Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Is Domain Adaptation Always Helpful? A Frozen-Backbone Study of Cross-Domain Sentiment Transfer

T0 review · 4 major / 5 minor · reviewed 2026-07-08 · grok-4.5

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

Pith's one-line read Explicit domain adaptation helps frozen sentiment models only when the backbone lacks target-domain coverage.

desk verdict Useful preliminary evidence that frozen-backbone DA helps mainly when the backbone lacks target coverage, with DANN risking damage to FinBERT; the financial half rides on a restricted subset that needs hard scrutiny. read the letter →

arxiv 2607.05937 v1 pith:5Q4PEPLD submitted 2026-07-07 cs.CL cs.LG

Phat Tran , Artin Lahni , Pranav Kulkarni , Yaolun Zhang This is my paper · ORCID
classification cs.CLcs.LG
keywords domainadaptationfrozenbackbonesentimentanalysiscross-domaintransferDANNsupervisedcontrastivelearningFinBERTPLMembeddings
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

This paper asks whether explicit domain adaptation is actually useful once a pre-trained language-model backbone is frozen and only a lightweight adapter is trained for sentiment. The authors freeze several embedding backbones that differ in size and domain specialization (Qwen3-Embedding 0.6B/4B/8B, RoBERTa-base, and FinBERT), train an MLP adapter on consumer reviews with three adaptation objectives—adversarial alignment (DANN), maximum mean discrepancy (MMD), and supervised contrastive learning (SCL)—and measure transfer to movie reviews (SST-2) and a heavily restricted financial-news subset (Financial PhraseBank). On SST-2, adaptation adds essentially nothing at any scale. On the financial subset, the same methods recover substantial accuracy for small general-purpose backbones, while adversarial alignment is associated with degraded performance for the finance-specialized FinBERT, consistent with erosion of pre-existing domain structure that supervised contrastive loss tends to preserve. The claim is that the practical value of explicit domain adaptation is highly contingent on whether the frozen backbone already encodes the target domain.

What carries the argument

A controlled family of frozen PLM embedding backbones paired with a lightweight MLP adapter trained under DANN, MMD, or supervised contrastive loss; the design isolates when adaptation recovers transfer versus when the backbone already supplies domain coverage.

What would settle it

Re-run the identical frozen-backbone protocol on the full Financial PhraseBank or an independent financial sentiment corpus and check whether small general backbones still gain substantially from DANN/MMD/SCL while FinBERT still loses under DANN.

Watch

Extended reading notes

Core claim

The efficacy of explicit domain adaptation for frozen-backbone cross-domain sentiment transfer depends on target-domain coverage already present in the backbone. On SST-2 the methods yield negligible gains regardless of scale; on a restricted Financial PhraseBank subset they recover substantial performance for small general-purpose backbones, while DANN degrades FinBERT and supervised contrastive loss better preserves specialized structure.

Load-bearing premise

The heavily restricted Financial PhraseBank subset is treated as representative enough of financial-domain transfer that the observed recovery for small general backbones and the DANN degradation for FinBERT can be attributed to domain coverage rather than to subset-selection artifacts.

Editorial extensions

If this is right

  • Practitioners can skip domain-adaptation overhead when transferring frozen backbones between nearby review domains such as consumer reviews to SST-2.
  • Small general-purpose frozen embeddings can still serve specialized financial sentiment if adapted with an appropriate objective.
  • Adversarial domain alignment should be avoided for already domain-specialized backbones such as FinBERT.
  • Supervised contrastive loss is a safer adapter objective when the backbone may already hold useful domain structure.
  • Diagnosing backbone domain coverage matters more than the choice of adaptation method alone.

Reading between the lines

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

  • A cheap probe of frozen-embedding nearest-neighbor accuracy on unlabeled target text could decide whether to run adaptation at all before training an adapter.
  • The same contingency—adaptation helps only when the backbone lacks coverage—may appear in other frozen-backbone transfer tasks beyond sentiment, such as topic or intent classification.
  • As general backbones grow and absorb more financial text in pretraining, the recovery regime observed for small models may shrink, leaving specialized models like FinBERT as the main case where adversarial methods are harmful.
  • Systematically varying subset size and label balance on Financial PhraseBank would test how much of the reported recovery is an artifact of the restricted sample.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript presents a preliminary empirical case study of whether explicit domain adaptation (DA) improves cross-domain sentiment transfer when the PLM backbone is frozen. The authors train a lightweight MLP adapter on consumer-review source data using three DA methods (DANN, MMD, SCL) atop a controlled family of frozen embedding backbones (Qwen3-Embedding 0.6B/4B/8B, RoBERTa-base, FinBERT), and evaluate transfer to SST-2 (movie reviews) and a heavily restricted subset of Financial PhraseBank. They report two patterns: on SST-2, DA yields negligible gain at all scales; on the financial subset, DA appears to recover substantial performance for small general-purpose backbones, while adversarial alignment (DANN) degrades the domain-specialized FinBERT backbone and supervised contrastive loss appears to preserve it. The central claim is that DA efficacy is highly contingent on whether the frozen backbone already encodes target-domain coverage.

Significance. If the contingency claim is robust, the work offers practical guidance for a common deployment pattern (frozen backbone + light adapter): skip explicit DA when the backbone already covers the target domain, and prefer structure-preserving objectives (e.g., SCL) over adversarial alignment for domain-specialized models. The controlled multi-scale backbone design and the side-by-side comparison of DANN/MMD/SCL under a frozen-backbone constraint are useful contributions to the DA literature. The manuscript is appropriately hedged as preliminary. The main significance risk is that the positive half of the contingency story rests on a single heavily restricted financial subset whose representativeness is not yet independently established.

major comments (4)
  1. [Abstract; financial evaluation setup / Financial PhraseBank subset] The central contingency claim (DA helps mainly when the frozen backbone lacks target-domain coverage) is load-bearing on the financial evaluation set. The abstract and framing repeatedly flag a 'heavily restricted subset' / 'constrained sample' of Financial PhraseBank, yet the recovery for small general backbones and the DANN degradation for FinBERT are the primary evidence that domain coverage—not subset size, label skew, lexical idiosyncrasy, or selection artifacts—drives the pattern. The manuscript needs a pre-specified subset definition (inclusion/exclusion rules, exact N, class balance, train/eval split protocol) and a control showing that the same backbone-coverage ranking predicts DA gain on a less-restricted financial set (or on multiple independent financial slices). Without that, SST-2 alone only supports the negative half of the claim, and the positive half can collapse into a
  2. [Results on FinBERT / DANN vs SCL comparison] Attribution of DANN degradation on FinBERT to 'erosion of pre-existing domain-specific structure' is a strong mechanistic claim. The manuscript should report diagnostics that separate structure erosion from ordinary optimization effects: e.g., source-domain accuracy before/after DANN, domain-classifier accuracy trajectories, representation similarity (CKA/cosine) of FinBERT embeddings pre- vs post-adapter training under DANN vs SCL/MMD, and a matched hyperparameter sweep showing the degradation is not simply under/over-regularization of the adversarial term. Absent such evidence, the FinBERT–DANN result remains an association, not a demonstrated mechanism.
  3. [Experimental results / tables for SST-2 and financial subset] Statistical reliability of the reported recovery and degradation patterns is not established from the abstract-level description. For each backbone × method × target cell, the paper should report mean ± std over multiple random seeds (adapter init and data order), sample sizes, and a simple significance test or confidence interval for the DA-vs-no-DA delta. This is especially important for the financial subset, which is described as heavily restricted and therefore high-variance. Without error bars and seed-level reporting, the 'substantial recovery' and 'degraded performance' claims cannot be assessed as load-bearing findings.
  4. [Central claim / backbone selection and analysis] The operational definition of 'target-domain coverage' of a frozen backbone is currently implicit (scale + FinBERT specialization). To make the contingency claim falsifiable rather than post-hoc, the paper should define a coverage proxy measured before any DA training—e.g., zero-shot or linear-probe accuracy of the frozen embeddings on the financial target, nearest-neighbor domain purity, or vocabulary/embedding overlap with the financial corpus—and show that this proxy ranks backbones in the same order as observed DA gain. If coverage is only inferred from the DA outcomes themselves, the argument is circular.
minor comments (5)
  1. [Data / experimental setup] Clarify the exact source corpus for 'consumer reviews' (Amazon product reviews? Yelp? multi-domain?) and whether source and target label spaces are identically defined (binary vs three-class Financial PhraseBank mapping).
  2. [Methods / adapter and DA hyperparameters] State MLP adapter architecture (depth, width, dropout), training budget, and the full hyperparameter grid for DANN λ, MMD kernel bandwidth, and SCL temperature so that the frozen-backbone comparison is reproducible.
  3. [Results tables] When reporting 'negligible gain' on SST-2 and 'substantial recovery' on finance, give absolute accuracies (or F1) for the no-DA baseline alongside deltas, so readers can judge ceiling effects vs true non-transfer.
  4. [Abstract and conclusion] The prose is carefully hedged ('preliminary', 'appears to', 'within this constrained sample'), which is appropriate; keep that tone in the title/abstract claims so they do not over-generalize beyond the two targets studied.
  5. [Related work] Add a brief related-work placement against prior frozen-backbone / PEFT domain-adaptation studies and against classic DANN/MMD sentiment-transfer benchmarks so the novelty of the multi-scale Qwen3 + FinBERT design is explicit.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The four major comments correctly identify where the contingency claim needs stronger grounding: a pre-specified financial subset definition and less-restricted control, diagnostics for the FinBERT–DANN association, multi-seed statistics, and an a-priori coverage proxy. We accept all four points as requiring manuscript changes. The revised version will add the requested subset protocol and control experiments, representation and optimization diagnostics, seed-level means/stds with confidence intervals, and a pre-DA coverage proxy that ranks backbones independently of the DA outcomes. We retain the preliminary framing and do not over-claim mechanism where evidence remains associative.

read point-by-point responses

  1. Referee: The central contingency claim is load-bearing on the financial evaluation set. The manuscript needs a pre-specified subset definition (inclusion/exclusion rules, exact N, class balance, train/eval split protocol) and a control showing that the same backbone-coverage ranking predicts DA gain on a less-restricted financial set (or on multiple independent financial slices). Without that, SST-2 alone only supports the negative half of the claim, and the positive half can collapse into a selection artifact.

    Authors: We agree. The positive half of the contingency claim currently rests on a single restricted Financial PhraseBank slice whose construction was not fully pre-specified in the text, which is a genuine limitation. In revision we will (i) document the exact inclusion/exclusion rules, final N, class balance, and train/eval split protocol used for the restricted subset, and (ii) add a control evaluation on a less-restricted Financial PhraseBank partition (and, where feasible, an independent financial slice) so that the backbone-coverage ranking can be checked for consistency beyond the original constrained sample. If the recovery pattern does not hold on the less-restricted set, we will narrow the claim accordingly and treat the restricted-subset result as exploratory. SST-2 already supports the negative half; the added control is needed for the positive half. revision: yes

  2. Referee: Attribution of DANN degradation on FinBERT to 'erosion of pre-existing domain-specific structure' is a strong mechanistic claim. The manuscript should report diagnostics that separate structure erosion from ordinary optimization effects: e.g., source-domain accuracy before/after DANN, domain-classifier accuracy trajectories, representation similarity (CKA/cosine) of FinBERT embeddings pre- vs post-adapter training under DANN vs SCL/MMD, and a matched hyperparameter sweep showing the degradation is not simply under/over-regularization of the adversarial term. Absent such evidence, the FinBERT–DANN result remains an association, not a demonstrated mechanism.

    Authors: The referee is correct that the current wording over-reaches from association to mechanism. We will revise the language to state that DANN is associated with degraded FinBERT transfer, and that this is consistent with (but not proven to be) erosion of domain-specific structure. To strengthen the evidence, the revision will report: (1) source-domain accuracy before vs after adapter training under each objective; (2) domain-classifier accuracy trajectories for DANN; (3) CKA and cosine similarity between frozen FinBERT embeddings and post-adapter representations under DANN vs MMD vs SCL; and (4) a matched sweep over the adversarial weight to check that degradation is not an artifact of a single regularization strength. If the diagnostics do not cleanly separate structure erosion from optimization effects, we will keep the claim associative and remove stronger causal language. revision: yes

  3. Referee: Statistical reliability of the reported recovery and degradation patterns is not established. For each backbone × method × target cell, the paper should report mean ± std over multiple random seeds (adapter init and data order), sample sizes, and a simple significance test or confidence interval for the DA-vs-no-DA delta. This is especially important for the financial subset, which is described as heavily restricted and therefore high-variance.

    Authors: We agree that single-run point estimates are insufficient, particularly on a restricted financial subset. The revised manuscript will report, for every backbone × method × target cell, mean ± standard deviation over multiple random seeds that vary both adapter initialization and data order, together with the evaluation sample size and a confidence interval (or paired test) on the DA-versus-no-DA delta. We will flag any cells where the interval includes zero and temper language about 'substantial recovery' or 'degraded performance' accordingly. This change is necessary for the patterns to be treated as load-bearing findings. revision: yes

  4. Referee: The operational definition of 'target-domain coverage' of a frozen backbone is currently implicit (scale + FinBERT specialization). To make the contingency claim falsifiable rather than post-hoc, the paper should define a coverage proxy measured before any DA training—e.g., zero-shot or linear-probe accuracy of the frozen embeddings on the financial target, nearest-neighbor domain purity, or vocabulary/embedding overlap with the financial corpus—and show that this proxy ranks backbones in the same order as observed DA gain. If coverage is only inferred from the DA outcomes themselves, the argument is circular.

    Authors: This is a fair and important critique: without a pre-DA coverage measure the contingency argument is circular. In revision we will define an explicit coverage proxy computed solely from the frozen embeddings before any adapter or DA training—specifically linear-probe (and, where applicable, zero-shot) accuracy on the financial target, optionally supplemented by nearest-neighbor domain purity or embedding/vocabulary overlap with the financial corpus. We will report the proxy ranking of the backbone family and test whether it predicts the observed DA-gain ordering. If the proxy fails to rank in the same order as DA gain, we will revise or withdraw the contingency framing rather than retrofit the definition after the fact. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical frozen-backbone train-and-evaluate study; contingency claims are observational, not forced by construction or self-citation.

full rationale

This paper is a controlled empirical case study of frozen PLM backbones (Qwen3-Embedding scales, RoBERTa-base, FinBERT) with lightweight MLP adapters trained under DANN, MMD, and SCL, then evaluated on SST-2 and a restricted Financial PhraseBank subset. The central claim—that explicit domain adaptation efficacy is contingent on whether the frozen backbone already encodes target-domain coverage—is presented as an observational pattern from the experiments (“we observe two distinct transfer patterns… These preliminary findings suggest…”), not as a first-principles derivation, uniqueness theorem, or fitted identity renamed as prediction. There are no equations that define a quantity in terms of the quantity being “predicted,” no load-bearing uniqueness result imported from the authors’ prior work, no ansatz smuggled via self-citation, and no renaming of a known empirical regularity as a novel derivation. Subset-selection and representativeness concerns (flagged by the authors themselves as “heavily restricted” / “constrained sample”) are validity/correctness risks for the attribution story, not circularity of the derivation chain. The study is self-contained against its reported train/eval protocol; score 0 is the appropriate honest finding.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new particles, forces, or theoretical entities. The ledger is dominated by standard ML design choices: which subset of Financial PhraseBank is used, which hyperparameters train the adapter, and the domain assumptions that freezing the backbone and using these three DA methods fairly answers whether domain adaptation is helpful.

free parameters (2)
  • Financial PhraseBank subset selection
    The financial target is a 'heavily restricted subset'; the selection rule is a free design choice that can materially change the reported recovery and degradation patterns.
  • MLP adapter and DA hyperparameters
    Learning rates, adapter width, DANN adversarial weight, MMD kernel bandwidth, SCL temperature, and training epochs are free training choices not fixed by theory; values not given in the abstract.
assumptions (3)
  • domain assumption Frozen PLM embeddings plus a lightweight MLP adapter are a sufficient and fair testbed for measuring the value of explicit domain adaptation in sentiment transfer.
    The entire experimental family freezes the backbone; if fine-tuning the backbone changes the contingency, the claim does not transfer.
  • domain assumption Consumer reviews as the sole source domain and SST-2 plus restricted Financial PhraseBank as targets adequately probe cross-domain sentiment transfer.
    Findings are conditioned on this source-target pair choice; other domain pairs could reverse the patterns.
  • domain assumption DANN, MMD, and SCL as implemented are representative of explicit domain adaptation methods for this setting.
    Conclusions about when DA helps inherit the strengths and failure modes of these three methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Domain Adaptation Always Helpful? A Frozen-Backbone Study of Cross-Domain Sentiment Transfer." pith.science (2026). https://pith.science/paper/5Q4PEPLD

@misc{pith2026260705937,
  author       = {Pith},
  title        = {Pith review of: Is Domain Adaptation Always Helpful? A Frozen-Backbone Study of Cross-Domain Sentiment Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5Q4PEPLD}},
  note         = {Machine review of arXiv:2607.05937}
}
read the original abstract

Sentiment analysis with frozen pre-trained language model (PLM) backbones has become a common paradigm, yet the practical benefit of explicit domain adaptation remains unclear, particularly when backbones encode varying degrees of target-domain knowledge. We present a preliminary case study evaluating a controlled family of frozen embedding backbones (Qwen3-Embedding 0.6B, 4B, 8B), alongside RoBERTa-base and FinBERT. We train a lightweight MLP adapter on consumer reviews using Domain-Adversarial Neural Networks (DANN), Maximum Mean Discrepancy (MMD), and Supervised Contrastive Learning (SCL), and evaluate transfer to movie reviews (SST-2) and a heavily restricted subset of financial news (Financial PhraseBank). Within this constrained sample, we observe two distinct transfer patterns. On SST-2, domain adaptation provides negligible gain regardless of scale. On the financial subset, explicit domain adaptation appears to recover substantial performance for small general-purpose backbones. Notably, we find that adversarial alignment (DANN) is associated with degraded performance for domain-specialized backbones like FinBERT, consistent with erosion of pre-existing domain-specific structure, whereas supervised contrastive loss appears to preserve it. These preliminary findings suggest that the efficacy of explicit domain adaptation is highly contingent on whether the frozen backbone already possesses target-domain coverage.

Figures

Figures reproduced from arXiv: 2607.05937 by the authors.

Figure 1
Figure 1. Overview of the frozen-backbone domain adaptation pipeline. Source and target texts are encoded by a [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. t-SNE of adapter features (Qwen3-0.6B), col [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Domain transfer heatmap for Qwen3-0.6B, with rows as source domains and columns as eval￾uation domains. Off-diagonal entries show zero-shot performance. C. Ablation Study To isolate each component’s contribution, we train the adapter once for each of six DA configurations per backbone, targeting only the financial domain from scratch. Table III reports macro F1 on Financial PhraseBank. Because this evaluation set co… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Confusion matrices for Qwen3-0.6B on zero [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    FinBERT: Financial Sentiment Analysis with Pre-trained Language Models

    Dogu Araci. FinBERT: Financial Sentiment Analy- sis with Pre-trained Language Models.arXiv preprint arXiv:1908.10063, 2019

    work page Pith review arXiv 1908

  2. [2]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle- moyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach.arXiv preprint arXiv:1907.11692, 2019

    work page Pith review arXiv 1907

  3. [3]

    How Contaminated Is Your Benchmark? Measuring Dataset Leakage in Large Language Models with Kernel Divergence

    Hyeong Kyu Choi, Maxim Khanov, Hongxin Wei, and Yixuan Li. How Contaminated Is Your Benchmark? Measuring Dataset Leakage in Large Language Models with Kernel Divergence. InForty-second International Conference on Machine Learning, 2025

    work page 2025

  4. [4]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive Deep Models for Semantic Composi- tionality Over a Sentiment Treebank. InProceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–1642, 2013

    work page 2013

  5. [5]

    Good Debt or Bad Debt: Detecting Semantic Orientations in Economic Texts.Journal of the Association for Information Science and Technology, 65(4):782–796, 2014

    Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Walle- nius, and Pyry Takala. Good Debt or Bad Debt: Detecting Semantic Orientations in Economic Texts.Journal of the Association for Information Science and Technology, 65(4):782–796, 2014. 8

    work page 2014

  6. [6]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Day- iheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176, 2025

    work page Pith review arXiv 2025

  7. [7]

    Domain-Adversarial Training of Neural Networks.J

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pas- cal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-Adversarial Training of Neural Networks.J. Mach. Learn. Res., 17(1), 2016

    work page 2016

  8. [8]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. A Kernel Two-Sample Test.Journal of Machine Learning Research, 13(25):723–773, 2012

    work page 2012

Show all 23 references
  1. [9]

    Supervised Contrastive Learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. InAdvances in Neural Information Processing Systems, volume 33, pages 18661–18673, 2020

    2020

  2. [10]

    Character- level Convolutional Networks for Text Classification

    Xiang Zhang, Junbo Zhao, and Yann LeCun. Character- level Convolutional Networks for Text Classification. In Advances in Neural Information Processing Systems, vol- ume 28, 2015

    2015

  3. [11]

    Hidden Factors and Hidden Topics: Understanding Rating Dimensions with ReviewText

    Julian McAuley and Jure Leskovec. Hidden Factors and Hidden Topics: Understanding Rating Dimensions with ReviewText. InProceedingsofthe7thACMConferenceon Recommender Systems, RecSys ’13, pages 165–172, 2013

    2013

  4. [12]

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A Multi- Task Benchmark and Analysis Platform for Natural Lan- guage Understanding. InProceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neu- ral Networks...

    2018

  5. [13]

    BERT Re- discovers the Classical NLP Pipeline

    Ian Tenney, Dipanjan Das, and Ellie Pavlick. BERT Re- discovers the Classical NLP Pipeline. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593–4601, 2019

    2019

  6. [14]

    Benchmark Probing: Inves- tigating Data Leakage in Large Language Models

    Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Ger- stein, and Arman Cohan. Benchmark Probing: Inves- tigating Data Leakage in Large Language Models. In NeurIPS 2023 Workshop on Backdoors in Deep Learning - The Good, the Bad, and the Ugly, 2024

    2023

  7. [15]

    Benchmarkdatacontaminationoflargelanguage models: A survey, 2024

    Cheng Xu, Shuhao Guan, Derek Greene, and M-Tahar Kechadi. Benchmarkdatacontaminationoflargelanguage models: A survey, 2024

    2024

  8. [16]

    The Emperor’s New Clothes in Benchmarking? A Rigorous Examination of Mitigation Strategies for LLM Benchmark Data Contamination

    Yifan Sun, Han Wang, Dongbai Li, Gang Wang, and Huan Zhang. The Emperor’s New Clothes in Benchmarking? A Rigorous Examination of Mitigation Strategies for LLM Benchmark Data Contamination. InForty-second Inter- national Conference on Machine Learning, 2025

    2025

  9. [17]

    Neural Unsupervised Domain Adaptation in NLP—A Survey

    Alan Ramponi and Barbara Plank. Neural Unsupervised Domain Adaptation in NLP—A Survey. InProceedings of the 28th International Conference on Computational Linguistics, pages 6838–6855, 2020

    2020

  10. [18]

    Rethink Maximum Mean Discrepancy for Domain Adap- tation, 2020

    Wei Wang, Haojie Li, Zhengming Ding, and Zhihui Wang. Rethink Maximum Mean Discrepancy for Domain Adap- tation, 2020

    2020

  11. [19]

    Freeze the Backbones: a Parameter-Efficient ContrastiveApproachtoRobustMedicalVision-Language Pre-Training

    JiumingQin,CheLiu,SiboCheng,YikeGuo,andRossella Arcucci. Freeze the Backbones: a Parameter-Efficient ContrastiveApproachtoRobustMedicalVision-Language Pre-Training. InICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 16...

    2024

  12. [20]

    No Train, all Gain: Self-Supervised Gradients Improve Deep Frozen Representations

    Walter Simoncini, Andrei Bursuc, Spyros Gidaris, and Yuki M Asano. No Train, all Gain: Self-Supervised Gradients Improve Deep Frozen Representations. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  13. [21]

    Domain Adversarial Training for Aspect-Based Sentiment Analysis

    Joris Knoester, Flavius Frasincar, and Maria Mihaela Truşcˇ a. Domain Adversarial Training for Aspect-Based Sentiment Analysis. InWeb Information Systems Engi- neering – WISE 2022, pages 21–37, 2022

    2022

  14. [22]

    Pseudo-Label Guided Unsupervised Domain Adaptation of Contextual Embeddings

    Tianyu Chen, Shaohan Huang, Furu Wei, and Jianxin Li. Pseudo-Label Guided Unsupervised Domain Adaptation of Contextual Embeddings. InProceedings of the Second Workshop on Domain Adaptation for NLP, pages 9–15, 2021

    2021

  15. [23]

    Decoupledweightdecay regularization, 2019

    IlyaLoshchilovandFrankHutter. Decoupledweightdecay regularization, 2019. 9

    2019

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.