019. Towards knowledge-based recommender dialog system. arXiv preprint arXiv:1908.05391(2019)

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  3. [3]

    Konstantina Christakopoulou, Filip Radlinski, and Katja Hofmann. 2016. Towards conversational recommender systems. InProceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 815–824

    work page 2016

  4. [4]

    Yang Deng, Yaliang Li, Fei Sun, Bolin Ding, and Wai Lam. 2021. Unified conversa- tional recommendation policy learning via graph-based reinforcement learning. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1431–1441

    work page 2021

  5. [5]

    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

    work page 2021

  6. [6]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt and predict paradigm (p5). InProceedings of the 16th ACM Conference on Recommender Systems. 299–315

    work page 2022

  7. [7]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Shuyu Guo, Shuo Zhang, Weiwei Sun, Pengjie Ren, Zhumin Chen, and Zhaochun Ren. 2023. Towards explainable conversational recommender systems. InProceed- ings of the 46th International ACM SIGIR conference on research and development in information retrieval. 2786–2795

    work page 2023

  9. [9]

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  10. [10]

    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 Information and Knowledge Management. 720–730

    work page 2023

  11. [11]

    Diana C Hernandez-Bocanegra and Jürgen Ziegler. 2023. Explaining recommen- dations through conversations: Dialog model and the effects of interface type and degree of interactivity.ACM Transactions on Interactive Intelligent Systems 13, 2 (2023), 1–47

    work page 2023

  12. [12]

    Chenhao Hu, Shuhua Huang, Yansen Zhang, and Yubao Liu. 2022. Learning to infer user implicit preference in conversational recommendation. InProceedings of the 45th International ACM SIGIR conference on research and development in information retrieval. 256–266

    work page 2022

  13. [13]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.ICLR1, 2 (2022), 3

    work page 2022

  14. [14]

    Ivica Kostric, Krisztian Balog, and Filip Radlinski. 2024. Generating usage-related questions for preference elicitation in conversational recommender systems. ACM Transactions on Recommender Systems2, 2 (2024), 1–24

    work page 2024

  15. [15]

    2018.Content analysis: An introduction to its methodology

    Klaus Krippendorff. 2018.Content analysis: An introduction to its methodology. Sage publications

    work page 2018

  16. [16]

    Wenqiang Lei, Xiangnan He, Yisong Miao, Qingyun Wu, Richang Hong, Min- Yen Kan, and Tat-Seng Chua. 2020. Estimation-action-reflection: Towards deep interaction between conversational and recommender systems. InProceedings of the 13th ACM International Conference on Web Search and Data Mining. 304–312

    work page 2020

  17. [17]

    Wendi Li, Wei Wei, Xiaoye Qu, Xian-Ling Mao, Ye Yuan, Wenfeng Xie, and Dangyang Chen. 2023. Trea: Tree-structure reasoning schema for conversational recommendation.arXiv preprint arXiv:2307.10543(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    Kevin McCarthy, James Reilly, Lorraine McGinty, and Barry Smyth. 2004. On the dynamic generation of compound critiques in conversational recommender systems. InInternational Conference on Adaptive Hypermedia and Adaptive Web- Based Systems. Springer, 176–184

    work page 2004

  19. [19]

    Kevin McCarthy, Yasser Salem, and Barry Smyth. 2010. Experience-based cri- tiquing: Reusing critiquing experiences to improve conversational recommenda- tion. InInternational Conference on Case-Based Reasoning. Springer, 480–494

    work page 2010

  20. [20]

    Bilih Priyogi. 2019. Preference elicitation strategy for conversational recom- mender system. InProceedings of the twelfth ACM International Conference on Web Search and Data Mining. 824–825

    work page 2019

  21. [21]

    Xuhui Ren, Hongzhi Yin, Tong Chen, Hao Wang, Zi Huang, and Kai Zheng

    work page

  22. [22]

    InProceedings of the 44th international ACM SIGIR Conference on Research and Development in Information Retrieval

    Learning to ask appropriate questions in conversational recommendation. InProceedings of the 44th international ACM SIGIR Conference on Research and Development in Information Retrieval. 808–817

    work page

  23. [23]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt- Thieme. 2012. BPR: Bayesian personalized ranking from implicit feedback.arXiv preprint arXiv:1205.2618(2012)

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  24. [24]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  25. [25]

    Yueming Sun and Yi Zhang. 2018. Conversational recommender system. InThe 41st international acm sigir conference on research and development in information retrieval. 235–244

    work page 2018

  26. [27]

    Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat-Seng Chua. 2019. Kgat: Knowledge graph attention network for recommendation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 950–958

    work page 2019

  27. [28]

    Xiaolei Wang, Kun Zhou, Ji-Rong Wen, and Wayne Xin Zhao. 2022. Towards unified conversational recommender systems via knowledge-enhanced prompt learning. InProceedings of the 28th ACM SIGKDD conference on knowledge dis- covery and data mining. 1929–1937

    work page 2022

  28. [29]

    Zhouhang Xie, Junda Wu, Hyunsik Jeon, Zhankui He, Harald Steck, Rahul Jha, Dawen Liang, Nathan Kallus, and Julian McAuley. 2024. Neighborhood-based collaborative filtering for conversational recommendation. InProceedings of the 18th ACM Conference on Recommender Systems. 1045–1050

    work page 2024

  29. [30]

    Zhihui Xie, Tong Yu, Canzhe Zhao, and Shuai Li. 2021. Comparison-based conversational recommender system with relative bandit feedback. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1400–1409

    work page 2021

  30. [31]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [32]

    Ting Yang and Li Chen. 2024. Unleashing the retrieval potential of large language models in conversational recommender systems. InProceedings of the 18th ACM Conference on Recommender Systems. 43–52

    work page 2024

  32. [33]

    Tong Yu, Yilin Shen, and Hongxia Jin. 2019. A visual dialog augmented interactive recommender system. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 157–165

    work page 2019

  33. [34]

    Yiming Zhang, Lingfei Wu, Qi Shen, Yitong Pang, Zhihua Wei, Fangli Xu, Bo Long, and Jian Pei. 2022. Multiple choice questions based multi-interest policy learning for conversational recommendation. InProceedings of the ACM Web Conference 2022. 2153–2162

    work page 2022

  34. [35]

    Kun Zhou, Wayne Xin Zhao, Shuqing Bian, Yuanhang Zhou, Ji-Rong Wen, and Jingsong Yu. 2020. Improving conversational recommender systems via knowl- edge graph based semantic fusion. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery and data mining. 1006–1014

    work page 2020

  35. [36]

    Jie Zou, Yifan Chen, and Evangelos Kanoulas. 2020. Towards question-based rec- ommender systems. InProceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 881–890

    work page 2020