Pith. sign in

REVIEW 4 major objections 4 minor 52 references

Rethinking Word Similarity: Semantic Similarity through Classification Confusion

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

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

Pith's one-line read This paper claims that a word-identity classifier's confusion probabilities—the chance it mistakes one word for another—can serve as a semantic similarity measure that matches or slightly exceeds cosine on human benchmarks and tracks…

desk verdict The core idea of measuring similarity as classifier confusion is genuinely novel and worth engaging, but the current evidence is too thin and partly confounded to support the benchmark claims. read the letter →

arxiv 2502.05704 v1 pith:KOCQJXBJ submitted 2025-02-08 cs.CL cs.AI

Kaitlyn Zhou , Haishan Gao , Sarah Chen , Dan Edelstein , Dan Jurafsky , Chen Shani This is my paper · ORCID
classification cs.CLcs.AI
keywords wordsimilarityclassificationconfusioncontextualembeddingsBERTsemanticchangeFrenchRevolutioncosine
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

The paper proposes that semantic similarity between two words can be measured by a classifier's confusion: the probability that a word-identity classifier, trained on contextual word representations, labels the embedding of one word as the other word. The method, Word Confusion, treats the classifier's class set as features the analyst chooses, so similarity is always relative to the dimensions that matter for a task. On the MEN, WordSim353, and SimLex human-judgment benchmarks, this confusion-based score matches or slightly exceeds cosine similarity between embedding vectors. The paper argues that this framing captures context-dependence, asymmetry, and polysemy better than geometric distance, and demonstrates it by tracing the French word 'révolution' from popular to state-associated meaning between 1789 and 1793.

What carries the argument

The central object is Word Confusion, a similarity measure built from the confusion matrix of a word-identity classifier. A one-vs-rest logistic regression is trained on contextual embeddings of chosen class words, averaged over the last four layers of BERT; at inference, the vector of class probabilities $p(w_i \mid e_j)$ defines how similar target $w_j$ is to each class $w_i$. The mechanism matters because classifier decision boundaries are not anchored at the origin the way cosine distances are, and the linear layer's parameters reweight the embedding geometry, so small, informative prediction errors become the similarity signal.

What would settle it

On the SimLex benchmark, split the word pairs into frequency bands and recompute Word Confusion's correlation with human ratings within each band; if the correlation largely disappears within bands, the reported match to humans is an artifact of word frequency rather than semantic confusion.

Watch

Extended reading notes

Core claim

Word Confusion defines similarity by $\mathrm{sim}(w_i, w_j) = p(w_i \mid e_j)$: the probability that a logistic-regression classifier, trained to map averaged BERT token embeddings to word identities over a chosen class set $W$, assigns the embedding $e_j$ of target word $w_j$ to class $w_i$. The confusion probabilities, not the predicted label, carry the semantic signal, and the class set acts as a dynamic feature set, operationalizing a feature-based view of similarity. The paper reports Spearman correlations with human judgments of 0.66 versus 0.59 for MEN, 0.67 versus 0.54 for WS353, and 0.44 versus 0.39 for SimLex, plus an average macro-F1 of 0.86 versus 0.80 across sentiment, grammatical-gender, and domain-classification tasks. In the historical case study, 'révolution' is classified as popular in 1789, 'contre-révolution' appears in the state cluster in 1792, and 'révolution' itself is classified with the state by 1793.

Load-bearing premise

The load-bearing premise is that the mistakes a word-identity classifier makes on averaged word representations reflect human-perceived semantic similarity, rather than artifacts of word frequency, polysemy, or the particular set of distractor classes chosen.

Editorial extensions

If this is right

  • Because similarity is defined as the probability assigned to each class, Word Confusion can measure how close an unseen word is to any set of analyst-chosen features, such as positive/negative sentiment or people/state in revolutionary discourse.
  • The measure is asymmetric: $p(\text{animal}\mid\text{cat}) \neq p(\text{cat}\mid\text{animal})$, so it can represent directional human judgments that cosine cannot.
  • It matches or slightly exceeds cosine similarity against human judgments on MEN, WordSim353, and SimLex, and outperforms three cosine baselines on sentiment, grammatical-gender, and concept-domain classification.
  • Applied to French parliamentary speeches from 1789 to 1793, it traces 'révolution' from popular to state-associated meaning, with 'contre-révolution' appearing first in the state cluster, offering a concrete temporal story about the concept.

Reading between the lines

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

  • If the central claim holds, any task that currently uses cosine similarity could be reframed around a task-specific class set, making similarity an analyst-controlled quantity rather than a fixed geometric distance.
  • The asymmetry property suggests Word Confusion could be used to surface directional relations such as hypernymy or connotation shifts, where 'cat is similar to animal' is not reversible.
  • Because the paper reports that classifier errors grow with frequency and polysemy, an immediate test is whether combining Word Confusion with cosine or frequency normalization improves robustness; this is an extension the authors do not run.
  • The French Revolution causal claim depends on the hand-picked seed words, so an independent validation should resample seed sets and check whether the people-to-state shift persists; the paper lists this as future work.
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 / 4 minor

Summary. The paper proposes Word Confusion, a semantic similarity measure in which the similarity of a target word w_j to a confound word w_i is defined as the probability p(w_i | e_j) assigned by a word-identity logistic regression classifier trained on contextual BERT embeddings (Eq. 1). The set of classifier classes acts as analyst-selectable features. The authors report that Word Confusion is comparable to or slightly better than cosine similarity on MEN, WS353, and SimLex (Table 1), achieves higher macro-F1 on feature classification tasks such as sentiment and grammatical gender (Table 2), and demonstrates a diachronic application to the French word 'révolution' in the Archives Parlementaires (Section 4.2), where they claim the state-oriented meaning first emerges via 'contre-révolution'.

Significance. If the empirical claims are validated, Word Confusion offers a useful alternative to cosine similarity for computational social science and cultural analytics: it is asymmetric, interpretable, and allows the analyst to choose the feature dimensions via the class set. The paper is clearly written as a proof-of-concept and provides code and data. The central premise, that classification confusion reflects human-perceived similarity, is plausible and worth testing. However, the current evidence is not yet sufficient to establish that the reported benchmark advantages are due to semantic signal rather than known classifier biases, and the theoretical derivation in Section 3.2 is mathematically flawed. The historical case study is intriguing but rests on a small and potentially unstable seed-word set.

major comments (4)
  1. [§2.1, Table 1, §A.1] The headline benchmark results in Table 1 may be an artifact of token frequency and polysemy. The paper's own error analysis (§A.1, Figures 4 and 5) shows that classification errors are far more likely for high-frequency and polysemous words: words with over 10 million occurrences have a 42% error rate versus 3% for rare words, and words with more than 10 WordNet senses are misidentified at 34% versus 4% for monosemous words. In a softmax classifier, such class-level imbalances can inflate the probability assigned to frequent or polysemous confound classes regardless of genuine semantic proximity. Table 1 reports only single-run correlations with no controls for word frequency, sense count, or class-balance in training sentences. The central claim that Word Confusion 'slightly outperforms' cosine similarity depends on ruling out this confound, e.g., by matching classes on frequency, by adding frequency as a covariate in the correlation analysis, or by reporting stratified results for low-frequency and monosemous pairs. Without such a control, the reported gains over cosine (0.66 vs 0.59 on MEN, 0.67 vs 0.54 on WS353, 0.44 vs 0.39 on SimLex) cannot be attributed to a better similarity measure.
  2. [§3.2] The SVD derivation of distance redefinition is mathematically garbled and does not support the text's conclusion. After writing A = UΣVᵀ, the paper states that the transformed vectors Ax and Ay can be rewritten as σ1u1v1ᵀx1 + σ2u2v2ᵀx2 and σ1u1v1ᵀy1 + σ2u2v2ᵀy2, which is incorrect: the correct expression is A x = Σ_i σ_i u_i (v_iᵀ x), which is a vector, not a sum of scalar terms. The subsequent claimed cosine distance formula omits cross terms between the singular components and incorrectly assumes σ1σ1 = 1 and σ1σ2 = 0. As written, this section does not convincingly show that a linear transformation redefines distance. The general point is true, but the derivation should be corrected or substantially rewritten.
  3. [§4.2] The historical case study's novel claim—that the state-oriented meaning of 'révolution' first emerges via 'contre-révolution'—is not supported by the quantitative evidence presented. The seed-word sets for 'the people' and 'the state' are listed with ellipses ({peuple, populaire, ...} and {conseil, gouvernement, ...}), and although the Limitations section states that different seed sets were explored 'without seeing drastic changes', those sets and the corresponding results are not reported. The narrative is based on a small set of occurrences (2,206 mentions of 'révolution' across 218 speeches) and a two-dimensional projection with no confidence intervals or significance tests. To make this case study persuasive, the authors should report the full seed sets, show stability of the temporal trend under seed-word variation, and provide at least a simple statistical test of the shift from popular to state association between 1789 and 1793. Alternatively, the claim should be explicitly framed as an illustrative, qualitative demonstration rather than a discovery.
  4. [Tables 1 and 2] All results in Tables 1 and 2 are single runs with no reported variance, confidence intervals, or significance tests. Given that the advantage over cosine is modest (e.g., SimLex 0.44 vs 0.39), these differences may not be statistically reliable. The authors should provide bootstrap confidence intervals over word pairs or repeated training runs with different random seeds, at least for the three benchmark datasets.
minor comments (4)
  1. [Abstract and §2.1] The abstract contains a typo: 'WirdSim353' should be 'WordSim353'. More importantly, §2.1 describes WordSim353 as containing '2000 word-pairs'; the actual WordSim353 dataset contains 353 word pairs. This should be corrected.
  2. [§3.2 and Figure 2] Figure 2 is difficult to interpret because the axes are not labeled and the method used to generate the data is not described. A brief caption explaining the artificial data and decision boundaries would improve clarity.
  3. [§6] In the Discussion, 'ofter' should be 'offers'. Also, the phrase 'captures different aspects of both similarity and multi-faceted words' is awkward and should be rephrased.
  4. [§4.2 and Appendix D] The corpus name is inconsistently written as 'Archive Parlementaires' in the text and 'Archives Parlementaires' in the footnote; use the standard 'Archives Parlementaires' throughout. In Appendix D, the years in Table 3 skip some periods (e.g., 1942-1955) without explanation; the segmentation should be described completely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Word Confusion is an empirical, self-supervised measurement; benchmarks and external historical data provide independent checks.

full rationale

The central derivation defines similarity as classifier confusion (Eq. 1) and measures it by training a word-identity logistic regression on BERT embeddings. The human similarity benchmarks (MEN, WS353, SimLex) supply external labels that are never used in training, so the reported correlations are genuine empirical outcomes rather than quantities forced by construction. The identifiability experiment in Section 3.1 is an independent sanity check, not an input to the similarity scores. The French Revolution analysis uses hand-selected seed words and reports classifier probabilities over temporal segments; the seed-word choice is acknowledged as a limitation, and the qualitative narrative is a post-hoc interpretation, not a fitted output. The inflation appendix is explicitly preliminary and inconclusive, and no claim that the central method predicts inflation is made. The same-author citations (Zhou et al. 2022a,b) support background claims about frequency and dispersion and are not load-bearing for the definition or the benchmark results. Concerns about frequency/polysemy confounds and the absence of error bars are validity issues, not circularity.

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

The method's results hinge on the analyst's choice of classes, the sampling schedule, and the embedding construction; these are inputs rather than derived quantities. The main axiomatic load is the assumption that confusion errors of a word-identity classifier are a faithful proxy for human similarity, which the benchmarks support only weakly. No new theoretical entities are introduced.

free parameters (5)
  • Analyst-chosen class set W = Benchmark vocabularies; seed sets such as {positive, negative}, {peuple, ...}, {gouvernement, ...}, {land, sea}…
    Similarity scores are defined only with respect to W (Section 2, Eq. 1). Different W produce different similarities, so all results are conditional on this choice, which is made by hand for each experiment.
  • Number of sampled occurrences per word = 30 for target words; 500-1000 for seed words; 2000 '$' tokens per segment
    Appendix B states 'having at least 1000 embeddings to train Word Confusion is necessary to get good and stable results', but no sensitivity analysis is reported for the benchmark or feature tasks.
  • Embedding construction = Average of the last four BERT hidden layers
    Chosen by the authors; alternative layer concatenations gave 'almost identical results' (Appendix A), but the choice is not derived from a principle.
  • Classifier hyperparameters = scikit-learn logistic regression, one-vs-rest, unspecified regularization
    The paper says no hyperparameters were tuned; this controls overfitting risk but leaves solver, C, and max iterations unspecified, hurting reproducibility.
  • Temporal segmentation and bucket count for inflation appendix = 17 segments based on DJI trends; 60 value buckets
    Appendix D: the segmentation and bucket count are arbitrary choices; the resulting correlations are weak (0.187 with CPI, 0.169 with DJI).
assumptions (5)
  • domain assumption Contextualized word embeddings contain enough word-identity information to be linearly separable.
    Section 3.1 reports 90% accuracy on a one-shot 1000-class task; the method relies on this separability, and the 90% is offered as support rather than a proof.
  • domain assumption A classifier's confusion probabilities, scaled through softmax of a linear model, correspond to human similarity judgments.
    This is the central premise of the paper; it is tested against MEN, WS353, and SimLex, but without error bars or multiple seeds.
  • ad hoc to paper The Archives Parlementaires corpus and the selected seed words represent the historical concepts 'the people' and 'the state' adequately.
    Section 4.2: seed words are given only partially ('{peuple, populaire, ...} and {conseil, gouvernement, ...}'); no checks for coverage or balance, and the authors acknowledge seed-word choice can affect similarities.
  • domain assumption Tversky's feature-based model of similarity is an appropriate cognitive foundation for word similarity.
    The paper's framing relies on Tversky (1977) but does not test whether the classifier's classes are cognitively analogous to features.
  • domain assumption Average pooled embeddings of a word across sentences, rather than individual contextualized tokens, preserve the similarity signal needed for the benchmarks.
    Section 2.1 and Appendix B use averaged embeddings; the paper does not compare against using full contextual distributions except in the one-shot identifiability experiment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Word Similarity: Semantic Similarity through Classification Confusion." pith.science (2026). https://pith.science/paper/KOCQJXBJ

@misc{pith2026250205704,
  author       = {Pith},
  title        = {Pith review of: Rethinking Word Similarity: Semantic Similarity through Classification Confusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KOCQJXBJ}},
  note         = {Machine review of arXiv:2502.05704}
}
read the original abstract

Word similarity has many applications to social science and cultural analytics tasks like measuring meaning change over time and making sense of contested terms. Yet traditional similarity methods based on cosine similarity between word embeddings cannot capture the context-dependent, asymmetrical, polysemous nature of semantic similarity. We propose a new measure of similarity, Word Confusion, that reframes semantic similarity in terms of feature-based classification confusion. Word Confusion is inspired by Tversky's suggestion that similarity features be chosen dynamically. Here we train a classifier to map contextual embeddings to word identities and use the classifier confusion (the probability of choosing a confounding word c instead of the correct target word t) as a measure of the similarity of c and t. The set of potential confounding words acts as the chosen features. Our method is comparable to cosine similarity in matching human similarity judgments across several datasets (MEN, WirdSim353, and SimLex), and can measure similarity using predetermined features of interest. We demonstrate our model's ability to make use of dynamic features by applying it to test a hypothesis about changes in the 18th C. meaning of the French word "revolution" from popular to state action during the French Revolution. We hope this reimagining of semantic similarity will inspire the development of new tools that better capture the multi-faceted and dynamic nature of language, advancing the fields of computational social science and cultural analytics and beyond.

Figures

Figures reproduced from arXiv: 2502.05704 by the authors.

Figure 1
Figure 1. Word Confusion: We predetermine a set of classes for our classifier, in this case {red, green, blue}. This choice of classes defines the similarity features used to describe the input word. At training, we extract sentences containing the chosen class words {red, green, blue}. We then train the classifier to map from a BERT contextual embeddings of these words to right class /feature (color, in this case). At infere… view at source ↗
Figure 2
Figure 2. Differences in decision boundaries between [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. In 1789, the word “revolution” was primarily associated with popular action (represented in orange). In 1792 “revolution” was now also seen as something that the government should lead (represented in blue) found in the “counter-revolution” cluster. In 1793, this new governmental meaning had spread back to the word “revolution” itself. We color-code the classes orange as “peuple” (“the people”) and blue as “gouverne… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The bar charts above highlight the percentage [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: The percentage of errors for words binned by [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Average CPI, DJI, and Word Confusion values between the years 1915-2009. For each temporal segment, the Word Confusion values were calcu￾lated using the mean predicted value for each item in the basket of goods. We can see that until the 1970s Word Confusion values fol…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 30 canonical work pages

  1. [1]

    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.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Eneko Agirre, Enrique Alfonseca, Keith Hall, Jana Kravalova, Marius Pa s ca, and Aitor Soroa. 2009. https://aclanthology.org/N09-1003 A study on similarity and relatedness using distributional and W ord N et-based approaches . In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North A merican Chapter of the Association for Co...

    work page 2009

  4. [4]

    Keith Michael Baker. 1990. https://doi.org/10.1017/CBO9780511625527 Inventing the French Revolution: Essays on French Political Culture in the Eighteenth Century . Ideas in Context. Cambridge University Press

    work page doi:10.1017/cbo9780511625527 1990

  5. [5]

    Paheli Bhattacharya, Kripabandhu Ghosh, Arindam Pal, and Saptarshi Ghosh. 2020. Methods for computing legal document similarity: A comparative study. arXiv preprint arXiv:2004.12307

    arXiv 2020

  6. [6]

    Terra Blevins and Luke Zettlemoyer. 2020. https://doi.org/10.18653/v1/2020.acl-main.95 Moving down the long tail of word sense disambiguation with gloss informed bi-encoders . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1006--1017, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.acl-main.95 2020

  7. [7]

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29

    2016

  8. [8]

    Arthur Bra z inskas, Serhii Havrylov, and Ivan Titov. 2017. Embedding words as distributions with a bayesian skip-gram model. arXiv preprint arXiv:1711.11027

    arXiv 2017

Show all 52 references
  1. [9]

    Elia Bruni, Nam-Khanh Tran, and Marco Baroni. 2014. Multimodal distributional semantics. Journal of artificial intelligence research, 49:1--47

    2014

  2. [10]

    Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183--186

    2017

  3. [11]

    Center for Bibliographic Studies and Research, University of California, Riverside . 2024. Courtesy of the california digital newspaper collection. Data retrieved from World Development Indicators, http://cdnc.ucr.edu

    2024

  4. [12]

    Vi \' e gas, and Martin Wattenberg

    Andy Coenen, Emily Reif, Ann Yuan, Been Kim, Adam Pearce, Fernanda B. Vi \' e gas, and Martin Wattenberg. 2019. http://arxiv.org/abs/1906.02715 Visualizing and measuring the geometry of BERT . CoRR, abs/1906.02715

    2019 arXiv

  5. [13]

    Fahim Dalvi, Abdul Rafae Khan, Firoj Alam, Nadir Durrani, Jia Xu, and Hassan Sajjad. 2022. https://openreview.net/forum?id=POTMtpYI1xH Discovering latent concepts learned in BERT . In International Conference on Learning Representations

    2022

  6. [14]

    Dan Edelstein. 2012. Do we want a revolution without revolution? reflections on political authority. French Historical Studies, 35(2):269--289

    2012

  7. [15]

    Katrin Erk and Marianna Apidianaki. 2024. Adjusting interpretable dimensions in embedding space with human judgments. arXiv preprint arXiv:2404.02619

    2024 arXiv

  8. [16]

    Kawin Ethayarajh. 2019. How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings. arXiv preprint arXiv:1909.00512

    2019 arXiv

  9. [17]

    Allyson Ettinger and Tal Linzen. 2016. Evaluating vector space models using human semantic priming results. In Proceedings of the 1st workshop on evaluating vector-space representations for NLP, pages 72--77

    2016

  10. [18]

    Ellen RK Evers and Dani \"e l Lakens. 2014. Revisiting tversky's diagnosticity principle. Frontiers in Psychology, 5:57776

    2014

  11. [19]

    Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences, 115(16):E3635--E3644

    2018

  12. [20]

    Gabriel Grand, Idan Asher Blank, Francisco Pereira, and Evelina Fedorenko. 2022. Semantic projection recovers rich human knowledge of multiple object features from word embeddings. Nature human behaviour, 6(7):975--987

    2022

  13. [21]

    William L Hamilton, Jure Leskovec, and Dan Jurafsky. 2016. Cultural shift or linguistic drift? comparing two computational measures of semantic change. In Proceedings of the conference on empirical methods in natural language processing. Conference on empirical methods in natu...

    2016

  14. [22]

    John Hewitt and Percy Liang. 2019. https://doi.org/10.18653/v1/D19-1275 Designing and interpreting probes with control tasks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Langua...

    2019 doi

  15. [23]

    Felix Hill, Roi Reichart, and Anna Korhonen. 2015. https://doi.org/10.1162/COLI_a_00237 S im L ex-999: Evaluating semantic models with (genuine) similarity estimation . Computational Linguistics, 41(4):665--695

    2015 doi

  16. [24]

    Nikhil Johri, Daniel Ramage, Daniel McFarland, and Daniel Jurafsky. 2011. https://aclanthology.org/W11-1516 A study of academic collaborations in computational linguistics using a latent mixture of authors model . In Proceedings of the 5th ACL - HLT Workshop on Language Techno...

    2011

  17. [25]

    Daphna Keidar, Andreas Opedal, Zhijing Jin, and Mrinmaya Sachan. 2022. https://doi.org/10.18653/v1/2022.acl-long.101 Slangvolution: A causal analysis of semantic change and frequency dynamics in slang . In Proceedings of the 60th Annual Meeting of the Association for Computati...

    2022 doi

  18. [26]

    Austin C Kozlowski, Matt Taddy, and James A Evans. 2019. The geometry of culture: Analyzing the meanings of class through word embeddings. American Sociological Review, 84(5):905--949

    2019

  19. [27]

    Liu, Matt Gardner, Yonatan Belinkov, Matthew E

    Nelson F. Liu, Matt Gardner, Yonatan Belinkov, Matthew E. Peters, and Noah A. Smith. 2019. https://doi.org/10.18653/v1/N19-1112 Linguistic knowledge and transferability of contextual representations . In Proceedings of the 2019 Conference of the North A merican Chapter of the ...

    2019 doi

  20. [28]

    Li Lucy, Divya Tadimeti, and David Bamman. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.228 Discovering differences in the representation of people using contextualized semantic axes . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processi...

    2022 doi

  21. [29]

    Qing Lyu, Marianna Apidianaki, and Chris Callison-burch. 2023. https://doi.org/10.18653/v1/2023.starsem-1.32 Representation of lexical stylistic features in language models ' embedding space . In Proceedings of the 12th Joint Conference on Lexical and Computational Semantics (...

    2023 doi

  22. [30]

    Navid Madani, Rabiraj Bandyopadhyay, Briony Swire-Thompson, Michael Miller Yoder, and Kenneth Joseph. 2023. http://arxiv.org/abs/2305.09548 Measuring social dimensions of self-presentation in social media biographies with an identity-based approach

    2023 arXiv

  23. [31]

    Thomas Manzini, Lim Yao Chong, Alan W Black, and Yulia Tsvetkov. 2019. https://doi.org/10.18653/v1/N19-1062 Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings . In Proceedings of the 2019 Conference of the North A merican ...

    2019 doi

  24. [32]

    Matej Martinc, Petra Kralj Novak, and Senja Pollak. 2020. https://aclanthology.org/2020.lrec-1.592 Leveraging contextual embeddings for detecting diachronic semantic shift . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4811--4819, Marseille...

    2020

  25. [33]

    Saif Mohammad, Svetlana Kiritchenko, and Xiaodan Zhu. 2013. https://aclanthology.org/S13-2053 NRC - C anada: Building the state-of-the-art in sentiment analysis of tweets . In Second Joint Conference on Lexical and Computational Semantics (* SEM ), Volume 2: Proceedings of the...

    2013

  26. [34]

    Bhat, and Pramod Viswanath

    Jiaqi Mu, S. Bhat, and Pramod Viswanath. 2018. All-but-the-top: Simple and effective postprocessing for word representations. ICLR, abs/1702.01417

    2018 arXiv

  27. [35]

    Aida Nematzadeh, Stephan C Meylan, and Thomas L Griffiths. 2017. Evaluating vector-space models of word representation, or, the unreasonable effectiveness of counting words near other words. In CogSci

    2017

  28. [36]

    B. New, C. Pallier, M. Brysbaert, and L. Ferrand. 2004. http://arxiv.org/abs/http://www.lexique.org/?page_id=294 Lexique 2 : A new french lexical database . Behavior Research Methods, Instruments, & Computers, 36(3):516--524

    2004

  29. [37]

    Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. https://doi.org/10.3115/1118693.1118704 Thumbs up? sentiment classification using machine learning techniques . In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing ( EMNLP 2002) , p...

    2002

  30. [38]

    Valentina Nicole Pescuma, Chiara Zanini, Davide Crepaldi, and Francesca Franzon. 2021. Form and function: A study on the distribution of the inflectional endings in italian nouns and adjectives. Frontiers in Psychology, page 4422

    2021

  31. [39]

    Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. https://doi.org/10.18653/v1/N19-1128 W i C : the word-in-context dataset for evaluating context-sensitive meaning representations . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association ...

    2019 doi

  32. [40]

    Sebasti \'a n A R \' os, Roberto A Silva, and Felipe Aguilera. 2012. A dissimilarity measure for automate moderation in online social networks. In Proceedings of the 4th International Workshop on Web Intelligence & Communities, pages 1--9

    2012

  33. [41]

    Saumya Sahai and Dravyansh Sharma. 2021. https://doi.org/10.18653/v1/2021.sigtyp-1.9 Predicting and explaining F rench grammatical gender . In Proceedings of the Third Workshop on Computational Typology and Multilingual NLP, pages 90--96, Online. Association for Computational ...

    2021 doi

  34. [42]

    Sewell Jr

    William H. Sewell Jr. and William Hamilton Sewell. 2005. Logics of history: Social theory and social transformation. University of Chicago Press

    2005

  35. [43]

    Pinky Sitikhu, Kritish Pahi, Pujan Thapa, and Subarna Shakya. 2019. A comparison of semantic similarity methods for maximum human interpretability. In 2019 artificial intelligence for transforming business and society (AITB), volume 1, pages 1--4. IEEE

    2019

  36. [44]

    Sahar Sohangir and Dingding Wang. 2017. Improved sqrt-cosine similarity measurement. Journal of Big Data, 4:1--13

    2017

  37. [45]

    Aina Garí Soler and Marianna Apidianaki. 2020. http://arxiv.org/abs/2010.02686 Bert knows punta cana is not just beautiful, it's gorgeous: Ranking scalar adjectives with contextualised representations

    2020 arXiv

  38. [46]

    William Timkey and Marten Van Schijndel. 2021. All bark and no bite: Rogue dimensions in transformer language models obscure representational quality. arXiv preprint arXiv:2109.04404

    2021 arXiv

  39. [47]

    Amos Tversky. 1977. Features of similarity. Psychological review, 84(4):327

    1977

  40. [48]

    Luke Vilnis and Andrew McCallum. 2014. Word representations via gaussian embedding. arXiv preprint arXiv:1412.6623

    2014 arXiv

  41. [49]

    Kelly Zhang and Samuel Bowman. 2018. https://doi.org/10.18653/v1/W18-5448 Language modeling teaches you more than translation does: Lessons learned through auxiliary syntactic task analysis . In Proceedings of the 2018 EMNLP Workshop B lackbox NLP : Analyzing and Interpreting ...

    2018 doi

  42. [50]

    Kaitlyn Zhou, Kawin Ethayarajh, Dallas Card, and Dan Jurafsky. 2022 a . Problems with cosine as a measure of embedding similarity for high frequency words. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), page...

    2022

  43. [51]

    Kaitlyn Zhou, Kawin Ethayarajh, Dallas Card, and Dan Jurafsky. 2022 b . https://doi.org/10.18653/v1/2022.acl-short.45 Problems with cosine as a measure of embedding similarity for high frequency words . In Proceedings of the 60th Annual Meeting of the Association for Computati...

    2022 doi

  44. [52]

    George Kingsley Zipf. 1945. The meaning-frequency relationship of words. The Journal of general psychology, 33(2):251--256

    1945

Pith tools

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