Improved Personalized Headline Generation via Denoising Fake Interests from Implicit Feedback

Junhong Lian; Kejin Liu; Ningtao Wang; Weiqiang Wang; Xiang Ao; Xing Fu; Xinyu Liu; Yu Cheng

REVIEW 4 major objections 3 minor 35 references

Improved Personalized Headline Generation via Denoising Fake Interests from Implicit Feedback

T0 review · 4 major / 3 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 Filtering out noisy clicks by dwell time and breaking-news bursts improves personalized headline generation, and the resulting PHG-DIF framework reaches state-of-the-art scores on the new DT-PENS benchmark.

desk verdict A plausible denoising framework and a genuinely new dwell-time benchmark, but the SOTA claim is weakened because the benchmark references were generated with the same dwell-time signal the model consumes. read the letter →

arxiv 2508.07178 v2 pith:ITBMELCQ submitted 2025-08-10 cs.CL cs.AI

Kejin Liu , Junhong Lian , Xiang Ao , Ningtao Wang , Xing Fu , Yu Cheng , Weiqiang Wang , Xinyu Liu This is my paper · ORCID

classification cs.CLcs.AI

keywords personalizedheadlinegenerationclicknoisedwelltimeimplicitfeedbackuserinterestmodelingbreakingnewspointer-generatornetworkDT-PENSbenchmark

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

Serving a reader the right news headline depends on knowing what that reader cares about; this paper argues that current systems misread their evidence. They treat every click as an expression of interest, even though many clicks are misclicks, instant exits, or reactions to trending stories that the reader does not actually care about. The authors show that dwell time — how long the reader stayed on the article — separates genuine interest from this "click noise," and that filtering out sub-10-second clicks and clicks on breaking-news items measurably improves headline quality. On that basis they build PHG-DIF, which denoises the clickstream, models instantaneous, evolving, and stable interests separately, and reduces personalization when the candidate article is itself breaking news. They also release DT-PENS, a benchmark of 1,000 users with dwell-time-annotated histories and nearly 10,000 human-validated personalized headlines, where PHG-DIF reports state-of-the-art scores. If the paper is right, interest is better measured by lingering than by clicking.

What carries the argument

Three components carry the argument. News-level filtering removes the top 0.1% click-through-rate news (the breaking set) from each user's history by masking their dwell times to zero; time-level filtering then re-weights what remains through three encoders — IPL (last $K=30$ clicks), IEA (one-week sliding window), SIM (clicks above the user's mean dwell time) — whose vectors dynamic attention fuses into $e_u$. On the generating side, a pointer-generator decoder (one that can copy words from the article or generate new ones) uses a BERT breaking-news classifier to produce $\alpha=B_\psi(v)$, and the gate $\lambda_t=\sigma(W_\lambda[c_t;s_t;\alpha e_u]+b_\lambda)$ mixes copying and generating

What would settle it

Regenerate the DT-PENS reference headlines with the same LLM pipeline but with dwell times removed from the prompts (keeping click order and the breaking-news filter), then compare PHG-DIF against the strongest baseline, GTP. If the reported margin largely disappears, the headline-quality gain is an artifact of the benchmark encoding the model's own dwell-time assumption; if the margin holds, the denoising mechanism is validated independently of how the references were written. A complementary check is blind A/B ranking by real readers of their own histories, pitting filtered-profile headlines

Watch

Extended reading notes

Core claim

Click equals interest is the false premise this paper attacks: in a sampled user's history, about 28% of clicks lasted under 10 seconds — too short to read even 100 words — and transient events create click bursts from users who were not really interested. PHG-DIF removes both noise types by dual-stage filtering: it drops the top 0.1% most-clicked breaking news from histories and re-weights the rest by dwell time through three time-aware encoders, fused into one user vector; a breaking-news classifier then gates each decoding step between copying facts and rewriting toward the profile. On DT-PENS it reports state-of-the-art scores (ROUGE-1 24.33, ROUGE-2 7.99, ROUGE-L 22.47, BLEURT 48.50, BA

Load-bearing premise

The load-bearing premise is that a click's dwell time and a news item's popularity reliably separate genuine interest from noise — and that DT-PENS's reference headlines, written by LLMs shown dwell times and then human-verified (Section 5.1), measure true user preference rather than echoing the model's own dwell-time logic; the paper itself (Appendix A) leaves sparse-interaction, cold-start, and cross-platform cases unaddressed.

Editorial extensions

If this is right

Filtering clicks with sub-10-second dwell times and clicks on breaking-news items measurably improves personalized headline quality on all five reported metrics.
Long-term stable interests carry more weight than recent or evolving ones: removing the stable-interest module costs the most in the ablations.
Breaking news should be reported factually rather than personalized: an inference-time predictor that switches the model toward copying improves quality over personalizing everything.
DT-PENS gives the field a shared, dwell-time-annotated benchmark (1,000 users, 9,823 instances) for studying click noise in generation, not just in recommendation.
Click noise harms headline generation, not merely ranking: noisy histories cause generated headlines to hallucinate content that matches no real user interest.

Reading between the lines

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

The benchmark's ground truth was written by LLMs that were shown dwell times and few-shot personalized examples, then human-verified (Section 5.1). If those references encode the same "short dwell means noise" logic the model uses, part of the state-of-the-art margin may reflect the benchmark validating the model's own assumptions; the paper's experiments do not separate these effects.
The three thresholds — 10 seconds, top-0.1% CTR, and mean dwell time — are hand-set rather than learned. Per-user calibration or a learned noise classifier could extend the same idea to sparse histories, cold-start users, and other platforms, which the paper lists as limitations.
The same two-signal denoising (linger time plus popularity bursts) transfers beyond headlines to any implicit-feedback personalization task — recommendation ranking, email subject lines, ad copy — wherever a click is not proof of interest.
A direct test: regenerate DT-PENS references with dwell times hidden from the LLM prompts, then re-run PHG-DIF against GTP. If the margin collapses, the gain is largely benchmark self-fulfillment; if it survives, the dwell-time filter is independently validated.

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

You should know two things before reading this one. First, the paper does something real: it builds DT-PENS, a new personalized-headline benchmark with dwell-time annotations for 1,000 users and ~10k human-validated references, and it proposes PHG-DIF, a clean dual-filtering plus multi-granular temporal fusion framework. Second, the headline result is not as solid as it looks. The benchmark construction and the model share the same causal assumption, so the reported SOTA advantage is partly self-fulfilling.

What is actually new and good: DT-PENS is a useful resource for a subfield that has lacked dwell-time data. The framework is sensible: filter clicks by short dwell time and breaking-news burstiness, then model instantaneous, evolving, and stable interests with three encoders and a dynamic fusion. The ablations show each module contributes, and the authors are transparent about many limitations. The paper is clearly written and builds directly on prior PENS/PNG/GTP work.

The soft spots are real and concentrated in evaluation. In Appendix B, the LLM prompt that generates every reference headline receives clicked headlines, dwell times, exposure times, and unclicked headlines, and is explicitly asked to identify user interests from that history. Human annotators then choose the headline that best reflects user preferences. That process encodes the same dwell-time heuristic PHG-DIF uses, so baselines that ignore dwell time are structurally disadvantaged. The claim that the LLM was "not explicitly informed" of the correlation is not convincing; providing the signal and asking the model to infer interests is enough. The paper also lacks confidence intervals or multiple seeds, selects hyperparameters K and M on the final test set, leaves the reward function R(Y) unspecified, and motivates the whole approach with a single-user case study.

None of this makes the framework wrong. It makes the SOTA claim premature. The fix is straightforward: validate on references built without dwell time in the prompt, or on the original PENS test set, or with a larger external human preference study. The current user study has five participants and simulated profiles, so it cannot carry that weight.

Who is this for? Researchers working on personalized headline generation or benchmark construction for implicit feedback. It deserves a serious referee, but the referee should insist on external validation before the empirical claims are taken at face value. I would not cite the benchmark as a gold standard until that circularity is addressed.

Referee Report

4 major / 3 minor

Summary. The paper argues that historical clickstreams contain personalized-irrelevant click noise, identifiable through short dwell times (user dimension) and abnormal click bursts on breaking news (news dimension). It proposes PHG-DIF, which applies dual-stage filtering, three time-aware encoders (IPL, IEA, SIM), dynamic attention fusion, and a breaking-news-aware pointer-generator, trained with MLE plus A2C policy-gradient fine-tuning. The authors release DT-PENS, a 1,000-user / 9,823-instance benchmark with dwell-time annotations and LLM-generated, human-filtered personalized headlines. Experiments report SOTA over non-personalized and personalized baselines, plus ablations, a small user study, and a case study.

Significance. If the results hold, the paper makes a useful empirical contribution: it releases DT-PENS with dwell-time annotations and a codebase, and its ablations show that removing any of the three time-aware encoders degrades performance, with SIM most important. The dual-filtering idea is well motivated by the click-noise literature. However, the benchmark-based SOTA claim is currently compromised by the self-referential construction of DT-PENS, hyperparameter selection on the test set, and the unspecified reward function. If these are remedied, the paper could be a valuable resource for the community.

major comments (4)

[§5.1 and Appendix B (Fig. 4a)] The DT-PENS reference construction is not an independent criterion for the paper's central claim. The LLM prompt receives the user's clicked headlines, dwell times, exposure times, and unclicked headlines, and explicitly instructs the model to 'Identify user interests from the provided history' and 'Match user interests to news content.' PHG-DIF's own novelty is that short dwell times and breaking-news bursts are click noise; the reference headlines are therefore generated under the same assumption that dwell time is a personalization signal, and human annotators then select the headline 'that best reflected the user's historical preferences' from this pool. Baselines that do not exploit dwell time cannot access this signal, so the reported ROUGE/BLEURT/BARTScore advantage may be an artifact of benchmark construction. A concrete remedy: construct a second reference set from prompts that
[§6.2.3–6.2.4] The hyperparameters K (IPL history length) and M (breaking-news CTR threshold) are selected by performance on the final DT-PENS benchmark (K peaks at 30 in Fig. 3a; M=0.10 is optimal in Fig. 3b). No held-out validation split or nested procedure is described. Choosing hyperparameters on the evaluation set inflates the reported scores and undermines the SOTA comparison. Please report results from a validation split and, ideally, multiple random seeds with variance estimates.
[§4.4, Eq. (11)] The policy-gradient objective L_RL = -E[R(Y)] is never specified: R(Y) is said to 'aggregate several headline-quality indicators' but no definition or list of indicators is given. This matters because if R(Y) includes ROUGE/BLEURT/BARTScore computed against DT-PENS references, the reward function overlaps with the evaluation metrics, creating train/eval leakage. Please define R(Y) precisely and state which components are used during training and which are withheld from tuning.
[Table 1 and Table 3] The claim of 'significantly improves' is not backed by adequate statistics. Table 1 reports p≤0.05 but gives no number of seeds, confidence intervals, or paired-test details; Table 3 is based on 5 participants and 20 articles each, with no inter-annotator agreement or significance test. Please provide run-level results, a statistical test with multiple seeds, and details for the user study.

minor comments (3)

[§1 and Figure 1] The motivating analysis uses one randomly sampled user (U362229); 'rigorous analysis' overstates the evidence. Please report aggregate statistics over many users, with error bars, before concluding that dwell-time filtering generally improves personalization.
[Table 3] The header uses downward arrows for all dimensions, but the text says PHG-DIF achieved the 'highest scores'; clarify that lower ranks are better for all three dimensions. Also specify whether ties are allowed and how ranks were normalized.
[§2 and author block] The related-work section contains unresolved citation placeholders 'sequential attention mechanisms [? ?]'. In addition, the author affiliation block contains a stray 'pilcrow' symbol that appears to be a formatting artifact.

Circularity Check

2 steps flagged · score 6.0 of 10

DT-PENS ground-truth headlines are generated from dwell-time-annotated histories via LLM prompts, and key hyperparameters are selected on the same test set, making the SOTA claim partly self-fulfilling.

self definitional [Section 5.1.1 and Appendix B, Figure 4a (compare Section 4.1)]
"To ensure fairness, LLMs are not explicitly informed of the correlation between user click history and dwell time. Instead, we adopted a few-shot prompting strategy... Inputs: - User history: { clicked headlines , dwell times , exposure times, uninterested headlines} ... Steps: 1. Identify user interests from the provided history. ... 3. Match user interests to news content."

PHG-DIF's central denoising mechanism is to treat short dwell times and abnormal click bursts as click noise (Sections 1 and 4.1). The DT-PENS reference headlines are produced by LLMs that are given dwell times and explicitly instructed to 'identify user interests from the provided history' and to 'match user interests to news content.' Thus the ground-truth labels are generated under the same dwell-time-is-interest assumption that PHG-DIF implements. Human annotators then select 'the headline that best reflected the user's historical preferences' from the same dwell-time-annotated history. A model that exploits dwell-time filtering is therefore aligned with the label-generation process, while baselines that ignore dwell times are systematically disadvantaged. The reported SOTA gain is par
fitted input called prediction [Sections 6.2.3, 6.2.4, and Table 1]
"As illustrated in Figure 3a, our experimental results show that varying K across the set {5, 10, 20, 30, 40, 50} leads to a rapid improvement in model performance with increasing K, which peaks at K=30. ... The results revealed optimal model performance at M=0.10 (corresponding to 0.10% CTR threshold)."

The final SOTA numbers in Table 1 are obtained after selecting the IPL history length K = 30 and the breaking-news CTR threshold M = 0.10 by maximizing performance on the DT-PENS evaluation set. No separate held-out validation split is described. The same DT-PENS test set is then used to report the headline results. This means the reported improvements are not independent held-out predictions; the two key design choices, including the definition of breaking news, are fitted to the evaluation data. The comparison is therefore statistically forced and the SOTA claim is not fully external.

full rationale

The paper's internal model derivation—dual-stage dwell-time/breaking-news filtering, multi-granular temporal fusion, and pointer-generator decoding—is largely self-contained and is not circular in its equations. However, the central empirical claim of SOTA performance rests on DT-PENS, whose reference headlines are generated by LLMs that receive dwell-time-annotated histories and are prompted to infer user interests, which is exactly the signal PHG-DIF uses to denoise clicks. This makes the benchmark partly self-referential: a dwell-time-aware model is rewarded by construction, while dwell-time-ignoring baselines are handicapped. Additionally, the two most important hyperparameters (K and M) are selected by performance on the same DT-PENS test set used for final reporting, further compromising the independence of the headline results. These issues are not mere speculation; they follow directly from the quoted construction details and sensitivity analyses. The method may still be a useful engineering contribution, and the small human user study provides some external signal, but the paper's headline SOTA claim is partially circular, so a score of 6 is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical entities (particles, forces, dimensions) are introduced. 'Click noise', 'fake interests', and 'breaking-news set' are conceptual constructs used to structure the method, not independent entities with falsifiable handles.

free parameters (6)

Dwell time noise threshold = 10 seconds
Clicks with dwell time under 10 seconds are removed as noise in time-level filtering (Section 4.1). Stated without sensitivity analysis.
Breaking-news CTR threshold M = 0.10%
News in the top 0.10% of CTR are treated as breaking news and filtered out; chosen by sensitivity analysis on DT-PENS (Section 6.2.4).
IPL history length K = 30
Number of recent clicks used for instantaneous preference learning; chosen by sensitivity analysis on DT-PENS (Section 6.2.3).
IEA time window n = 1 week
Time window for interest evolution analysis; set by hand with no sensitivity analysis (Section 5.4).
SIM dwell-time outlier cutoff = 3000s
Dwell times above 3000s are excluded when computing the mean threshold for stable interests (Section 5.4).
Reward function R(Y) = unspecified
The policy-gradient reward is described only as 'several headline-quality indicators' with no explicit composition (Section 4.4).

assumptions (5)

domain assumption Dwell time is a reliable proxy for genuine user interest
Introduced in Section 1 and used by all three time-aware encoders.
ad hoc to paper Clicks with dwell time under 10 seconds are click noise
Stated in Section 1; underlies time-level filtering.
ad hoc to paper News in the top 0.1% of CTR are breaking news, and clicks on them are noise
Defined in Section 4.1, tuned in Section 6.2.4.
domain assumption User interests decompose into instantaneous, evolving, and stable components
The IPL/IEA/SIM design in Section 4.1 assumes this decomposition.
ad hoc to paper LLM-generated and human-filtered headlines in DT-PENS are valid ground truth for personalization
Dataset construction in Section 5.1 relies on this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improved Personalized Headline Generation via Denoising Fake Interests from Implicit Feedback." pith.science (2026). https://pith.science/paper/ITBMELCQ

@misc{pith2026250807178,
  author       = {Pith},
  title        = {Pith review of: Improved Personalized Headline Generation via Denoising Fake Interests from Implicit Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITBMELCQ}},
  note         = {Machine review of arXiv:2508.07178}
}

read the original abstract

Accurate personalized headline generation hinges on precisely capturing user interests from historical behaviors. However, existing methods neglect personalized-irrelevant click noise in entire historical clickstreams, which may lead to hallucinated headlines that deviate from genuine user preferences. In this paper, we reveal the detrimental impact of click noise on personalized generation quality through rigorous analysis in both user and news dimensions. Based on these insights, we propose a novel Personalized Headline Generation framework via Denoising Fake Interests from Implicit Feedback (PHG-DIF). PHG-DIF first employs dual-stage filtering to effectively remove clickstream noise, identified by short dwell times and abnormal click bursts, and then leverages multi-level temporal fusion to dynamically model users' evolving and multi-faceted interests for precise profiling. Moreover, we release DT-PENS, a new benchmark dataset comprising the click behavior of 1,000 carefully curated users and nearly 10,000 annotated personalized headlines with historical dwell time annotations. Extensive experiments demonstrate that PHG-DIF substantially mitigates the adverse effects of click noise and significantly improves headline quality, achieving state-of-the-art (SOTA) results on DT-PENS. Our framework implementation and dataset are available at https://github.com/liukejin-up/PHG-DIF.

Figures

Figures reproduced from arXiv: 2508.07178 by the authors.

**Figure 2.** Overview of the proposed PHG-DIF framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Impact of IPL history length 𝐾 and breaking news threshold 𝑀 on model performance [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Prompt templates for instructing LLMs in the DT-PENS dataset construction. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages

[1]

Xiang Ao, Ling Luo, Xiting Wang, Zhao Yang, Jiun-Hung Chen, Ying Qiao, Qing He, and Xing Xie. 2023. Put Your Voice on Stage: Personalized Headline Generation for News Articles.ACM Transactions on Knowledge Discovery from Data18, 3 (2023), 1–20

work page 2023
[2]

Xiang Ao, Xiting Wang, Ling Luo, Ying Qiao, Qing He, and Xing Xie. 2021. PENS: A dataset and generic framework for personalized news headline generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 82–92

2021
[3]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Jill Burstein, Christy...

2019
[4]

Yijiang River Dong, Tiancheng Hu, and Nigel Collier. 2024. Can LLM be a Personalized Judge?arXiv preprint arXiv:2406.11657(2024)

arXiv 2024
[5]

Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A fam- ily of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793(2024)

arXiv 2024
[6]

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276(2024)

arXiv 2024
[7]

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720(2024)

arXiv 2024
[8]

Hao Jiang, Chuanzhen Li, and Mingxiao An. 2024. Time Matters: Enhancing Pre- trained News Recommendation Models with Robust User Dwell Time Injection. arXiv preprint arXiv:2405.12486(2024)

work page Pith review arXiv 2024

Show all 35 references

[9]

Hyoung R Kim and Philip K Chan. 2003. Learning implicit user interest hierarchy for context in personalization. InProceedings of the 8th international conference on Intelligent user interfaces. 101–108

2003
[10]

Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension.. InProceedings of the 58t...

2020 doi
[11]

Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. 2024. Llms-as-judges: a comprehensive survey on llm-based evaluation methods.arXiv preprint arXiv:2412.05579(2024)

2024 arXiv
[12]

Zhengpeng Li, Jiansheng Wu, Jiawei Miao, and Xinmiao Yu. 2022. News headline generation based on improved decoder from transformer.Scientific Reports12, 1 (2022), 11648

2022
[13]

Junhong Lian, Xiang Ao, Xinyu Liu, Yang Liu, and Qing He. 2025. Panoramic In- terests: Stylistic-Content Aware Personalized Headline Generation. InCompanion Proceedings of the ACM on Web Conference 2025. 1109–1112

2025
[14]

Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. InText summarization branches out. 74–81

2004
[15]

Ling Luo, Xiang Ao, Yan Song, Feiyang Pan, Min Yang, and Qing He. 2019. Reading like HER: Human reading inspired extractive summarization. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Nat...

2019
[16]

Ramesh Nallapati, Bowen Zhou, Cicero Nogueira Dos Santos, Caglar Gulcehre, and Bing Xiang. 2016. Abstractive Text Summarization using Sequence-to- sequence RNNs and Beyond. InProceedings of the 20th SIGNLL Conference on Computational Natural Language Learning. 280–290

2016
[17]

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (...

2022
[18]

Tao Qi, Fangzhao Wu, Chuhan Wu, Peiru Yang, Yang Yu, Xing Xie, and Yongfeng Huang. 2021. HieRec: Hierarchical User Interest Modeling for Personalized News Recommendation. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Int...

2021
[19]

Mingjie Qian, Yongsen Zheng, Jinghui Qin, and Liang Lin. 2023. HutCRS: Hi- erarchical user-interest tracking for conversational recommender system. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 10281–10290

2023
[20]

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research21, 140 (2020), 1–67. http:/...

2020
[21]

Keith Rayner, Elizabeth R Schotter, Michael EJ Masson, Mary C Potter, and Rebecca Treiman. 2016. So much to read, so little time: How do we read, and can speed reading help?Psychological Science in the Public Interest17, 1 (2016), 4–34

2016
[22]

Alexander M Rush, Sumit Chopra, and Jason Weston. 2015. A Neural Atten- tion Model for Abstractive Sentence Summarization. InProceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. 379–389

2015
[23]

Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7370–7392

2024
[24]

Liu, and Christopher D

Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get To The Point: Summarization with Pointer-Generator Networks. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). doi:10.18653/v1/p17-1099

2017 doi
[25]

Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning Robust Metrics for Text Generation. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 7881–7892

2020
[26]

Yun-Zhu Song, Yi-Syuan Chen, Lu Wang, and Hong-Han Shuai. 2023. General then Personal: Decoupling and Pre-training for Personalized Headline Generation. Transactions of the Association for Computational Linguistics11 (2023), 1588–1607

2023
[27]

Xiaoyu Tan, Leijun Cheng, Xihe Qiu, Shaojie Shi, Yuan Cheng, Wei Chu, Yinghui Xu, and Yuan Qi. 2024. Enhancing Personalized Headline Generation via Offline Goal-conditioned Reinforcement Learning with Large Language Models. InPro- ceedings of the 30th ACM SIGKDD Conference on ...

2024
[28]

Wenjie Wang, Fuli Feng, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2021. Denoising implicit feedback for recommendation. InProceedings of the 14th ACM international conference on web search and data mining. 373–381

2021
[29]

Ruobing Xie, Cheng Ling, Yalong Wang, Rui Wang, Feng Xia, and Leyu Lin. 2021. Deep feedback network for recommendation. InProceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence. 2519–2525

2021
[30]

Ruobing Xie, Lin Ma, Shaoliang Zhang, Feng Xia, and Leyu Lin. 2023. Reweighting Clicks with Dwell Time in Recommendation. InCompanion Proceedings of the ACM Web Conference 2023. 341–345

2023
[31]

An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Me...

2024 arXiv
[32]

Zhao Yang, Junhong Lian, and Xiang Ao. 2023. Fact-Preserved Personalized News Headline Generation. In2023 IEEE International Conference on Data Mining (ICDM). IEEE, 1493–1498

2023
[33]

Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. Bartscore: Evaluating generated text as text generation.Advances in Neural Information Processing Systems34 (2021), 27263–27277

2021
[34]

Kui Zhang, Guangquan Lu, Guixian Zhang, Zhi Lei, and Lijuan Wu. 2022. Per- sonalized headline generation with enhanced user interest perception. InInter- national Conference on Artificial Neural Networks. Springer, 797–809

2022
[35]

Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 5941–5948

2019

Pith tools

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

[1] [1]

Xiang Ao, Ling Luo, Xiting Wang, Zhao Yang, Jiun-Hung Chen, Ying Qiao, Qing He, and Xing Xie. 2023. Put Your Voice on Stage: Personalized Headline Generation for News Articles.ACM Transactions on Knowledge Discovery from Data18, 3 (2023), 1–20

work page 2023

[2] [2]

Xiang Ao, Xiting Wang, Ling Luo, Ying Qiao, Qing He, and Xing Xie. 2021. PENS: A dataset and generic framework for personalized news headline generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 82–92

2021

[3] [3]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Jill Burstein, Christy...

2019

[4] [4]

Yijiang River Dong, Tiancheng Hu, and Nigel Collier. 2024. Can LLM be a Personalized Judge?arXiv preprint arXiv:2406.11657(2024)

arXiv 2024

[5] [5]

Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A fam- ily of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793(2024)

arXiv 2024

[6] [6]

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276(2024)

arXiv 2024

[7] [7]

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720(2024)

arXiv 2024

[8] [8]

Hao Jiang, Chuanzhen Li, and Mingxiao An. 2024. Time Matters: Enhancing Pre- trained News Recommendation Models with Robust User Dwell Time Injection. arXiv preprint arXiv:2405.12486(2024)

work page Pith review arXiv 2024