Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Mechanistic Decomposition of Sentence Representations

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that sentence embeddings can be decomposed into sparse, interpretable atoms via supervised dictionary learning on token representations, and that mean pooling preserves the atoms best aligned with the sentence direction.

desk verdict A useful proof-of-concept for decomposing sentence embeddings, but the 'mechanistic' label oversells a descriptive, label-aligned dictionary method. read the letter →

arxiv 2506.04373 v2 pith:ZRHW2P7N submitted 2025-06-04 cs.CL cs.AI

Matthieu Tehenan , Vikram Natarajan , Jonathan Michala , Milton Lin , Juri Opitz This is my paper · ORCID
classification cs.CLcs.AI
keywords sentenceembeddingsdictionarylearningmechanisticinterpretabilitymeanpoolingsparsecodingprobingpart-of-speechtagginglinearencoding
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

Sentence embeddings are central to modern language AI, but their content is normally opaque: a sentence becomes a dense vector that can be compared but not read. This paper argues that the content can be recovered by decomposing each token's embedding into a sparse combination of learned dictionary atoms, guided by linguistic labels such as part of speech and dependency relations. Because the final sentence vector is obtained by mean pooling the tokens, the pooled representation inherits a weighted average of those atoms, and the authors show that each atom's contribution can be scored by how often it is used and how well it aligns with the sentence direction. The central finding is that many semantic and syntactic properties are linearly encoded and survive pooling, which makes sentence embeddings inspectable and opens a path to controlling what they represent.

What carries the argument

The central object is the sparse dictionary model $X \approx DZ$, in which each token embedding $x_t$ is written as a linear combination of $k$ dictionary atoms with a code $z_t$ satisfying $\|z_t\|_0 \le 5$. The dictionary is trained with a supervised objective combining reconstruction error, cross-entropy losses for part-of-speech and dependency prediction, reconstruction of the static word embedding, and a sparsity penalty. The pooling analysis then rests on the identity $s = D\bar{z}$, where $\bar{z}$ is the mean of the token codes; this makes the sentence vector a weighted combination of the same atoms. Each atom's contribution is scored as $a_k = \bar{z}_k \langle d_k, s\rangle$ — the product of how frequently the atom is used and how strongly it aligns with the overall sentence direction — which lets the authors rank atoms and linguistic classes by their impact on the final embedding.

What would settle it

Take a sentence, identify its dominant atom by the attribution score, then intervene by setting that atom's activation to zero (or to a large value) in the decomposed code and feed the reconstructed embedding into a downstream similarity or classification task. If the decomposition is mechanistic, the change in output should be predictable from the atom's labeled feature; if the output barely moves or changes erratically, the atoms are only a low-rank approximation with no causal force.

Watch

Extended reading notes

Core claim

The paper's central claim is that a sentence embedding is not an opaque mixture but a weighted sum of a small number of interpretable directions. Concretely, the authors model token representations as $x_t = D z_t$, where $D$ is a learned dictionary of atoms and each sparse code $z_t$ has at most five nonzero entries, and they learn $D$ together with a classifier so that atoms line up with syntactic roles. Applying this to three sentence encoders on a Brown-corpus sample, they find that linear probes nearly match nonlinear probes for part-of-speech prediction (up to 0.89 on MiniLM), that individual atoms specialize in features such as numerals or adjectival modifiers, and that after mean pooling the dominant contributors are semantically central categories like nouns, verbs, and roots. The authors take this as evidence that much semantic and syntactic information is linearly encoded in the embeddings, with the pooled vector inheriting a weighted average of interpretable atoms.

Load-bearing premise

The entire pipeline assumes that token embeddings are well approximated as sparse linear combinations of a fixed dictionary, with at most five nonzero coefficients per token; if that sparse-linear model is wrong, the learned atoms and the pooling attributions do not describe the encoder's actual computation.

Editorial extensions

If this is right

  • A sentence embedding can be read as a small set of weighted interpretable atoms, giving a natural way to explain similarity, retrieval, or classification decisions in terms of specific linguistic features.
  • Mean pooling is transparent rather than lossy in this view: it linearly aggregates sparse codes, so the pooled vector's content is expressible as an average of atom activations.
  • The diagnostic transfers to any mean-pooled sentence encoder, offering a scalable route to feature attribution without retraining the encoder.
  • The dominance of nouns, verbs, and syntactic roots in the attribution scores gives empirical weight to linguistic theories that place core semantic roles near the root of a sentence.
  • Rare categories such as interjections appear to occupy dedicated singular vectors in the probe weights, suggesting low-frequency features have their own linear subspaces even when they contribute little to the pooled representation.

Reading between the lines

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

  • If the sparse-linear hypothesis holds broadly, interpretability tools learned on one encoder or language could transfer to others, since the underlying dictionaries may share structural similarities.
  • The contribution score $a_k = \bar{z}_k \langle d_k, s\rangle$ implies a testable prediction: sentences whose dominant atoms belong to a given semantic class should cluster along those atoms' directions, which contrastive similarity experiments could verify.
  • Comparing the same contribution score under max or attention pooling would quantify which pooling strategy best preserves rare or spike-activating atoms, a natural extension the paper leaves open.
  • If rare tags such as interjections really live in their own subspaces, targeted manipulation of those directions could enable few-shot adaptation or bias control without retraining.
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

4 major / 5 minor

Summary. The manuscript proposes a supervised dictionary-learning framework for token-level representations of sentence encoders and uses mean pooling of the learned sparse codes to attribute content to sentence embeddings. Experiments are run on the Brown corpus with three sentence-transformers (multilingual-e5-large, all-mpnet-base-v2, all-MiniLM-L6-v2). The paper reports that linear probes recover part-of-speech and dependency information, that supervised dictionary learning yields POS/DEP-aligned atoms, and that a contribution score ranks atoms and linguistic classes by their role in the pooled representation. The central claims are that many semantic and syntactic aspects are linearly encoded and that the method provides a mechanistic decomposition of sentence embeddings.

Significance. If the central claims held, the paper would offer a useful bridge between token-level dictionary-learning interpretability and sentence-level embedding analysis, with potential applications in transparency and controllable representation. Strengths of the manuscript include the use of three publicly available models, a reproducible hyperparameter sweep reported in full in the appendix, and the explicit probing baselines in Section 4.1. However, the evidence currently supports a descriptive, label-aligned decomposition rather than a mechanistic account: the dictionary is trained with the same POS/DEP labels later used to interpret it, no causal interventions are performed, and the quantitative results are reported without variance and on a single corpus. These gaps directly affect the strength of the abstract's 'mechanistically decompose' claim.

major comments (4)
  1. [§4.2, Figures 5/6/14] The supervised dictionary-learning objective in §3.1 includes the term λ L_sup(f_θ(z_i), y_i), and §4.2 states that the objective uses cross-entropy losses for POS and dependency prediction. The interpretability evidence in Figures 5, 6, and 14 then evaluates atom alignment with these very POS and DEP labels. This means that the apparent interpretability is at least partly a consequence of the supervision signal, rather than an independent discovery about the encoder's internal structure. An unsupervised dictionary-learning baseline, or an evaluation on held-out labels not used during training, is needed to support the claim that the atoms capture genuine latent features of the representations.
  2. [§3.2, Eq. (3)] The contribution score is defined as a_k = \bar{z}_k ⟨d_k, s⟩. Since s = D\bar{z} = Σ_j \bar{z}_j d_j, and Figure 12 shows that the dictionary atoms are nearly orthogonal, ⟨d_k, s⟩ ≈ \bar{z}_k ||d_k||^2, so a_k is approximately proportional to \bar{z}_k^2. Under the reported near-orthogonality, the 'directional alignment' factor mostly restates the atom's usage, and the rankings in Figures 8–10 become largely rankings by squared mean activation. The paper should use an alignment measure that excludes the atom's own contribution to s, or explicitly qualify that a_k is a usage-weighted self-alignment score rather than an independent directional measure.
  3. [§3.1 and §4.2] The formal objective in §3.1 is min_{D,Z,θ} Σ_i ||x_i − Dz_i||² + λ L_sup(f_θ(z_i), y_i) subject to ||z_i||₀ ≤ ε, with no static-embedding reconstruction term and no explicit sparsity penalty. Section 4.2, however, states that the objective combines four terms: contextual reconstruction, POS/DEP cross-entropy, static-embedding reconstruction, and a sparsity penalty. Appendix C further lists hyperparameters such as alpha-pos, alpha-dep, alpha-static, and alpha-sparse (Figures 14b/14c) that do not appear in the formal equation. As written, these inconsistencies prevent the method from being reproduced from the equations alone, and the relationship among λ, the alpha weights, and the l1 coefficients should be specified precisely.
  4. [§5, Tables 1–3] The central claim that the method 'mechanistically decompose[s]' sentence embeddings is not supported by the present experiments. The dictionary is fit to token activations, and the contribution metric is descriptive; no intervention shows that removing, scaling, or otherwise manipulating an atom's contribution changes the encoder's output in a predicted direction, and no comparison against a dense low-rank baseline (e.g., PCA) establishes that the sparse-linear model describes the encoder's actual computation. In addition, the paper selects the best hyperparameter run from Tables 1–3 and reports all interpretability results for that single configuration, without variance across seeds or a held-out corpus. At minimum, the title, abstract, and Section 6 should be reframed to 'descriptive decomposition,' or causal tests and multi-run evaluation should be added.
minor comments (5)
  1. [§5] The text contains a typo: 'Meal pooling compresses a sequence of token codes' should read 'Mean pooling compresses a sequence of token codes.'
  2. [Limitations] The Limitations section begins with 'This work is a bluerprint'; 'bluerprint' should be 'blueprint.'
  3. [§2.2] The word 'interpreatbility' appears in 'input-saliency interpreatbility methods'; it should be 'interpretability.'
  4. [Figure 6 and §5.1] There is an inconsistency in the description of atom 58: Figure 6's caption says 'atom 58 detects adjectival modifiers,' while §5.1 says 'atom 58, which activates for adverbs describing actions.' Please clarify which category is intended.
  5. [Figure 14 caption] The caption says 'POS-aligned heatmaps for four sentence embedding models,' but the figure contains only three panels: all-MiniLM-L6-v2, all-mpnet-base-v2, and multilingual-e5-base. Either a fourth panel is missing or the caption should be corrected.

Circularity Check

2 steps flagged · score 6.0 of 10

Supervised dictionary learning trains atoms on POS/DEP labels, then reports label-aligned atoms as mechanistic evidence; the pooling-contribution analysis re-expresses those same fitted codes, so the interpretability claim is partially circular.

  1. self definitional [Section 3.1 (Interpretability), dictionary-learning objective and following paragraph]
    "To steer the learned atoms toward interpretable features, we introduce supervision via an auxiliary predictor fθ (e.g. for syntactic features we have identified with probes). ... This joint objective ensures that the resulting atoms are both reconstructive and predictive of the target labels, hence interpretable."

    Interpretability is defined here as being 'predictive of the target labels,' and the joint objective contains exactly the cross-entropy loss L_sup(fθ(z_i), y_i) over POS and DEP labels. The paper later presents atom–POS/DEP alignment (Figures 5, 6, 14) as a discovery about the encoder, but that alignment is inserted by the training objective itself: the dictionary and codes are optimized to make the labels recoverable. Observing label-aligned atoms after such training is therefore the optimization criterion restated, not independent evidence that the decomposition reflects a causal mechanism of the model.

  2. fitted input called prediction [Section 5.2 (Interpreting Pooled Features), Setup and Results]
    "To account for the fact that atoms may respond to multiple syntactic roles, we assign each atom a set of fractional weights π_i,c, reflecting how often it is activated by tokens of class c (POS or DEP). We then compute the total contribution of each class by summing the corresponding fractions of atom-level contributions. We find that the most influential parts of speech are nouns, verbs, and proper nouns..."

    The per-class contribution ranking is computed from the same atom activations and the same POS/DEP labels that were used to supervise the dictionary in Section 4.2. Because the contribution score is defined as a_k = ar z_k⟨d_k,s⟩ with s = Dar z, it is an algebraic rearrangement of the fitted sparse codes and dictionary, not an independent behavioral or causal measurement. Aggregating these scores by POS/DEP activation frequencies thus re-expresses the supervised fit plus token-frequency statistics; the resulting dominance of nouns and verbs is partly forced by the choice of supervision and the metric definition rather than discovered from the encoder's actual computation.

full rationale

The paper has a genuinely independent component: the linear-probe results in Section 4.1 are empirical and do not rely on circular reasoning, and the self-citations to Opitz et al. are not load-bearing for the main derivation. However, the central 'mechanistic decomposition' claim rests on the supervised dictionary-learning pipeline. There, interpretability is defined as label-predictiveness, the labels are put into the training loss, and the same labels are later used to demonstrate atom interpretability and to rank POS/DEP contributions. Those demonstrations verify the training objective rather than providing independent evidence about the encoder's mechanisms. This is partial circularity in a central claim, so the score is 6 rather than a lower score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a sparse linear model of token embeddings, on the correctness of automatic syntactic annotations, and on the chosen attribution measure. The dictionary size, sparsity budget, and loss weights are tuned rather than derived, and the interpretability evaluation is partially self-confirming because the same labels shape and validate the atoms.

free parameters (4)
  • Dictionary size k = 64 to 512 (best MiniLM: 64)
    Number of atoms; swept via Optuna and directly determines how fine-grained the decomposition is.
  • Sparsity budget epsilon (l0 bound) = 5 nonzero coefficients
    Set a priori; bounds the number of active atoms per token and shapes the decomposition.
  • Loss weights alpha_pos, alpha_dep, alpha_static, alpha_sparse = e.g., MPNet best trial: 0.827365, 0.718469, 0.587198, 0.939217
    Balance reconstruction, POS prediction, DEP prediction, static embedding reconstruction, and sparsity; tuned by Optuna per model.
  • Learning rate = e.g., 0.000729 (MPNet best)
    Adam learning rate tuned per trial; affects convergence and final atoms.
assumptions (5)
  • domain assumption Token embeddings are well approximated by sparse linear combinations of a fixed dictionary (X ≈ DZ, ||z_i||_0 ≤ 5).
    Stated in §4.2 as the dictionary-learning model; the entire decomposition and pooling analysis depends on this approximation being meaningful.
  • domain assumption Stanza's POS and dependency annotations on the Brown corpus are correct and sufficient supervision for interpretable atoms.
    These labels y_i are used as the supervised target in both probing (§4.1) and dictionary learning (§4.2).
  • domain assumption Mean pooling is the relevant pooling operation and equals a simple arithmetic mean of token vectors (s = 1/n Σ x_t).
    The paper explicitly restricts to mean pooling and uses this equation in §3.2; other pooling methods are not analyzed.
  • domain assumption High linear probe accuracy on POS/DEP implies the property is linearly encoded in the embeddings.
    This interpretation of probing results is standard in the literature but still an assumption about what probe accuracy measures.
  • ad hoc to paper The contribution score a_k = \bar z_k ⟨d_k, s⟩ meaningfully quantifies an atom's role in the pooled representation.
    Introduced in §3.2 as the attribution definition; it is a chosen measure, not derived from an independent criterion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mechanistic Decomposition of Sentence Representations." pith.science (2026). https://pith.science/paper/ZRHW2P7N

@misc{pith2026250604373,
  author       = {Pith},
  title        = {Pith review of: Mechanistic Decomposition of Sentence Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZRHW2P7N}},
  note         = {Machine review of arXiv:2506.04373}
}
read the original abstract

Sentence embeddings are central to modern NLP and AI systems, yet little is known about their internal structure. While we can compare these embeddings using measures such as cosine similarity, the contributing features are not human-interpretable, and the content of an embedding seems untraceable, as it is masked by complex neural transformations and a final pooling operation that combines individual token embeddings. To alleviate this issue, we propose a new method to mechanistically decompose sentence embeddings into interpretable components, by using dictionary learning on token-level representations. We analyze how pooling compresses these features into sentence representations, and assess the latent features that reside in a sentence embedding. This bridges token-level mechanistic interpretability with sentence-level analysis, making for more transparent and controllable representations. In our studies, we obtain several interesting insights into the inner workings of sentence embedding spaces, for instance, that many semantic and syntactic aspects are linearly encoded in the embeddings.

Figures

Figures reproduced from arXiv: 2506.04373 by the authors.

Figure 1
Figure 1. Overview of our method. (1) A sentence is tok￾enized into individual token embeddings xt. (2) Each token embedding is decomposed into a set of interpretable latent features using dictionary learning Dzk. (3) The decomposed features are aggregated via mean pooling of the sparse codes. This yields an interpretable decomposition of the pooled sen￾tence vector s. 2020), the issue is particularly acute in the case of sen… view at source ↗
Figure 2
Figure 2. Dependency probing accuracy of linear, nonlinear, shuffled, and random probes across three embedding models (MPNet, E5, MiniLM). Shuffled and random baselines remain near chance [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Part-of-speech probing accuracy of linear, non￾linear, shuffled, and random probes across our three embed￾ding models (multilingual-e5-large, all-mpnet-base-v2, all￾MiniLM-L6-v2). Nonlinear probes outperforms by a slight amount all other settings, with shuffled and random probes showing minimal signal. the values for SVD similarities are shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Heatmap of POS-specific activation patterns across dictionary atoms for the best-performing hyperparameter con￾figuration of all-MiniLM-L6-v2 with a dictionary size k of 64, and a linear encoder [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Top atoms for our best performing dictionary con￾figuration for all-MiniLM-L6-v2. Our token level atoms are interpretable, and specialize in specific syntactic features. The ten atoms with the highest label-assignment confidence (up to 92%) each specialize in a single …
Figure 7
Figure 7. Figure 7: Activation mean and variance across dictio [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: ) clearly dominates the final sentence rep￾resentation. These contributions are interpretable. For example, atom 5, associated with numerical tokens, has limited influence. Similarly, atom 58, which activates for adverbs describing actions, con￾tributes weakly, as seen…
Figure 9
Figure 9. Figure 9: Relative contributions of the top 12 dependencies tag to the overall sentence representation on our best trial of all-MiniLM-L6-v2 [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Relative contributions of part-of-speech tags to the overall sentence representation in our best-performing run using all-MiniLM-L6-v2. The plot includes only tags that account for more than 2% of the total attribution. mirror semantic importance, and we observe a lon…
Figure 11
Figure 11. Figure 11: Probes for token position across models. • Probe Results (subsection E.1): Learned atoms show near-orthogonality with a few tight clus￾ters. • Atom Orthogonality (subsection E.2): Pairwise cosine similarities confirm separation between atoms. • POS Activation Maps (su…
Figure 13
Figure 13. Figure 13: Cosine similarities between dictionary atoms, for the best-performing configuration of all-MiniLM (with k=64 atoms and a linear encoder). E.3 Atom activiation heat map aligned with POS In [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 12
Figure 12. Figure 12: Cosine similarity of dictionary atoms, all-MiniLM In [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 14
Figure 14. Figure 14: POS-aligned heatmaps for four sentence embedding models. E.4 Training Results [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Riemannian Geometry for Pre-trained Language Model Embeddings

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Aggregating per-token pullback metrics via the Fréchet mean on the SPD manifold outperforms Euclidean mean pooling for sentence classification, with most of the gain attributable to geometric aggregation rather than l...

Reference graph

Works this paper leans on

52 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

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

    2019

  2. [2]

    Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but tough-to-beat baseline for sentence embeddings. In International conference on learning representations

    work page 2017

  3. [3]

    Giorgio Barnab \`o , Antonio Uva, Sandro Pollastrini, Chiara Rubagotti, and Davide Bernardi. 2023. https://doi.org/10.18653/v1/2023.findings-ijcnlp.36 Supervised clustering loss for clustering-friendly sentence embeddings: an application to intent clustering . In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2023 (Findings), pages...

    work page doi:10.18653/v1/2023.findings-ijcnlp.36 2023

  4. [4]

    Yonatan Belinkov. 2021. https://arxiv.org/abs/2102.12452 Probing classifiers: Promises, shortcomings, and advances . Preprint, arXiv:2102.12452

    arXiv 2021

  5. [5]

    Vinamra Benara, Chandan Singh, John X Morris, Richard J Antonello, Ion Stoica, Alexander G Huth, and Jianfeng Gao. 2024. Crafting interpretable embeddings for language neuroscience by asking llms questions. Advances in neural information processing systems, 37:124137

    2024

  6. [6]

    Leonard Bereska and Efstratios Gavves. 2024. https://doi.org/10.48550/arXiv.2404.14082 Mechanistic Interpretability for AI Safety -- A Review . ArXiv:2404.14082 [cs]

    arXiv doi:10.48550/arxiv.2404.14082 2024

  7. [7]

    Rishi Bommasani, Kelly Davis, and Claire Cardie. 2020. https://doi.org/10.18653/v1/2020.acl-main.431 I nterpreting P retrained C ontextualized R epresentations via R eductions to S tatic E mbeddings . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4758--4781, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.acl-main.431 2020

  8. [8]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Ch...

    work page 2023

Show all 52 references
  1. [9]

    Deng Cai and Wai Lam. 2019. https://doi.org/10.18653/v1/D19-1393 Core semantic first: A top-down approach for AMR parsing . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language ...

    2019 doi

  2. [10]

    Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. https://doi.org/10.48550/arXiv.1906.04341 What Does BERT Look At ? An Analysis of BERT 's Attention . ArXiv:1906.04341 [cs]

    2019 doi arXiv

  3. [11]

    Alexis Conneau, German Kruszewski, Guillaume Lample, Loïc Barrault, and Marco Baroni. 2018. https://doi.org/10.18653/v1/P18-1198 What you can cram into a single \ &!\#* vector: Probing sentence embeddings for linguistic properties . In Proceedings of the 56th Annual Meeting of...

    2018 doi

  4. [12]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse autoencoders find highly interpretable features in language models. arXiv:2309.08600

    2023 arXiv

  5. [13]

    James Fodor, Simon De Deyne, and Shinsuke Suzuki. 2025. https://doi.org/10.1162/coli_a_00536 Compositionality and sentence meaning: Comparing semantic parsing and transformers on a challenging sentence similarity dataset . Computational Linguistics, 51(1):139--190

    2025 doi

  6. [14]

    W. N. Francis and H. Kucera. 1979. http://icame.uib.no/brown/bcm.html Brown corpus manual . Technical report, Department of Linguistics, Brown University, Providence, Rhode Island, US

    1979

  7. [15]

    Gangeh, Ahmed K

    Mehrdad J. Gangeh, Ahmed K. Farahat, Ali Ghodsi, and Mohamed S. Kamel. 2015. https://doi.org/10.48550/arXiv.1502.05928 Supervised Dictionary Learning and Sparse Representation - A Review . ArXiv:1502.05928

    2015 doi arXiv

  8. [16]

    Leo Gao, Tom Dupr \'e la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024. Scaling and evaluating sparse autoencoders. arXiv:2406.04093

    2024 arXiv

  9. [17]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. arXiv:2104.08821

    2021 arXiv

  10. [18]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

    2023 arXiv

  11. [19]

    Kailash A Hambarde and Hugo Proenca. 2023. Information retrieval: recent advances and beyond. IEEE Access

    2023

  12. [20]

    Yikun Han, Chunjiang Liu, and Pengfei Wang. 2023. A comprehensive survey on vector database: Storage and retrieval technique, challenge. arXiv:2310.11703

    2023

  13. [21]

    John Hewitt and Christopher D. Manning. 2019. https://doi.org/10.18653/v1/N19-1419 A Structural Probe for Finding Syntax in Word Representations . In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics : Human Langu...

    2019 doi

  14. [22]

    Huang, Wenlin Yao, Kaiqiang Song, Hongming Zhang, Muhao Chen, and Dong Yu

    James Y. Huang, Wenlin Yao, Kaiqiang Song, Hongming Zhang, Muhao Chen, and Dong Yu. 2023. https://doi.org/10.48550/arXiv.2305.14599 Bridging Continuous and Discrete Spaces : Interpretable Sentence Representation Learning via Compositional Operations . ArXiv:2305.14599 [cs]

    2023 doi arXiv

  15. [23]

    Ganesh Jawahar, Benoît Sagot, and Djamé Seddah. 2019. https://doi.org/10.18653/v1/P19-1356 What Does BERT Learn about the Structure of Language ? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 3651--3657, Florence, Italy. Ass...

    2019 doi

  16. [24]

    Abhinav Ramesh Kashyap, Thanh-Tung Nguyen, Viktor Schlegel, Stefan Winkler, See-Kiong Ng, and Soujanya Poria. 2023. A comprehensive survey of sentence representations: From the bert epoch to the chatgpt era and beyond. arXiv:2305.12641

    2023 arXiv

  17. [25]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  18. [26]

    Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. arXiv:2011.05864

    2020 arXiv

  19. [27]

    Leo Z Liu, Yizhong Wang, Jungo Kasai, Hannaneh Hajishirzi, and Noah A Smith. 2021. Probing across time: What does roberta know and when? arXiv:2104.07885

    2021 arXiv

  20. [28]

    Qi Liu, Matt J Kusner, and Phil Blunsom. 2020. A survey on contextual embeddings. arXiv:2003.07278

    2020 arXiv

  21. [29]

    Julien Mairal, Jean Ponce, Guillermo Sapiro, Andrew Zisserman, and Francis Bach. 2008. Supervised dictionary learning. Advances in neural information processing systems, 21

    2008

  22. [30]

    Lucas Moeller, Dmitry Nikolaev, and Sebastian Pad \'o . 2023. https://doi.org/10.18653/v1/2023.emnlp-main.980 An attribution method for S iamese encoders . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15818--15827, Singapore....

    2023 doi

  23. [31]

    Lucas Moeller, Dmitry Nikolaev, and Sebastian Pad \'o . 2024. https://aclanthology.org/2024.eacl-long.125 Approximate attributions for off-the-shelf S iamese transformers . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Lingu...

    2024

  24. [32]

    Vivi Nastase and Paola Merlo. 2024. https://aclanthology.org/2024.repl4nlp-1.15 Tracking linguistic information in transformer-based sentence embeddings through targeted sparsification . In Proceedings of the 9th Workshop on Representation Learning for NLP (RepL4NLP-2024), pag...

    2024

  25. [33]

    Juri Opitz and Anette Frank. 2022. https://doi.org/10.18653/v1/2022.aacl-main.48 SBERT studies Meaning Representations : Decomposing Sentence Embeddings into Explainable Semantic Features . In Proceedings of the 2nd Conference of the Asia - Pacific Chapter of the Association f...

    2022 doi

  26. [34]

    Juri Opitz, Lucas Möller, Andrianos Michail, and Simon Clematide. 2025. https://doi.org/10.48550/arXiv.2502.14862 Interpretable Text Embeddings and Text Similarity Explanation : A Primer . ArXiv:2502.14862 [cs]

    2025 doi

  27. [35]

    Tiago Pimentel, Josef Valvoda, Rowan Hall Maudslay, Ran Zmigrod, Adina Williams, and Ryan Cotterell. 2020. https://doi.org/10.18653/v1/2020.acl-main.420 Information- Theoretic Probing for Linguistic Structure . In Proceedings of the 58th Annual Meeting of the Association for C...

    2020 doi

  28. [36]

    Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. arXiv:2003.07082

    2020 arXiv

  29. [37]

    Abhilasha Ravichander, Yonatan Belinkov, and Eduard Hovy. 2021. https://doi.org/10.18653/v1/2021.eacl-main.295 Probing the probing paradigm: Does probing accuracy entail task relevance? ACL, pages 3363--3377

    2021 doi

  30. [39]

    Nils Reimers and Iryna Gurevych. 2019. http://arxiv.org/abs/1908.10084 Sentence- BERT : Sentence Embeddings using Siamese BERT - Networks . ArXiv:1908.10084 [cs]

    2019 arXiv

  31. [40]

    Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. https://doi.org/10.48550/arXiv.1905.05950 BERT Rediscovers the Classical NLP Pipeline . ArXiv:1905.05950 [cs]

    2019 doi arXiv

  32. [41]

    L Tesni \`e re. 1959. El \'e ments de syntaxe structurale

    1959

  33. [42]

    Alexandros Vasileiou and Oliver Eberle. 2024. https://doi.org/10.18653/v1/2024.naacl-long.435 Explaining text similarity in transformer models . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language...

    2024 doi

  34. [43]

    Elena Voita and Ivan Titov. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.14 Information- Theoretic Probing with Minimum Description Length . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing ( EMNLP ) , pages 183--196, Online. Assoc...

    2020 doi

  35. [44]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv:2212.03533

    2022 arXiv

  36. [45]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv:2402.05672

    2024 arXiv

  37. [46]

    John Wieting, Mohit Bansal, Kevin Gimpel, and Karen Livescu. 2015. Towards universal paraphrastic sentence embeddings. arXiv:1511.08198

    2015 arXiv

  38. [47]

    Jinming Xing, Dongwen Luo, Chang Xue, and Ruilin Xing. 2025. https://arxiv.org/abs/2411.14654 Comparative analysis of pooling mechanisms in llms: A sentiment analysis perspective . Preprint, arXiv:2411.14654

    2025 arXiv

  39. [48]

    Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. 2024. https://doi.org/10.18653/v1/2024.emnlp-industry.103 mGTE : Generalized long-context text representation and r...

    2024 doi

  40. [49]

    Judit Ács, Ákos Kádár, and András Kornai. 2021. https://doi.org/10.48550/arXiv.2102.10864 Subword Pooling Makes a Difference . ArXiv:2102.10864 [cs]

    2021 doi arXiv

  41. [50]

    online" 'onlinestring :=

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

  42. [51]

    write newline

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

  43. [52]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...

  44. [53]

    write newline

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

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.