Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization

Manato Tajiri; Michimasa Inaba

REVIEW 3 major objections 5 minor 43 references

Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A DPO-fine-tuned dialogue summary and item-explanation generator improves conversational recommendation ranking on two Japanese datasets, with summary refinement carrying most of the gain.

desk verdict A competent but modest DPO-on-SumRec paper whose headline gains are undermined by using the same predictor for reward and evaluation; worth refereeing, not worth believing as-is. read the letter →

arxiv 2508.19918 v3 pith:E7OH5DLI submitted 2025-08-27 cs.IR

Manato Tajiri , Michimasa Inaba This is my paper

classification cs.IR

keywords conversationalrecommendersystemsdirectpreferenceoptimizationdialoguesummarizationlargelanguagemodelsitemrecommendationinformationhitrateandmeanreciprocalrankTabidachicorpusChatRecdataset

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

Conversational recommender systems should learn what a user wants before suggesting items, but many benchmark systems recommend quickly in short exchanges. The paper proposes a two-stage pipeline in which an LLM turns the dialogue history into a summary and turns each candidate item's description into a recommendation explanation, and a frozen score predictor scores the item. Direct Preference Optimization (DPO) then fine-tunes both text generators, using the predictor's error as the preference signal: the winner and loser texts are generated candidates whose predicted scores are closest to and furthest from the ground-truth score. The paper reports higher Hit Rate and MRR at most cutoffs against no-DPO baselines on the Tabidachi and ChatRec datasets, with the dialogue-summary generator's DPO training being the main source of improvement. This matters because it offers a label-free way to make LLM-generated intermediate text carry the information a downstream recommender actually needs.

What carries the argument

The load-bearing mechanism is DPO with a frozen score predictor as the reward model. In Stage 1, DeBERTa predicts the ground-truth recommendation score from a dialogue summary, an item recommendation explanation, and the item description. In Stage 2, for each input the LLM samples several summaries or explanations; the sample whose predicted score is closest to the ground truth is the winner, the furthest is the loser, and the DPO loss widens the log-probability gap between them. Because the losers are well-formed 'almost good' texts, DPO learns fine-grained distinctions between phrasing that helps the downstream scorer and phrasing that does not.

What would settle it

Retrain the Stage-1 score predictor on outputs of the DPO-tuned generators, or replace it with an independently trained encoder, and re-run the Tabidachi evaluation; if the Hit Rate and MRR gains vanish or reverse, the DPO improvements were tuned to the original predictor's idiosyncrasies rather than to semantically better summaries and explanations.

Watch

Extended reading notes

Core claim

The paper's central claim is that DPO fine-tuning of both LLM generators improves item recommendation on realistic conversational datasets, with the dialogue-summary generator carrying most of the benefit. The ablation shows that removing DPO from the summary model costs more than removing it from the item-information model; human evaluation finds DPO summaries more consistent, fluent, and useful, while DPO item explanations are rated worse on all criteria—a trade-off the paper accepts because those explanations are internal features, not user-facing text. The connecting mechanism is that preference pairs are not hand-written: they are generated candidates ranked by a frozen DeBERTa score pr

Load-bearing premise

The load-bearing premise is that the frozen scoring model trained on texts from the untuned language model gives an honest score for texts from the tuned model; if DPO only exploits the scoring model's blind spots, the ranking gains are not gains in recommendation quality.

Editorial extensions

If this is right

Refining the dialogue summary is the highest-leverage intervention in this pipeline: the ablation's 'w/o Rec-DPO' row shows most of the gain survives when only the summary is DPO-trained.
The preference construction needs no human annotation: winner/loser pairs are sampled generations scored by the frozen predictor, so the same recipe can be reused whenever an LLM produces intermediate text for a downstream scorer.
On short, high-precision dialogues like ChatRec, DPO's benefit shows up as better top-rank ordering (MRR) rather than larger recall sets, since Hit Rate at high cutoffs stays flat.
The item recommendation explanation becomes longer, less lexically diverse, and lower-rated by humans after DPO even though it helps the predictor, so treating it as user-facing text would require a separate quality control step.

Reading between the lines

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

Because the same frozen predictor supplies both the DPO reward and the test-time judge, part of the reported gain could be predictor-specific; retraining the predictor on DPO-generated texts or swapping in a different encoder would test whether the gain is semantic or an artifact of exploitation.
The two datasets are both Japanese travel-related dialogues, so applying the method to other languages and domains may require recalibrating the chunked-summary and score-predictor stages, and the gains may shrink.
A natural next step the paper does not run is iterative DPO: use the tuned generators to rebuild the preference data and retrain the predictor, which would show whether the improvements compound or hit diminishing returns.
If item recommendation information is ever shown to users, the observed human-eval decline becomes a real cost; a user-facing variant would need a factual-consistency constraint, since the case study already shows an item recommendation containing a detail not present in the description.

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.

Desk Editor's Note

read the letter

Quick take: this is a straightforward, clearly-written application of DPO to SumRec. The contribution is incremental—training the two LLM text generators with preference pairs derived from score differences—and it mostly works about as well as you'd expect. However, the central evaluation loop is too tight to support the headline claim.

What's genuinely good: the method is simple and described well. They build on SumRec, keep the score predictor frozen while generating DPO data, include ablations showing summary-DPO drives most of the gain, run human eval on generated text, and release code. The Tabidachi summary gains in human usefulness are plausible. The recommendation-info gains are smaller, and the human eval shows quality drops.

The big problem is circularity. The same frozen DeBERTa scores are used to create preference pairs (Eqs. 3–5) and to rank items at test time. DPO is trained to make texts score well according to that judge, so improved HR/MRR partly measures how well the generators exploit that judge, not recommendation quality. This isn't just a theoretical worry: their own human eval (Sec. 4.7) shows DPO item-recommendation info rated worse than SumRec on all four criteria, and the case study shows hallucinated details. They hand-wave by saying item info is internal, but the item info is exactly what feeds the judge.

There are smaller issues. The conclusion claims superior performance across all cutoffs, but Table 1 shows HR@5 lower than SumRec on both datasets. No variance or significance tests are reported, even though each model was trained five times. The ChatRec gains are tiny, though the MRR improvement is consistent. The preference construction using absolute distance to ground truth is reasonable, but there's no analysis of how often the 'winner' is actually closer semantically. The Limitations section does acknowledge hallucination and narrow evaluation, which is to their credit, but it does not acknowledge the circularity.

Who's it for: people building on SumRec-style CRS pipelines, or working on RL/DPO for recommendation text. It deserves a serious referee because the method is coherent and the weakness can be fixed. In current form I'd expect major revision or rejection, but the authors could salvage it with an independent evaluation judge, human-item-judgment validation, or at minimum error bars.

Referee Report

3 major / 5 minor

Summary. The paper extends the SumRec conversational recommender by applying Direct Preference Optimization (DPO) to its two LLM-based generators: the dialogue summary generator and the item recommendation information generator. Preference pairs are constructed by generating multiple candidate texts, scoring them with a frozen DeBERTa regression model, and pairing the candidate whose score is closest to the ground-truth item score against the one farthest away (Eqs. 3-5). At test time, the same frozen DeBERTa is used to score the generated summary, item recommendation information, and item description for ranking. Experiments on Tabidachi and ChatRec report HR/MRR, ablations, automatic text analyses, and a CrowdWorks human evaluation. The authors conclude that DPO fine-tuning improves recommendation performance, mainly through the dialogue summary branch, and that degraded item recommendation information is acceptable because it is internal.

Significance. If the empirical claims were sound, the paper would make a useful contribution: it addresses realistic long-dialogue recommendation, demonstrates a DPO-based method to align generated text with recommendation utility, and ships public code with detailed hyperparameter and licensing information. The human evaluation of generated texts is a valuable addition. However, the current support is undermined by a circular evaluation loop: the same frozen DeBERTa predictor supplies both the DPO reward signal and the test-time ranking. Improvements on that predictor may reflect reward hacking rather than better recommendation quality, and the paper's own human evaluation shows that the DPO item recommendation information is judged worse than SumRec on all four criteria. The mixed results in Table 1 also contradict the paper's strong claims. These issues need to be resolved before the central claim can be accepted.

major comments (3)

[§3.3.1-§3.3.3, Eq. (5), §4.4] The evaluation is circular in a way that is load-bearing for the headline claim. Preference pairs are selected using the absolute difference between the frozen DeBERTa predictor's output and the ground-truth score (Eqs. 3-4), and the final HR/MRR evaluation uses the same frozen predictor (Eq. 1). Thus the generators are explicitly trained to minimize error with respect to that judge and then evaluated by that same judge. The reported gains may reflect adaptation to the predictor's idiosyncrasies (e.g., keyword repetition, length) rather than improved recommendation quality. Section 4.7 amplifies the concern: human raters judged the DPO item recommendation information worse than SumRec on all four criteria, yet Table 1 reports the full method as best. The paper dismisses this because item information is 'internal,' but that information is a direct input to the predictor; if the text is wo
[§4.4, Table 1, §5] The results do not support the claim of superior performance at all rank cutoffs. Section 4.4 says the proposed method 'outperformed existing methods across all rank cutoffs' on Tabidachi, but Table 1 shows HR@5 is 0.7231 for Ours versus 0.7574 for SumRec. On ChatRec, HR@5 is 0.9933 for Ours versus 1.0 for SumRec. Moreover, Appendix C states that each model was trained five times and the averages are reported, but no standard deviations, confidence intervals, or significance tests are given. Please report variance and run pairwise significance tests, and qualify the claims to match the actual table values.
[§3.3.2, Eqs. (3)-(5)] The preference-pair construction for the dialogue summary model appears ill-posed. For each candidate item m, the 'best' summary is the one whose score is closest to that item's ground-truth score, and the 'worst' is the farthest. The same summary can therefore be a winner for one item and a loser for another, and the DPO pairs are not necessarily coherent preferences over summaries. Additionally, the absolute-distance criterion optimizes a regression objective, while the final evaluation is ranking-based (HR/MRR). The authors should either constrain the pair construction (e.g., aggregate preferences across items or use a ranking-aware margin) or provide evidence that the resulting preference pairs are consistent and aligned with the ranking objective.

minor comments (5)

[Eq. (5)] The notation in Eq. (5) is garbled: 'P Sn,snm,+' should be cleaned, and the subscript/superscript placement of sn m,+ should be made consistent.
[§4.5, Table 2] The interpretation of decreased BLEU/ROUGE-L as 'greater emphasis on explanatory elements' is presented without statistical support or comparison to human judgments. Please provide significance tests or qualitative evidence for this claim.
[Figure 3] The stage labels are confusing: Stage 1 trains the score predictor, Stage 2 trains the generator models. The arrows from the frozen models to the score predictor in Stage 1 are not visually distinguished from the trainable paths. Please revise for clarity.
[Table 4] The column headings 'T', 'E', 'N' for Tabidachi are not explained in the text or appendix. Please define them explicitly.
[§4.4] The phrase 'particularly significant performance improvements' is used although no significance tests are reported. Please rephrase or add significance testing.

Circularity Check

1 steps flagged · score 6.0 of 10

DPO reward and HR/MRR evaluation use the same frozen DeBERTa score; Section 4.7 shows human-judged item-info quality drops, so the headline gains are a closed-loop measure of fitting the predictor.

fitted input called prediction [Section 3.3.1-3.3.3 (Eqs. 3-5); Section 3.2.3; Section 4.2-4.4 (Table 1)]
"The preference data used for DPO training of the dialogue summary and item recommendation information generation models are created based on the output of the score predictor. Therefore, the score predictor is trained first."

Equations 3-5 define DPO winners/losers by the absolute difference |y - DeBERTa(s,r,d)|, selecting the generated text closest/furthest to the ground-truth score. The final recommendation evaluation (Sec. 4.2-4.4) ranks candidate items using the same frozen DeBERTa score (Eq. 1). Thus the DPO objective directly trains the LLM to produce text that moves the fixed predictor's output toward ground truth, and the reported HR/MRR measures exactly that same predictor-output gap. The 'prediction' is therefore not an independent test of recommendation quality; it is the optimized training objective renamed as evaluation. The paper's own human evaluation (Sec. 4.7) confirms this: for item recommendation information, SumRec is judged better on all four criteria after DPO, yet Table 1 credits DPO with

full rationale

The central claim of Section 5 ('our method achieved superior recommendation performance over baselines on both datasets') rests on a closed loop. Stage 1 trains DeBERTa on texts from the original, unfine-tuned Llama generator. Stage 2 uses that same DeBERTa to construct DPO preference pairs (Eqs. 3-5) and to score candidate items for HR/MRR. Consequently, the reported improvement tells us the generator has learned to satisfy DeBERTa's scoring function, not that it produces better recommendations in an external sense. The human evaluation in Section 4.7 is decisive supporting evidence: after DPO, item recommendation information is rated worse than SumRec's on Consistency, Conciseness, Fluency, and Usefulness, yet Table 1 reports the DPO model as best—exactly the signature of reward hacking on the frozen predictor. This is partial rather than full circularity because the DPO training could in principle fail to improve even the predictor's ranking (it is an empirical optimization, not a logical identity), so the paper still contains an experimental result. But that result is an in-distribution measure of fit to the training-time reward, not an independent first-principles prediction of recommendation quality. The reuse of SumRec (Asahara et al., 2023, a same-author baseline) is a self-citation, but it is not the load-bearing source of the closed loop; the loop is created by the paper's own equations. No other circular steps (self-definition, imported uniqueness, renamed known result) are present.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method introduces no new entities or forces. The key ledger entries are the DPO hyperparameters tuned to the validation set and the domain assumptions that the frozen score predictor is a faithful reward model and evaluator. The latter is the main source of circularity risk.

free parameters (7)

DPO beta for dialogue summary model (Tabidachi) = 0.1768
Selected by Optuna for best validation recommendation performance (Appendix C, Table 9).
DPO beta for item recommendation model (Tabidachi) = 0.06109
Selected by Optuna (Table 9).
DPO beta for dialogue summary model (ChatRec) = 0.1253
Selected by Optuna (Table 10).
DPO beta for item recommendation model (ChatRec) = 0.03949
Selected by Optuna (Table 10).
Learning rates for DPO models = 1.1593e-7 and 8.7340e-6 (Tabidachi), 6.4087e-7 and 1.7718e-7 (ChatRec)
Selected via Optuna; see Tables 9 and 10.
Chunk size for partial dialogue summaries = 30 utterances
Chosen by hand for Tabidachi; see Appendix B.1.
ChatRec like/dislike score threshold = 3 (like), 2 or less (dislike)
Converted 5-point human-predicted scores to binary labels; Section A.2.

assumptions (5)

standard math DPO loss (Eq. 5) is a valid training objective and the reference policy contributes as intended.
Relies on Rafailov et al. 2023; used as a black-box optimizer.
domain assumption Ground-truth label y=1 for the item in the next operator utterance and y=0 for others is a faithful measure of recommendation quality.
Section 3.2.3 defines the regression target; any labeling noise or ambiguity in which item is 'correct' propagates into the DPO signals.
domain assumption The frozen score predictor trained in Stage 1 is an accurate reward model for texts generated after DPO modification.
Section 3.3.1 trains DeBERTa on texts from the original LLM; Section 3.3.2/3.3.3 then selects preference pairs using this frozen predictor, assuming its scores remain meaningful for out-of-distribution DPO texts.
domain assumption The argmin/argmax winner-loser selection over |y - y_hat| produces preference pairs that capture text quality for recommendation.
Equations 3 and 4 define the preference data; if the distance metric is noisy, the DPO signal is noisy.
domain assumption HR@N and MRR computed with the frozen predictor reflect real-world recommendation quality.
Section 4.2 uses HR/MRR over candidate lists; the same predictor that generated the DPO preferences is used for ranking, so this assumption is the key evaluative premise.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization." pith.science (2026). https://pith.science/paper/E7OH5DLI

@misc{pith2026250819918,
  author       = {Pith},
  title        = {Pith review of: Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7OH5DLI}},
  note         = {Machine review of arXiv:2508.19918}
}

read the original abstract

Conversational Recommender Systems (CRSs) aim to elicit user preferences via natural dialogue to provide suitable item recommendations. However, current CRSs often deviate from realistic human interactions by rapidly recommending items in brief sessions. This work addresses this gap by leveraging Large Language Models (LLMs) to generate dialogue summaries from dialogue history and item recommendation information from item description. This approach enables the extraction of both explicit user statements and implicit preferences inferred from the dialogue context. We introduce a method using Direct Preference Optimization (DPO) to ensure dialogue summary and item recommendation information are rich in information crucial for effective recommendations. Experiments on two public datasets validate our method's effectiveness in fostering more natural and realistic conversational recommendation processes. Our implementation is publicly available at: https://github.com/UEC-InabaLab/Refining-LLM-Text

Figures

Figures reproduced from arXiv: 2508.19918 by the authors.

**Figure 2.** Item recommendation flow in SumRec. Dia [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Process flow of the proposed method during training. The method employs a two-stage training procedure. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Human evaluation of the proposed method and SumRec on Tabidachi Corpus, assessing dialogue summaries and item recommendation information. evaluators rated our summaries as superior, with “Usefulness” showing the most significant difference. This suggests DPO enhanced …

**Figure 5.** Figure 5: Requests to Crowd Workers E Case Study [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 7.** Figure 7: Requests to Crowd Workers(English Version, [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

**Figure 8.** Figure 8: Crowdworker response screen (English Ver [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 21 canonical work pages

[1]

Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

work page 2019
[2]

Amatriain and Justin D

X. Amatriain and Justin D. Basilico. 2015. https://api.semanticscholar.org/CorpusID:43176140 Recommender systems in industry: A netflix case study . In Recommender Systems Handbook

work page 2015
[3]

Ryutaro Asahara, Masaki Takahashi, Chiho Iwahashi, and Michimasa Inaba. 2023. https://aclanthology.org/2023.paclic-1.35 S um R ec: A framework for recommendation using open-domain dialogue . In Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation, pages 349--363, Hong Kong, China. Association for Computational Linguistics

work page 2023
[4]

Keyu Chen and Shiliang Sun. 2023. https://doi.org/10.1609/aaai.v37i11.26487 Cp-rec: contextual prompting for conversational recommender systems . In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances ...

doi:10.1609/aaai.v37i11.26487 2023
[5]

Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. https://doi.org/10.18653/v1/2023.acl-long.4 Explainable recommendation with personalized review retrieval and aspect learning . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 51--64, Toront...

work page doi:10.18653/v1/2023.acl-long.4 2023
[6]

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sashank Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, and 48 others. 2023. Palm: scaling language modeling with ...

work page 2023
[7]

Konstantina Christakopoulou, Filip Radlinski, and Katja Hofmann. 2016. https://doi.org/10.1145/2939672.2939746 Towards conversational recommender systems . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, page 815–824, New York, NY, USA. Association for Computing Machinery

arXiv 2016
[8]

Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual llm adaptation: Enhancing japanese language capabilities. In Proceedings of the First Conference on Language Modeling, COLM, page (to appear), University of Pennsy...

2024

Show all 43 references

[9]

Gomez-Uribe and Neil Hunt

Carlos A. Gomez-Uribe and Neil Hunt. 2016. https://doi.org/10.1145/2843948 The netflix recommender system: Algorithms, business value, and innovation . ACM Trans. Manage. Inf. Syst., 6(4)

2016 doi
[10]

Shirley Anugrah Hayati, Dongyeop Kang, Qingxiaoyang Zhu, Weiyan Shi, and Zhou Yu. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.654 INSPIRED : Toward sociable recommendation dialog systems . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Pro...

2020 doi
[11]

Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. https://openreview.net/forum?id=XPZIaotutsD \ DEBERTA \ : \ DECODING \ - \ enhanced \ \ bert \ \ with \ \ disentangled \ \ attention \ . In International Conference on Learning Representations

2021
[12]

Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian Mcauley. 2023. https://doi.org/10.1145/3583780.3614949 Large language models as zero-shot conversational recommenders . In Proceedings of the 32nd ...

2023
[13]

Michimasa Inaba, Yuya Chiba, Zhiyang Qi, Ryuichiro Higashinaka, Kazunori Komatani, Yusuke Miyao, and Takayuki Nagai. 2024. https://doi.org/10.1145/3675166 Travel agency task dialogue corpus: A multimodal dataset with age-diverse speakers . ACM Trans. Asian Low-Resour. Lang. In...

2024 doi
[14]

Minjin Kim, Minju Kim, Hana Kim, Beong-woo Kwak, SeongKu Kang, Youngjae Yu, Jinyoung Yeo, and Dongha Lee. 2024. https://doi.org/10.18653/v1/2024.findings-acl.65 Pearl: A review-driven persona-knowledge grounded conversational recommendation dataset . In Findings of the Associa...

2024 doi
[15]

Heejin Kook, Junyoung Kim, Seongmin Park, and Jongwuk Lee. 2025. https://aclanthology.org/2025.naacl-long.392/ Empowering retrieval-based conversational recommendation with contrasting user preferences . In Proceedings of the 2025 Conference of the Nations of the Americas Chap...

2025
[16]

Chuang Li, Yang Deng, Hengchang Hu, Min-Yen Kan, and Haizhou Li. 2025. https://aclanthology.org/2025.findings-naacl.17/ C hat CRS : Incorporating external knowledge and goal guidance for LLM -based conversational recommender systems . In Findings of the Association for Computa...

2025
[17]

Jiacheng Li, Zhankui He, Jingbo Shang, and Julian McAuley. 2023. https://doi.org/10.1145/3580305.3599535 Ucepic: Unifying aspect planning and lexical constraints for generating explanations in recommendation . In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Disco...

2023
[18]

Raymond Li, Samira Ebrahimi Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, and Chris Pal. 2018. https://proceedings.neurips.cc/paper_files/paper/2018/file/800de15c79c8d840f4e78d3af937d4d4-Paper.pdf Towards deep conversational recommendations . In Advances in Neural ...

2018
[19]

Tingting Liang, Chenxin Jin, Lingzhi Wang, Wenqi Fan, Congying Xia, Kai Chen, and Yuyu Yin. 2024. https://doi.org/10.18653/v1/2024.findings-acl.529 LLM - REDIAL : A large-scale dataset for conversational recommender systems created from user behaviors with LLM s . In Findings ...

2024 doi
[20]

Dongding Lin, Jian Wang, and Wenjie Li. 2023. https://doi.org/10.1609/aaai.v37i4.25567 Cola: improving conversational recommender systems by collaborative augmentation . In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference...

2023 doi
[21]

Linden, B

G. Linden, B. Smith, and J. York. 2003. https://doi.org/10.1109/MIC.2003.1167344 Amazon.com recommendations: item-to-item collaborative filtering . IEEE Internet Computing, 7(1):76--80

2003 arXiv
[22]

Zeming Liu, Haifeng Wang, Zheng-Yu Niu, Hua Wu, and Wanxiang Che. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.356 D u R ec D ial 2.0: A bilingual parallel corpus for conversational recommendation . In Proceedings of the 2021 Conference on Empirical Methods in Natural Lan...

2021 doi
[23]

Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Chris Leung, Jiajie Tang, and Jiebo Luo. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.39 LLM -rec: Personalized recommendation via prompting large language models . In Findings of the ...

2024 doi
[24]

Qiyao Ma, Xubin Ren, and Chao Huang. 2024. https://arxiv.org/abs/2406.02377 Xrec: Large language models for explainable recommendation . Preprint, arXiv:2406.02377

2024 arXiv
[25]

Wenchang Ma, Ryuichi Takanobu, and Minlie Huang. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.139 CR -walker: Tree-structured graph reasoning and dialog acts for conversational recommendation . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language...

2021 doi
[26]

Naoaki Okazaki, Kakeru Hattori, Hirai Shota, Hiroki Iida, Masanari Ohi, Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Rio Yokota, and Sakae Mizuki. 2024. Building a large japanese web corpus for large language models. In Proceedings of the First Conference on Language Modeling,...

2024
[27]

Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf Improving language understanding by generative pre-training . Technical report, OpenAI

2018
[28]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. https://openreview.net/forum?id=HPuSIXJaa9 Direct preference optimization: Your language model is secretly a reward model . In Thirty-seventh Conference on Neural Infor...

2023
[29]

Mathieu Ravaut, Hao Zhang, Lu Xu, Aixin Sun, and Yong Liu. 2024. https://aclanthology.org/2024.eacl-long.9/ Parameter-efficient conversational recommender system as a language processing task . In Proceedings of the 18th Conference of the European Chapter of the Association fo...

2024
[30]

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. https://arxiv.org/abs/2302.13971 Llama:...

2023 arXiv
[31]

Xiaolei Wang, Xinyu Tang, Xin Zhao, Jingyuan Wang, and Ji-Rong Wen. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.621 Rethinking the evaluation for conversational recommendation in the era of large language models . In Proceedings of the 2023 Conference on Empirical Method...

2023 doi
[32]

Xiaolei Wang, Kun Zhou, Ji-Rong Wen, and Wayne Xin Zhao. 2022. https://doi.org/10.1145/3534678.3539382 Towards unified conversational recommender systems via knowledge-enhanced prompt learning . In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data M...

2022
[33]

Bowen Yang, Cong Han, Yu Li, Lei Zuo, and Zhou Yu. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.4 Improving conversational recommendation systems' quality with context-aware item meta-information . In Findings of the Association for Computational Linguistics: NAACL 20...

2022 doi
[34]

Ching-Wen Yang, Che Wei Chen, Kun da Wu, Hao Xu, Jui-Feng Yao, and Hung-Yu Kao. 2024. https://arxiv.org/abs/2408.09865 Maple: Enhancing review generation with multi-aspect prompt learning in explainable recommendation . Preprint, arXiv:2408.09865

2024 arXiv
[35]

Ting Yang and Li Chen. 2024. https://doi.org/10.1145/3640457.3688146 Unleashing the retrieval potential of large language models in conversational recommender systems . In Proceedings of the 18th ACM Conference on Recommender Systems, RecSys '24, page 43–52, New York, NY, USA....

2024
[36]

Yusen Zhang, Ansong Ni, Ziming Mao, Chen Henry Wu, Chenguang Zhu, Budhaditya Deb, Ahmed Awadallah, Dragomir Radev, and Rui Zhang. 2022. https://doi.org/10.18653/v1/2022.acl-long.112 S umm ^n : A multi-stage summarization framework for long input dialogues and documents . In Pr...

2022 doi
[37]

Ming Zhong, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. 2022. https://arxiv.org/abs/2109.02492 Dialoglm: Pre-trained model for long dialogue understanding and summarization . Preprint, arXiv:2109.02492

2022 arXiv
[38]

Jinfeng Zhou, Bo Wang, Ruifang He, and Yuexian Hou. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.355 CRFR : Improving conversational recommender systems via flexible fragments reasoning on knowledge graphs . In Proceedings of the 2021 Conference on Empirical Methods in Na...

2021 doi
[39]

Kun Zhou, Yuanhang Zhou, Wayne Xin Zhao, Xiaoke Wang, and Ji-Rong Wen. 2020. https://doi.org/10.18653/v1/2020.coling-main.365 Towards topic-guided conversational recommender system . In Proceedings of the 28th International Conference on Computational Linguistics, pages 4128--...

2020 doi
[40]

Rongxin Zhu, Jey Han Lau, and Jianzhong Qi. 2025 a . https://aclanthology.org/2025.coling-main.302/ Factual dialogue summarization via learning from large language models . In Proceedings of the 31st International Conference on Computational Linguistics, pages 4474--4492, Abu ...

2025
[41]

Yaochen Zhu, Chao Wan, Harald Steck, Dawen Liang, Yesu Feng, Nathan Kallus, and Jundong Li. 2025 b . https://doi.org/10.1145/3696410.3714908 Collaborative retrieval for large language model-based conversational recommender systems . In Proceedings of the ACM on Web Conference ...

2025
[42]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[43]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

work page 2019

[2] [2]

Amatriain and Justin D

X. Amatriain and Justin D. Basilico. 2015. https://api.semanticscholar.org/CorpusID:43176140 Recommender systems in industry: A netflix case study . In Recommender Systems Handbook

work page 2015

[3] [3]

Ryutaro Asahara, Masaki Takahashi, Chiho Iwahashi, and Michimasa Inaba. 2023. https://aclanthology.org/2023.paclic-1.35 S um R ec: A framework for recommendation using open-domain dialogue . In Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation, pages 349--363, Hong Kong, China. Association for Computational Linguistics

work page 2023

[4] [4]

Keyu Chen and Shiliang Sun. 2023. https://doi.org/10.1609/aaai.v37i11.26487 Cp-rec: contextual prompting for conversational recommender systems . In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances ...

doi:10.1609/aaai.v37i11.26487 2023

[5] [5]

Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. https://doi.org/10.18653/v1/2023.acl-long.4 Explainable recommendation with personalized review retrieval and aspect learning . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 51--64, Toront...

work page doi:10.18653/v1/2023.acl-long.4 2023

[6] [6]

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sashank Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, and 48 others. 2023. Palm: scaling language modeling with ...

work page 2023

[7] [7]

Konstantina Christakopoulou, Filip Radlinski, and Katja Hofmann. 2016. https://doi.org/10.1145/2939672.2939746 Towards conversational recommender systems . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, page 815–824, New York, NY, USA. Association for Computing Machinery

arXiv 2016

[8] [8]

Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual llm adaptation: Enhancing japanese language capabilities. In Proceedings of the First Conference on Language Modeling, COLM, page (to appear), University of Pennsy...

2024