Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Large Language Models for Token-Efficient and Semantic-Preserving Opinion Summarization

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Stratified sampling of opinion facets lets LLMs summarize huge review corpora with ~97% fewer tokens while keeping viewpoint diversity.

desk verdict Solid journal extension of their ECML work: three formalized stratified samplers cut tokens ~97% and beat random on multi-domain opinion corpora, with public code; evaluation gold standards are LLM-derived and baselines thinner than the abstract claims. read the letter →

arxiv 2607.10825 v1 pith:V6HYVGLN submitted 2026-07-12 cs.CL cs.AIcs.LG

Fabrizio Marozzo , Stefano Iannicelli This is my paper
classification cs.CLcs.AIcs.LG
keywords LargeLanguageModelsOpinionSummarizationStratifiedSamplingTokenEfficiencySemanticPreservationReviewMiningMultidimensionalClassification
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

Huge collections of product reviews, hotel feedback, and social posts are redundant, imbalanced, and expensive to feed whole into a large language model. This paper argues that the right fix is not a bigger context window or retrieval, but a corpus-level selection step: first label every opinion on multiple semantic facets (sentiment, topics, emotion, and domain-specific labels), then use stratified samplers to pick a small, distribution-matched subset before the LLM ever sees the text. On Amazon electronics reviews, Tripadvisor New York hotels, and 2024 election posts from swing-state Twitter users, samples of roughly twenty opinions produce summaries whose topic coverage and embedding similarity to full-corpus summaries far exceed random selection, while cutting input tokens by about 97 percent. The practical payoff is scalable, viewpoint-balanced opinion digests that stay faithful to minority as well as majority voices without paying full-corpus token cost.

What carries the argument

Multidimensional stratified sampling: each opinion is first given probabilistic labels on sentiment, topics, emotion (and optional domain facets); three strategies (relevance-constrained Knapsack, KL-regularized Knapsack-KL, and density-based KDE) then select a fixed-size subset that preserves those class distributions under a hard token budget.

What would settle it

If, on the same three datasets, random samples of equal size matched or beat the stratified methods on both topic-coverage F1 and ModernBERT cosine similarity to a carefully constructed human or multi-model reference summary, the claimed advantage of distribution-aware selection would collapse.

Watch

Extended reading notes

Core claim

A compact subset of opinions selected by multidimensional stratified sampling (Knapsack, Knapsack-KL, or KDE) can be summarized by an LLM so that topic coverage and summary-level cosine similarity stay close to the full-corpus summary, while token usage falls by roughly two orders of magnitude.

Load-bearing premise

The evaluation treats the LLM summary of the entire unfiltered corpus as a reliable semantic gold standard, even though very large contexts can themselves degrade quality.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a corpus-level framework for LLM-based opinion summarization that first annotates opinions with multidimensional probabilistic facets (sentiment, emotion, BERTopic topics, and optional domain facets) and then selects a compact subset under a fixed size N via three stratified samplers—relevance-constrained Knapsack, KL-regularized Knapsack-KL (Eqs. 1–2, Algorithms 1–2), and KDE—before facet-aware prompting. The central claim is that these distribution-aligned subsets (N≈20) preserve topical structure and yield summaries whose ModernBERT cosine similarity to full-corpus GPT-5 summaries substantially exceeds random sampling, while cutting input tokens by roughly 97% on Amazon product reviews, Tripadvisor hotel reviews, and X/Twitter political posts. An ablation on semantic dimensions and a public implementation are provided.

Significance. If the semantic-preservation claim holds under stronger evaluation, the work is practically significant: it offers a clear, implementable alternative to brute-force long-context ingestion or query-specific RAG for large opinion corpora, with explicit multi-facet balance and documented token savings. Strengths include formalized sampling algorithms with pseudocode, three heterogeneous domains, a dimension ablation (Table 2), efficiency trade-off discussion, and a public GitHub release. These elements make the contribution reproducible and useful for applied opinion mining even if some evaluation choices need tightening. The advance over the authors’ prior conference version (broader samplers, multi-domain protocol, token-focused metrics) is real but incremental.

major comments (3)
  1. Section 4.2.2 and Figures 4(a)–4(c): the primary semantic-preservation metric is cosine similarity of sample-based summaries to full-corpus GPT-5 summaries. The paper itself cites evidence that very large contexts can degrade summary quality [29] and reports full sets of 28k–100k+ tokens. Treating the full-context output as the gold standard therefore risks measuring fidelity to a potentially biased or incomplete reference rather than to the true opinion distribution. This is load-bearing for the abstract’s “semantic preservation” claim and should be addressed by (i) human preference or aspect-coverage judgments, (ii) an independent reference (e.g., extractive multi-document baselines or human-written summaries on a subset), or (iii) explicit analysis showing that full-context quality does not degrade on these corpora.
  2. Abstract and Section 4: the abstract states that the method “consistently outperform[s] traditional AI-based and standard LLM summarization baselines,” yet the reported experiments compare mainly against random sampling and full-context LLM input (plus internal sampler variants). No classical multi-document summarizers (e.g., LexRank, MMR, clustering-based selection), no sentiment-only or single-facet filters, and no standard long-context or map-reduce LLM pipelines appear as quantitative baselines. The claim as written is not supported by the experimental section and should either be narrowed or backed by those comparisons.
  3. Section 4.2.1: ground-truth topics for coverage F1 are free-form themes extracted by ChatGPT-Thinking and matched via ModernBERT embeddings. Without human topic labels or inter-annotator agreement on a sample, both the absolute F1 numbers and the ranking of samplers remain LLM-dependent. A modest human-validated topic set (or at least a sensitivity check against an alternative topic model) is needed to underwrite the “topic coverage” half of the main claim.
minor comments (6)
  1. Section 3.2: the relevance score (top-k BERTopic term counts, normalized) is simple; a short sensitivity analysis on k, or comparison to embedding-based relevance, would strengthen the design justification.
  2. Algorithm 2 / Eq. (1): the linear schedule for α(i) is stated but not ablated; reporting results for fixed α or alternative schedules would clarify robustness of Knapsack-KL and KDE.
  3. Figure 2 and the hotel walkthrough are helpful; adding analogous distribution plots for Amazon and X/Twitter (even in appendix) would make multi-domain balance more transparent.
  4. Table 1 averages across domains; per-domain rows (or appendix tables) would show whether KDE’s edge is consistent or driven by one corpus.
  5. Related work: coverage of recent long-context and opinion-summarization LLM papers is thin; a few additional 2024–2025 citations would better situate the contribution.
  6. Minor typos/notation: “KullLei” in figure legends vs “Knapsack-KL” in text; “T opic” spacing artifacts in subsection titles; ensure consistent N vs sample size wording.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-by-construction circularity in sampling or claims; mild evaluative self-reference from using full-context LLM summaries and ChatGPT topics as gold standards.

  1. other [Section 4.2.2 (Content similarity) and 4.2.1 (Topic coverage); also citation to [29]]
    "For each instance, we first generate a reference summary using all available opinions, then create sample-based summaries of increasing size using each sampling strategy. Both summaries are embedded with ModernBERT, and cosine similarity is computed between their embeddings. ... Although modern LLMs support increasingly wider context windows, using very large inputs can still degrade summary quality as the context grows [29]"

    The quantitative backbone of the strongest claim (high semantic fidelity at ~97% token reduction) treats the full-context GPT-5 summary itself as the gold reference that N≈20 stratified samples approach (0.83–0.88 cosine). Topic F1 likewise uses ChatGPT-Thinking free-form topics extracted from the full corpus. Because the paper simultaneously flags large-context degradation, high similarity primarily shows that the sample reproduces the LLM's own full-input output rather than an independent external ground truth; the evaluation loop is therefore partially self-referential. This is not definitional equivalence of method to input (random still scores lower), so it is only mild.

full rationale

The paper's core chain is methodological and empirical, not a mathematical derivation of a forced result: multidimensional probabilistic classification (sentiment/emotion/topics via transformers + BERTopic) produces facet distributions; stratified samplers (Knapsack feasibility on target counts, Knapsack-KL with scheduled KL+relevance score, KDE density matching) then select a fixed-size subset under token budget to approximate those distributions while maximizing relevance; an LLM is prompted on the subset. None of these steps defines the output summary or the reported metrics in terms of themselves. Outperformance vs random on topic F1 and ModernBERT cosine is an empirical observation, not an identity. Self-citation to the authors' conference precursor [2] is explicit extension (new samplers, formalization, multi-domain eval, token-focused protocol) and is not load-bearing uniqueness or an ansatz that forbids alternatives. The only mild circularity is evaluative: full-corpus GPT-5 summaries are treated as the semantic reference for cosine (Sec. 4.2.2) and ChatGPT-Thinking free-form topics as ground truth for coverage (Sec. 4.2.1), even while the paper cites evidence that large contexts degrade quality. This makes 'semantic preservation' partly self-referential to LLM behavior under the very conditions flagged as problematic, but it does not collapse the method or the comparative claims into tautology by construction. Score remains low because the sampling objective and the random baseline comparison stand independently.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The central empirical claim rests on standard ML tooling plus a few design choices (sample size N, relevance definition, α schedule, density estimation) and the modeling assumption that facet-balanced subsets under a token budget preserve summary semantics relative to full-context LLM output. No new physical entities are postulated; the 'invented' pieces are the named sampling procedures.

free parameters (4)
  • sample size N
    Primary operating point for all main tables and figures; N=20 is the headline setting for Table 1 and the ablation, chosen by the authors rather than derived.
  • α(i) scheduling coefficient in Knapsack-KL / KDE score
    Linear schedule from near 1 to 0 in Eq. (1) trades relevance vs KL; functional form and endpoints are design choices that affect selection order.
  • top-k BERTopic terms for relevance_score
    Relevance is defined by counting top-k topic terms and normalizing; k and the term list are free design parameters of the value function.
  • histogram / density estimation settings for KDE
    KDE uses histogram-based density approximations over probability space; binning and update details are not fully fixed by theory and affect the scoring surface.
assumptions (4)
  • domain assumption Transformer classifiers and BERTopic produce probabilistic facet labels that adequately represent the semantic structure needed for balanced selection.
    Section 3.1 treats these scores as the ground structure for all samplers; accuracy of that structure is assumed rather than proven for every domain.
  • domain assumption Preserving empirical class distributions across sentiment/topic/emotion under a fixed token budget improves LLM summary fidelity versus random selection.
    This is the load-bearing methodological premise of §3.2 and the experimental claims in §4.
  • ad hoc to paper Full-corpus LLM summaries are a suitable reference for measuring semantic preservation via embedding cosine similarity.
    Section 4.2.2 defines the main content-similarity metric this way despite the paper noting long-context degradation [29].
  • standard math Standard optimization and information-theoretic tools (multi-constraint knapsack-style selection, KL divergence, kernel/histogram density estimation) are valid selection criteria.
    Used as background machinery in Algorithms 1–2 and Eqs. (1)–(2).
invented entities (3)
  • Knapsack stratified sampler for multi-facet opinion selection
    purpose: Select N opinions by relevance under per-class count capacities derived from corpus proportions.
    Named procedure in Algorithm 1; algorithmic construct, not an external physical entity.
  • Knapsack-KL stratified sampler
    purpose: Among feasible candidates, minimize a scheduled mix of KL drift and inverse relevance (Eq. 1).
    Algorithm 2; paper-specific combination of knapsack feasibility with KL regularization.
  • KDE stratified sampler for opinion subsets
    purpose: Greedily pick the globally best remaining item under the same score using density estimates of class probabilities.
    Described in §3.2 as the most faithful but costliest strategy; construct of the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models for Token-Efficient and Semantic-Preserving Opinion Summarization." pith.science (2026). https://pith.science/paper/V6HYVGLN

@misc{pith2026260710825,
  author       = {Pith},
  title        = {Pith review of: Large Language Models for Token-Efficient and Semantic-Preserving Opinion Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6HYVGLN}},
  note         = {Machine review of arXiv:2607.10825}
}
read the original abstract

Opinionated text - spanning product reviews, hotel feedback, and social posts - captures rich signals about user experiences, preferences, and concerns. However, the scale, redundancy, and imbalance of such corpora make it challenging to analyze opinions effectively, particularly when the goal is to generate summaries that remain faithful to the diversity of viewpoints expressed. This paper presents a framework that preserves semantics in LLM-based opinion summarization while minimizing token usage. We combine multidimensional classification (e.g., sentiment, topics) with a family of stratified sampling strategies to select compact yet representative subsets of opinions before prompting the LLM. Tailored prompts then produce balanced summaries that surface the salient aspects expressed in the opinions (e.g., strengths and weaknesses of products/hotels). Experiments on Amazon product reviews, Tripadvisor hotel reviews, and X/Twitter posts demonstrate that our method significantly reduces token usage and computational cost while consistently outperforming traditional AI-based and standard LLM summarization baselines in terms of content coverage, balance, and semantic preservation.

Figures

Figures reproduced from arXiv: 2607.10825 by the authors.

Figure 1
Figure 1. Execution flow of the proposed framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Class distributions for the full set of reviews of the example hotel and for the balanced [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Topic coverage (F1 score) vs. sample size for different sampling strategies. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Summary-level cosine similarity vs. sample size for different sampling strategies. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Token usage vs. sample size across sampling strategies and random selection. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 5 canonical work pages

  1. [29]

    Lv-eval: A balanced long-context benchmark with 5 length levels up to 256k.arXiv preprint arXiv:2402.05136, 2024

    Tao Yuan, Xuefei Ning, Dong Zhou, Zhijie Yang, Shiyao Li, Minghui Zhuang, Zheyue Tan, Zhuyu Yao, Dahua Lin, Boxun Li, et al. Lv-eval: A balanced long-context benchmark with 5 length levels up to 256k.arXiv preprint arXiv:2402.05136, 2024

    arXiv 2024

  2. [1]

    Enhancing sentiment analysis classification for amazon product reviews using cnn-sigtan-beta activation function.Multimedia Tools and Applications, 83(19):56719–56736, 2024

    P Anbumani and K Selvaraj. Enhancing sentiment analysis classification for amazon product reviews using cnn-sigtan-beta activation function.Multimedia Tools and Applications, 83(19):56719–56736, 2024

    2024

  3. [2]

    Balanced and token-efficient summarization of user reviews via stratified sampling and large language models

    Loris Belcastro and et al. Balanced and token-efficient summarization of user reviews via stratified sampling and large language models. InEuropean Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), page 290–306, 2025. 13

    2025

  4. [3]

    A literature survey of recent advances in chatbots.Information, 13(1), 2022

    Guendalina Caldarini, Sardar Jaf, and Kenneth McGarry. A literature survey of recent advances in chatbots.Information, 13(1), 2022

    2022

  5. [4]

    Harnessing prompt-based large language models for disaster monitoring and automated reporting from social media feedback.Online Social Networks and Media, 45:100295, 2025

    Riccardo Cantini and et al. Harnessing prompt-based large language models for disaster monitoring and automated reporting from social media feedback.Online Social Networks and Media, 45:100295, 2025

    2025

  6. [5]

    Multi-dimensional classification on social media data for detailed reporting with large language models

    Riccardo Cantini, Cristian Cosentino, and Fabrizio Marozzo. Multi-dimensional classification on social media data for detailed reporting with large language models. InInt. Conf. on Artificial Intelligence Applications and Innovations, pages 100–114, 2024

    2024

  7. [6]

    The use of mmr, diversity-based reranking for reordering documents and producing summaries

    Jaime Carbonell and Jade Goldstein. The use of mmr, diversity-based reranking for reordering documents and producing summaries. InProceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, pages 335–336, 1998

    1998

  8. [7]

    Topically diversified summarization of customer reviews

    Florian Carichon and Gilles Caporossi. Topically diversified summarization of customer reviews. In Proceedings of the 6th International Conference on Natural Language and Speech Processing (ICNLSP 2023), pages 178–191, 2023

    2023

Show all 33 references
  1. [8]

    From reviews to results: Generative ai for review-driven product and service comparisons

    Cristian Cosentino, Merve Gunduz-Cure, Fabrizio Marozzo, and Sule Ozturk-Birim. From reviews to results: Generative ai for review-driven product and service comparisons. In28th International Conference on Discovery Science (DS2025), page 78–93, 2025

    2025

  2. [9]

    A topic modeling comparison between lda, nmf, top2vec, and bertopic to demystify twitter posts.Frontiers in sociology, 7, 2022

    Roman Egger and Joanne Yu. A topic modeling comparison between lda, nmf, top2vec, and bertopic to demystify twitter posts.Frontiers in sociology, 7, 2022

    2022

  3. [10]

    Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2023

    Yunfan Gao and et al. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2023. URLhttps://arxiv.org/abs/2312.10997

    2023 arXiv

  4. [11]

    Bertopic: Neural topic modeling with a class-based tf-idf procedure

    Maarten Grootendorst. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv:2203.05794, 2022

    2022 arXiv

  5. [12]

    Mining customer product reviews for product development: A summarization process.Expert Systems with Applications, 132:141–150, October 2019

    Tianjun Hou, Bernard Yannou, Yann Leroy, and Emilie Poirson. Mining customer product reviews for product development: A summarization process.Expert Systems with Applications, 132:141–150, October 2019. ISSN 0957-4174. doi: 10.1016/j.eswa.2019.04.069. URL http://dx.doi.org/10....

    2019 doi

  6. [13]

    Mo, and Hai Liu

    Shaoqin Huang, Yue Wang, Daniel Y. Mo, and Hai Liu. Mining novel customer needs from online product review. InProceedings of the International Conference on Decision Science & Management, ICDSM ’24, page 168–172, New York, NY, USA, 2024. Association for Computing Machinery. IS...

    2024 doi

  7. [14]

    Large-scale and multi-perspective opinion summarization with diverse review subsets

    Han Jiang, Rui Wang, Zhihua Wei, Yu Li, and Xinpeng Wang. Large-scale and multi-perspective opinion summarization with diverse review subsets. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 5641–5656, 2023

    2023

  8. [15]

    Beyond opinion mining: Summarizing opinions of customer reviews

    Reinald Kim Amplayo, Arthur Brazinskas, Yoshi Suhara, Xiaolan Wang, and Bing Liu. Beyond opinion mining: Summarizing opinions of customer reviews. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’22, page...

    2022 doi

  9. [16]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis and et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems (NeurIPS), 2020. URLhttps://arxiv.org/ abs/2005.11401

    2020 arXiv

  10. [17]

    Coverage-based fairness in multi- document summarization.arXiv preprint arXiv:2412.08795, 2025

    Haoyuan Li, Yusen Zhang, Rui Zhang, and Snigdha Chaturvedi. Coverage-based fairness in multi- document summarization.arXiv preprint arXiv:2412.08795, 2025. URLhttps://arxiv.org/abs/ 2412.08795

    2025 arXiv

  11. [18]

    Rouge: A package for automatic evaluation of summaries

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

    2004

  12. [19]

    Opinion observer: analyzing and comparing opinions on the web

    Bing Liu, Minqing Hu, and Junsheng Cheng. Opinion observer: analyzing and comparing opinions on the web. In14th Int. Conf. on World Wide Web, pages 342–351, 2005

    2005

  13. [20]

    A survey of automatic text summarization: concepts, advances and future prospects.International Journal of Speech Technology, 28:801–824, 10 2025

    Chengyao Lv, Yiwen Tang, Lian Ao, Yanxia Huang, Simin Zhang, Junqing Fan, and Wei Han. A survey of automatic text summarization: concepts, advances and future prospects.International Journal of Speech Technology, 28:801–824, 10 2025. doi: 10.1007/s10772-025-10215-y

    2025 doi

  14. [21]

    Chatbots applications in education: A systematic review.Computers and Education: Artificial Intelligence, 2:100033, 2021

    Chinedu Wilfred Okonkwo and Abejide Ade-Ibijola. Chatbots applications in education: A systematic review.Computers and Education: Artificial Intelligence, 2:100033, 2021

    2021

  15. [22]

    Thumbs up? sentiment classification using machine learning techniques.cs/0205070, 2002

    Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. Thumbs up? sentiment classification using machine learning techniques.cs/0205070, 2002

    2002

  16. [23]

    Cognitive hybrid deep learning-based multi-modal sentiment analysis for online product reviews.ACM Trans

    Ashwin Perti, Amit Sinha, and Ankit Vidyarthi. Cognitive hybrid deep learning-based multi-modal sentiment analysis for online product reviews.ACM Trans. Asian Low-Resour. Lang. Inf. Process., 14 23(8), aug 2024. ISSN 2375-4699. doi: 10.1145/3615356. URLhttps://doi.org/10.1145/3615356

    2024 doi

  17. [24]

    Roumeliotis, Nikolaos D

    Konstantinos I. Roumeliotis, Nikolaos D. Tselikas, and Dimitrios K. Nasiopoulos. Llms in e-commerce: A comparative analysis of gpt and llama models in product review evaluation.Natural Language Processing Journal, 6:100056, 2024. ISSN 2949-7191. doi: https://doi.org/10.1016/j....

    2024 doi

  18. [25]

    Llms in e-commerce: a comparative analysis of gpt and llama models in product review evaluation.Natural Language Processing Journal, 6:100056, 2024

    Konstantinos I Roumeliotis, Nikolaos D Tselikas, and Dimitrios K Nasiopoulos. Llms in e-commerce: a comparative analysis of gpt and llama models in product review evaluation.Natural Language Processing Journal, 6:100056, 2024

    2024

  19. [26]

    Automatic text summarization methods: A comprehensive review.SN Computer Science, 4(1):33, 2022

    Grishma Sharma and Deepak Sharma. Automatic text summarization methods: A comprehensive review.SN Computer Science, 4(1):33, 2022

    2022

  20. [27]

    Text classification via large language models.arXiv:2305.08377, 2023

    Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, and Guoyin Wang. Text classification via large language models.arXiv:2305.08377, 2023

    2023 arXiv

  21. [28]

    Exploiting user experience from online customer reviews for product design.Int

    Bai Yang and et al. Exploiting user experience from online customer reviews for product design.Int. J. of Information Management, 46:173–186, 2019

    2019

  22. [30]

    Survey of transformers and towards ensemble learning using transformers for natural language processing.Journal of big Data, 11(1):25, 2024

    Hongzhi Zhang and M Omair Shafiq. Survey of transformers and towards ensemble learning using transformers for natural language processing.Journal of big Data, 11(1):25, 2024

    2024

  23. [31]

    Examining the influence of online reviews on consumers’ decision-making: A heuristic–systematic model.Decision support systems, 67:78–89, 2014

    Kem ZK Zhang, Sesia J Zhao, Christy MK Cheung, and Matthew KO Lee. Examining the influence of online reviews on consumers’ decision-making: A heuristic–systematic model.Decision support systems, 67:78–89, 2014

    2014

  24. [32]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. InICLR, 2020. URLhttps://arxiv.org/abs/1904.09675

    2020 arXiv

  25. [33]

    Clustering sentences with density peaks for multi-document summarization

    Yang Zhang, Yunqing Xia, Yi Liu, and Wenmin Wang. Clustering sentences with density peaks for multi-document summarization. InProceedings of the 2015 conference of the north american chapter of the association for computational linguistics: Human language technologies, pages 1...

    2015

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.