pith. sign in

arxiv: 2510.12534 · v4 · pith:JGFNA3PRnew · submitted 2025-10-14 · 💻 cs.AI

ProtoSiTex: Learning Semi-Interpretable Prototypes for Multi-label Text Classification

Utsav Kumar Nareti , Suraj Kumar , Soumya Pandey , Soumi Chattopadhyay , Chandranath Adak , Sankha Subhra Mullick This is my paper

Pith reviewed 2026-05-21 20:09 UTC · model grok-4.3

classification 💻 cs.AI
keywords prototype learningmulti-label classificationtext classificationinterpretable modelssemi-interpretable AIhierarchical lossfine-grained explanationshotel reviews dataset
0
0 comments X

The pith

ProtoSiTex learns semantically coherent prototypes unsupervised then maps them to multiple labels supervised to classify fine-grained multi-label text with explanations.

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

The paper introduces ProtoSiTex as a framework that alternates between an unsupervised phase to discover diverse and coherent prototypes from text and a supervised phase to align those prototypes to class labels. It adds a hierarchical loss that enforces consistency from subsentence up to document level while using multi-head attention to handle overlapping or conflicting meanings in multi-label settings. A new subsentence-annotated hotel review dataset serves as one testbed alongside existing benchmarks. If the approach holds, it would let models deliver both accurate multi-label predictions and human-readable explanations at a granularity finer than whole sentences or documents. This addresses a gap where prior prototype methods stayed coarse or single-label and could not reliably explain cases with mixed labels.

Core claim

ProtoSiTex is a semi-interpretable model that first runs an unsupervised prototype discovery phase to extract semantically coherent and diverse prototypes, then switches to a supervised classification phase that maps those prototypes to multiple class labels. Adaptive prototypes combined with multi-head attention capture overlapping semantics, while a hierarchical loss function maintains consistency across subsentence, sentence, and document levels. On the introduced hotel-review dataset annotated at the subsentence level with multiple labels, as well as two public benchmarks, the method reaches state-of-the-art accuracy and produces faithful explanations that align with human judgments.

What carries the argument

Dual-phase alternate training that discovers prototypes unsupervised and then maps them to labels supervised, guided by a hierarchical loss across text levels and multi-head attention for overlapping semantics.

If this is right

  • Multi-label predictions become possible at subsentence granularity instead of sentence or document level.
  • Explanations remain faithful to the model's decisions and align with how humans assign overlapping labels.
  • The same architecture handles binary, multi-class, and true multi-label tasks on text.
  • A new publicly usable benchmark dataset exists for training and evaluating subsentence multi-label models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The learned prototypes could be inspected to surface recurring patterns in review text that businesses might use for product improvement.
  • The hierarchical consistency mechanism might transfer to other sequential data such as time-stamped sensor logs that also carry multiple labels.
  • If the unsupervised discovery step can be made more robust, the overall framework could reduce the need for large labeled datasets in new domains.

Load-bearing premise

The unsupervised prototype discovery phase produces prototypes that remain semantically coherent and diverse enough to map reliably onto class labels without major loss of fidelity or spurious alignments.

What would settle it

Human raters judging that the prototype-based explanations on the subsentence-annotated hotel reviews fail to match the reasons annotators gave for assigning multiple overlapping labels at that level.

Figures

Figures reproduced from arXiv: 2510.12534 by Chandranath Adak, Sankha Subhra Mullick, Soumi Chattopadhyay, Soumya Pandey, Suraj Kumar, Utsav Kumar Nareti.

Figure 1
Figure 1. Figure 1: Proposed solution architecture: ProtoSiTex [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of eight aspects in HR [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance analysis of ProtoSiTex on HR: (a) with various prototype initializers; (b) without alternate training (w/o [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Impact of hyperparameters αi and λi on HR for clustering (α1, α2, α3), and classification (λ1, λ2, λ3). To systematically investigate the influence of these hyperparameters on model performance, we conducted a series of experiments on the HR. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Correctly predicted examples of ProtoSiTex on (a) HR, (b) TweetEVAL [ [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Misprediction examples of ProtoSiTex on (a) HR, (b) TweetEVAL [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

The rapid growth of user-generated text across digital platforms has intensified the need for interpretable models capable of fine-grained text classification and explanation. Existing prototype-based models offer intuitive explanations but typically operate at coarse granularity (sentence or document level) and fail to address the multi-label nature of real-world text classification. We propose ProtoSiTex, a semi-interpretable framework designed for fine-grained multi-label text classification. ProtoSiTex employs a dual-phase alternate training strategy: an unsupervised prototype discovery phase that learns semantically coherent and diverse prototypes, and a supervised classification phase that maps these prototypes to class labels. A hierarchical loss function enforces consistency across subsentence, sentence, and document levels, enhancing interpretability and alignment. Unlike prior approaches, ProtoSiTex captures overlapping and conflicting semantics using adaptive prototypes and multi-head attention. We also introduce a benchmark dataset of hotel reviews annotated at the subsentence level with multiple labels. Experiments on this dataset and two public benchmarks (binary and multi-class) show that ProtoSiTex achieves state-of-the-art performance while delivering faithful, human-aligned explanations, establishing it as a robust solution for semi-interpretable multi-label text classification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes ProtoSiTex, a semi-interpretable dual-phase framework for fine-grained multi-label text classification. It alternates between an unsupervised prototype discovery phase to learn semantically coherent and diverse prototypes and a supervised classification phase that maps prototypes to labels using adaptive prototypes, multi-head attention, and a hierarchical loss enforcing consistency across subsentence, sentence, and document levels. The authors introduce a new subsentence-annotated hotel review dataset and report state-of-the-art performance on this dataset plus two public benchmarks, along with faithful, human-aligned explanations.

Significance. If the central performance and fidelity claims hold after verification, the work would address a clear gap in prototype-based models for multi-label text by enabling fine-grained explanations in overlapping-semantics settings. The introduction of a subsentence-level multi-label benchmark dataset is a concrete contribution that could support future research, even if the current empirical support remains limited.

major comments (3)
  1. [Abstract and §3] Abstract and §3 (method description): the central claim that the unsupervised prototype discovery phase yields semantically coherent prototypes that the supervised phase maps to labels without meaningful fidelity loss or spurious alignments is load-bearing for the interpretability and SOTA assertions, yet the manuscript provides no quantitative fidelity metrics (e.g., prototype-label alignment scores before vs. after supervised training) or ablation on the effect of classification gradients on prototype semantics.
  2. [§4] §4 (experiments): the abstract asserts SOTA results and faithful explanations, but the available description contains no quantitative performance tables, ablation studies on the hierarchical loss or multi-head attention, or error analysis; this leaves the performance claims unverified and prevents assessment of whether the adaptive prototypes actually handle overlapping/conflicting semantics better than baselines.
  3. [§3.2] §3.2 (hierarchical loss and training): the dual-phase alternate training is presented as solving the multi-label challenge, but no analysis is given of whether the supervised phase distorts the unsupervised prototype semantics; a concrete test (e.g., measuring prototype diversity or semantic coherence post-training) is needed to address the risk of spurious alignments.
minor comments (2)
  1. [Abstract and §3] The abstract and method sections would benefit from explicit notation for the number of prototypes and the diversity regularization weight, as these appear to be free parameters.
  2. [§4] Clarify how the new hotel-review dataset's subsentence annotations are used in evaluation; the current high-level description leaves the mapping from fine-grained labels to prototype explanations unclear.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We have revised the manuscript to address the concerns regarding quantitative support for our claims on prototype fidelity, performance, and training dynamics. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method description): the central claim that the unsupervised prototype discovery phase yields semantically coherent prototypes that the supervised phase maps to labels without meaningful fidelity loss or spurious alignments is load-bearing for the interpretability and SOTA assertions, yet the manuscript provides no quantitative fidelity metrics (e.g., prototype-label alignment scores before vs. after supervised training) or ablation on the effect of classification gradients on prototype semantics.

    Authors: We agree that explicit quantitative fidelity metrics would strengthen the interpretability claims. In the revised manuscript we have added a new analysis subsection in §4 reporting prototype-label alignment scores (computed via cosine similarity between prototype embeddings and label embeddings) before versus after the supervised phase, together with an ablation that isolates the effect of classification gradients on prototype semantics. These results show only marginal shifts in coherence, supporting the original design rationale. revision: yes

  2. Referee: [§4] §4 (experiments): the abstract asserts SOTA results and faithful explanations, but the available description contains no quantitative performance tables, ablation studies on the hierarchical loss or multi-head attention, or error analysis; this leaves the performance claims unverified and prevents assessment of whether the adaptive prototypes actually handle overlapping/conflicting semantics better than baselines.

    Authors: We acknowledge that the experimental section required more granular reporting. The revised §4 now contains full quantitative performance tables across all three datasets with statistical significance tests, dedicated ablation tables for the hierarchical consistency loss and multi-head attention components, and a new error-analysis subsection that examines cases involving overlapping or conflicting labels. These additions directly verify the SOTA claims and the benefit of adaptive prototypes. revision: yes

  3. Referee: [§3.2] §3.2 (hierarchical loss and training): the dual-phase alternate training is presented as solving the multi-label challenge, but no analysis is given of whether the supervised phase distorts the unsupervised prototype semantics; a concrete test (e.g., measuring prototype diversity or semantic coherence post-training) is needed to address the risk of spurious alignments.

    Authors: This concern is closely related to the first comment. We have extended §3.2 and the new analysis in §4 with concrete post-training measurements of prototype diversity (average pairwise cosine distance among prototypes) and semantic coherence (alignment with unsupervised discovery-phase centroids). The added results demonstrate that the alternate training preserves the original semantic structure while improving label mapping, thereby mitigating the risk of spurious alignments. revision: yes

Circularity Check

0 steps flagged

No circularity detected in ProtoSiTex derivation or claims

full rationale

The paper describes an empirical dual-phase training procedure (unsupervised prototype discovery followed by supervised label mapping) evaluated on held-out test splits of a new subsentence-annotated dataset plus two public benchmarks. No equations, loss terms, or architectural choices are shown to reduce by construction to their own fitted inputs or to a self-citation chain; performance and fidelity claims rest on external experimental results rather than internal redefinition. The architecture employs standard components (adaptive prototypes, multi-head attention, hierarchical loss) whose behavior is measured against independent test data, satisfying the criteria for a self-contained, non-circular derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the assumption that prototypes discovered unsupervised will align with label semantics once supervised training begins; no explicit free parameters or invented entities are named in the abstract, though prototype count, diversity regularization strength, and attention head count are implicit hyperparameters.

free parameters (1)
  • number of prototypes and diversity regularization weight
    Typical in prototype learning; required to control coherence and coverage but not numerically specified in abstract.
axioms (1)
  • domain assumption Semantically coherent prototypes can be discovered in an unsupervised phase and then mapped to multiple overlapping labels without contradiction.
    Invoked when describing the dual-phase strategy and adaptive prototypes for conflicting semantics.

pith-pipeline@v0.9.0 · 5759 in / 1257 out tokens · 44161 ms · 2026-05-21T20:09:01.887801+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    ProtoSiTex employs a dual-phase alternate training strategy: an unsupervised prototype discovery phase that learns semantically coherent and diverse prototypes, and a supervised classification phase that maps these prototypes to class labels. A hierarchical loss function enforces consistency across subsentence, sentence, and document levels

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 4 internal anchors

  1. [1]

    Modeling and prediction of online product review helpfulness: a survey,

    G. O. Diazet al., “Modeling and prediction of online product review helpfulness: a survey,” inACL, 2018, pp. 698–708

    work page 2018

  2. [2]

    Do online reviews matter?—an empirical investigation of panel data,

    W. Duanet al., “Do online reviews matter?—an empirical investigation of panel data,”Decision support systems, vol. 45, pp. 1007–1016, 2008

    work page 2008

  3. [3]

    Whose online reviews to trust? understanding reviewer trustworthiness and its impact on business,

    S. Banerjeeet al., “Whose online reviews to trust? understanding reviewer trustworthiness and its impact on business,”Decision Support Systems, vol. 96, pp. 17–26, 2017

    work page 2017

  4. [4]

    How online product reviews affect retail sales: A meta- analysis,

    K. Floydet al., “How online product reviews affect retail sales: A meta- analysis,”Journal of retailing, vol. 90, no. 2, pp. 217–232, 2014

    work page 2014

  5. [5]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlinet al., “BERT: Pre-training of deep bidirectional transformers for language understanding,” inNAACL-HLT, Jun. 2019, pp. 4171–4186

    work page 2019

  6. [6]

    A survey on sentiment analysis methods, applications, and challenges,

    M. Wankhadeet al., “A survey on sentiment analysis methods, applications, and challenges,”Artificial Intelligence Review, vol. 55, no. 7, pp. 5731–5780, 2022

    work page 2022

  7. [7]

    QA dataset explosion: A taxonomy of nlp resources for question answering and reading comprehension,

    A. Rogerset al., “QA dataset explosion: A taxonomy of nlp resources for question answering and reading comprehension,”ACM CSUR, vol. 55, no. 10, pp. 1–45, 2023

    work page 2023

  8. [8]

    Deep learning–based text classification: a comprehensive review,

    S. Minaeeet al., “Deep learning–based text classification: a comprehensive review,”ACM CSUR, vol. 54, no. 3, pp. 1–40, 2021

    work page 2021

  9. [9]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,”Nature machine intelligence, vol. 1, no. 5, pp. 206–215, 2019

    work page 2019

  10. [10]

    Can we open the black box of ai?

    D. Castelvecchi, “Can we open the black box of ai?”Nature News, vol. 538, no. 7623, p. 20, 2016

    work page 2016

  11. [11]

    "why should i trust you?

    M. T. Ribeiroet al., “"why should i trust you?" explaining the predictions of any classifier,” inACM SIGKDD, 2016, pp. 1135–1144

    work page 2016

  12. [12]

    A unified approach to interpreting model predictions,

    S. M. Lundberget al., “A unified approach to interpreting model predictions,”NIPS, vol. 30, 2017

    work page 2017

  13. [13]

    Evaluating the robustness of interpretability methods through explanation invariance and equivariance,

    J. Crabbéet al., “Evaluating the robustness of interpretability methods through explanation invariance and equivariance,” inNIPS, A. Ohet al., Eds., vol. 36, 2023, pp. 71 393–71 429

    work page 2023

  14. [14]

    On the Robustness of Interpretability Methods

    D. Alvarez-Meliset al., “On the robustness of interpretability methods,” arXiv:1806.08049, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    Molnar,Interpretable machine learning

    C. Molnar,Interpretable machine learning. Lulu. com, 2020

    work page 2020

  16. [16]

    Cognitive representations of semantic categories

    E. Rosch, “Cognitive representations of semantic categories.”Journal of experimental psychology: General, vol. 104, no. 3, p. 192, 1975

    work page 1975

  17. [17]

    This looks like that: deep learning for interpretable image recognition,

    C. Chenet al., “This looks like that: deep learning for interpretable image recognition,”NIPS, vol. 32, 2019

    work page 2019

  18. [18]

    PIP-Net: Patch-based intuitive prototypes for interpretable image classification,

    M. Nautaet al., “PIP-Net: Patch-based intuitive prototypes for interpretable image classification,” inCVPR, 2023, pp. 2744–2753

    work page 2023

  19. [19]

    ProtoryNet - interpretable text classification via prototype trajectories,

    D. Honget al., “ProtoryNet - interpretable text classification via prototype trajectories,”JMLR, vol. 24, no. 1, Jan. 2023

    work page 2023

  20. [20]

    GAProtoNet: A multi-head graph attention-based prototypical network for interpretable text classification,

    X. Wenet al., “GAProtoNet: A multi-head graph attention-based prototypical network for interpretable text classification,” inCOLING, Jan. 2025, pp. 9891–9901

    work page 2025

  21. [21]

    Robust text classification: Analyzing prototype-based networks,

    Z. Souratiet al., “Robust text classification: Analyzing prototype-based networks,” inEMNLP, Nov. 2024, pp. 12 736–12 757

    work page 2024

  22. [22]

    A survey on text classification: From traditional to deep learning,

    Q. Liet al., “A survey on text classification: From traditional to deep learning,”ACM Trans. Intell. Syst. Technol., vol. 13, no. 2, Apr. 2022

    work page 2022

  23. [23]

    Protolens: Advancing prototype learning for fine-grained interpretability in text classification,

    B. Weiet al., “Protolens: Advancing prototype learning for fine-grained interpretability in text classification,” inACL, 2025, pp. 4503–4523

    work page 2025

  24. [24]

    SemEval-2016 task 5: Aspect based sentiment analysis,

    M. Pontikiet al., “SemEval-2016 task 5: Aspect based sentiment analysis,” inSemEval, 2016, pp. 19–30

    work page 2016

  25. [25]

    A survey of text classification algorithms,

    K. Schoutenet al., “A survey of text classification algorithms,”CICLing, pp. 1–14, 2015

    work page 2015

  26. [26]

    Explainable ai for text classification: Lessons from a comprehensive evaluation of post hoc methods,

    M. Cesariniet al., “Explainable ai for text classification: Lessons from a comprehensive evaluation of post hoc methods,”Cognitive Computation, pp. 1–19, 2024

    work page 2024

  27. [27]

    Prototypical networks for few-shot learning,

    J. Snellet al., “Prototypical networks for few-shot learning,” inNIPS, 2017, p. 4080–4090

    work page 2017

  28. [28]

    Interpretable and steerable sequence learning via prototypes,

    Y . Minget al., “Interpretable and steerable sequence learning via prototypes,” inACM SIGKDD, 2019, p. 903–913

    work page 2019

  29. [29]

    ClassVector: A parameterized prototype-based model for text classification

    J. Yaoet al., “ClassVector: A parameterized prototype-based model for text classification.” ACM ICMLC, 2019, p. 322–326

    work page 2019

  30. [30]

    The emerging trends of multi-label learning,

    W. Liuet al., “The emerging trends of multi-label learning,”IEEE TPAMI, vol. 44, no. 11, pp. 7955–7974, 2021

    work page 2021

  31. [31]

    A review on multi-label learning algorithms,

    M.-L. Zhanget al., “A review on multi-label learning algorithms,”IEEE TKDE, vol. 26, no. 8, pp. 1819–1837, 2013

    work page 2013

  32. [32]

    Hierarchical attention networks for document classification,

    Z. Yanget al., “Hierarchical attention networks for document classification,” inNAACL-HLT, 2016, pp. 1480–1489

    work page 2016

  33. [33]

    Deep learning for extreme multi-label text classification,

    J. Liuet al., “Deep learning for extreme multi-label text classification,” inACM SIGIR, 2017, pp. 115–124

    work page 2017

  34. [34]

    Label-specific dual graph neural network for multi-label text classification,

    Q. Maet al., “Label-specific dual graph neural network for multi-label text classification,” inACL, 2021, pp. 3855–3864

    work page 2021

  35. [35]

    Gemini: A Family of Highly Capable Multimodal Models

    G. Teamet al., “Gemini: A Family of Highly Capable Multimodal Models,” vol. arXiv:2312.11805, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  36. [36]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library,

    A. Paszkeet al., “PyTorch: An Imperative Style, High-Performance Deep Learning Library,” inNeurIPS, 2019, pp. 8024–8035

    work page 2019

  37. [37]

    Learning word vectors for sentiment analysis,

    A. L. Maaset al., “Learning word vectors for sentiment analysis,” in ACL, 2011, pp. 142–150

    work page 2011

  38. [38]

    TweetEval: Unified benchmark and comparative evaluation for tweet classification,

    F. Barbieriet al., “TweetEval: Unified benchmark and comparative evaluation for tweet classification,” inEMNLP, 2020, pp. 1644–1650

    work page 2020

  39. [39]

    ARTICLE: annotator reliability through in-context learning,

    S. Duttaet al., “ARTICLE: annotator reliability through in-context learning,” inAAAI, 2025, pp. 14 230–14 237

    work page 2025

  40. [40]

    Estimating the reliability, systematic error and random error of interval data,

    K. Krippendorff, “Estimating the reliability, systematic error and random error of interval data,”Educational and Psychological Measurement, vol. 30, no. 1, pp. 61–70, 1970

    work page 1970

  41. [41]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Y . Liuet al., “RoBERTa: A robustly optimized bert pretraining approach,”arXiv:1907.11692, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  42. [42]

    Graph convolutional networks for text classification,

    L. Yaoet al., “Graph convolutional networks for text classification,” in AAAI, vol. 33, no. 01, 2019, pp. 7370–7377

    work page 2019

  43. [43]

    Bertgcn: Transductive text classification by combining gnn and bert,

    Y . Linet al., “Bertgcn: Transductive text classification by combining gnn and bert,” inACL-IJCNLP, 2021, pp. 1456–1462

    work page 2021

  44. [44]

    ALBERT: A lite bert for self-supervised learning of language representations,

    Z. Lanet al., “ALBERT: A lite bert for self-supervised learning of language representations,” inICLR, 2020

    work page 2020

  45. [45]

    DeBERTav3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing,

    P. Heet al., “DeBERTav3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing,” inICLR, 2023

    work page 2023

  46. [46]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

    V . Sanhet al., “DistilBERT, a distilled version of bert: smaller, faster, cheaper and lighter,”arXiv:1910.01108, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  47. [47]

    ELECTRA: pre-training text encoders as discriminators rather than generators,

    K. Clarket al., “ELECTRA: pre-training text encoders as discriminators rather than generators,” inICLR, 2020

    work page 2020

  48. [48]

    Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference,

    B. Warneret al., “Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference,” inACL, 2025, pp. 2526–2547

    work page 2025

  49. [49]

    XLNet: Generalized autoregressive pretraining for language understanding,

    Z. Yanget al., “XLNet: Generalized autoregressive pretraining for language understanding,”NIPS, vol. 32, 2019

    work page 2019

  50. [50]

    Proto-lm: A prototypical network-based framework for built-in interpretability in large language models,

    S. Xieet al., “Proto-lm: A prototypical network-based framework for built-in interpretability in large language models,” inEMNLP, 2023, pp. 3964–3979

    work page 2023

  51. [51]

    ProtoTEx: Explaining model decisions with prototype tensors,

    A. Daset al., “ProtoTEx: Explaining model decisions with prototype tensors,” inACL, 2022, pp. 2986–2997. SUPPLEMENTARYAPPENDIX Appendices A, B, C, D, E, and F can be found in https://github.com/Utsav30/ProtoSiTex1

    work page 2022