Pith. sign in

REVIEW 4 major objections 6 minor 46 references

RiverText: A Python Library for Training and Evaluating Incremental Word Embeddings from Text Data Streams

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read RiverText standardizes training and evaluation of incremental word embeddings from text data streams in one open-source Python library.

desk verdict RiverText is a credible open-source resource for incremental word embeddings; the benchmark is suggestive but not conclusive because of OOV imputation and reporting gaps. read the letter →

arxiv 2506.23192 v1 pith:ZRO36426 submitted 2025-06-29 cs.CL cs.LG

Gabriel Iturra-Bocaz , Felipe Bravo-Marquez This is my paper · ORCID
classification cs.CLcs.LG
keywords incrementalwordembeddingstextdatastreamsstreamingmachinelearningsimilaritycategorizationnegativesamplingintrinsicevaluationPythonlibrary
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

RiverText is a Python library for training and evaluating word embeddings on text data streams, so language models can keep up with new words and changing meanings without full retraining. The paper claims this is the first open-source library to standardize incremental word embedding training and evaluation, bringing three methods (a count-based word-context matrix, incremental skip-gram with negative sampling, and an incremental continuous bag-of-words) under one interface. A periodic evaluation module adapts standard intrinsic word-similarity and word-categorization tests to the streaming setting, scoring the current vectors at regular intervals. The authors report a benchmark of 243 hyperparameter configurations and conclude that all three architectures are competitive after tuning. This matters for applications like social-media analysis, where hashtags, brand names, and word senses shift continuously.

What carries the argument

Three mechanisms carry the argument. The periodic evaluation loop (Algorithm 1) is the measurement device: after every p training instances it calls an intrinsic-task evaluator on the current vocabulary and vectors, producing a time series of similarity correlations and categorization purity scores. On the training side, IWCM maintains a PPMI-weighted word-context matrix with fixed-size vocabulary and contexts, reduced by incremental PCA; ISG and ICBOW share an adaptive unigram table from Kaji and Kobayashi, where word frequencies and a fixed-size sampling table are updated in a single pass, and a Misra-Gries algorithm bounds the vocabulary under constant memory. PyTorch provides the SGD updates for the two neural models. The combination of bounded vocabulary and periodic evaluation is what makes embeddings queryable and comparable at any point in the stream.

What would settle it

Run the same training streams with injected semantic change, such as a target word whose context distribution flips halfway through, and evaluate with both the paper's static gold datasets and time-stamped gold relations; if one model's scores collapse or diverge widely while the others hold, the claim that all three architectures are competitive would be contradicted. A simpler version: re-run the benchmark after replacing out-of-vocabulary averaging with discarding those test words and see whether the ranking in Table 1 changes.

Watch

Extended reading notes

Core claim

The paper's central claim is that incremental word embeddings can be packaged into one standardized open-source tool that both trains and evaluates them in a streaming setting. RiverText extends the interfaces of an existing stream machine-learning library, adding learn_one and learn_many methods so models process text instance-by-instance or in mini-batches, with bounded vocabulary memory enforced by a Misra-Gries sketch. It implements IWCM, ISG, and the first ICBOW with negative sampling, and it wraps the Word Embeddings Benchmark to run periodic intrinsic evaluations against MEN, Mturk, and AP datasets. The empirical comparison ranks the models, finding that the neural ISG and ICBOW models outperform the count-based IWCM on average, and the paper concludes that all three are competitive once their hyperparameters are tuned.

Load-bearing premise

The benchmark conclusion rests on the assumption that periodically scoring embeddings against static human-annotated similarity and category data measures quality in an evolving stream; the paper itself notes that out-of-vocabulary test words are averaged into the vocabulary and that the gold relations are treated as static.

Editorial extensions

If this is right

  • Text-stream deployments can keep embedding models current without periodically retraining on accumulated data, because the library updates vectors online and exposes them at any time.
  • The three implemented methods can be compared under identical conditions; the paper's benchmark gives each model a recommended hyperparameter setting.
  • Intrinsic word-similarity and word-categorization tests become stream-monitoring tools, producing a per-period quality signal rather than a single end-of-training score.
  • The bounded-vocabulary design means memory use stays roughly constant even for indefinitely long streams, with rare or old words evicted by the Misra-Gries sketch.
  • Because the interface follows the stream-learning library it extends, incremental embedding models can be slotted into larger streaming pipelines for tasks like classification or sentiment analysis.

Reading between the lines

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

  • If periodic static evaluation is accepted as a proxy, the same loop could double as a drift detector: a sustained drop in similarity correlation over the stream would signal semantic change before downstream task accuracy reports it.
  • The explicit word-context matrix in IWCM is inspectable, so a case study on a known semantic shift (for example, a word whose dominant context flips) could reveal which associations actually changed, whereas the neural models hide this structure.
  • The library currently implements only Misra-Gries for vocabulary sketching, so adding alternative eviction policies would likely change which rare words survive and could shift the benchmark rankings.
  • Comparing ICBOW's new negative-sampling implementation against a SpaceSaving-based streaming word2vec would isolate whether the adaptive unigram table or the neural architecture drives its performance.
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 / 6 minor

Summary. RiverText is a Python library for training and evaluating incremental word embeddings from text streams. It implements three models—Incremental Word Context Matrix (IWCM), Incremental Skip-gram with Negative Sampling (ISG), and Incremental Continuous Bag-of-Words with Negative Sampling (ICBOW)—within the River streaming-ML framework, and it provides a periodic intrinsic evaluation module for word similarity and categorization. The paper reports a benchmark of 243 hyperparameter configurations across the three architectures and concludes that all three are competitive after tuning, with concrete hyperparameter recommendations in Table 1.

Significance. If the resource claim holds, RiverText fills a real gap: no existing stream-ML library standardizes incremental word embedding training and evaluation. The paper's explicit contributions—open-source code, integration with River, and a unified evaluation protocol—are appropriate for a resource paper. The library design, including the use of Misra-Gries sketches and an incremental unigram table, is sensible and plausible. However, the experimental evidence that supports the comparative rankings and the "all three architectures are competitive" claim is not yet convincing; the evaluation protocol has an internal inconsistency and an unvalidated imputation step that could materially change the reported conclusions. The library itself appears to be a useful contribution, but the paper's experimental validation needs substantial strengthening before the benchmark claims can be accepted.

major comments (4)
  1. [§4.2.1 vs. Figure 1] The evaluation period p is inconsistent: §4.2.1 says p = 320,000 instances, while the Figure 1 caption says p = 3,200,000 instances. This factor-of-10 discrepancy changes the evaluation schedule and therefore the time-series means used to build Table 1. Please correct the inconsistency and ensure that the text, captions, and any released code or configuration files agree.
  2. [§3.2, §4.3, Table 1] The periodic evaluation assigns out-of-vocabulary test words the average embedding of in-vocabulary words. Because each model's Misra-Gries vocabulary evolves differently and is bounded, the OOV sets are not comparable across models or checkpoints: a word may be absent from one model's vocabulary because it was discarded, while it may never have been seen by another. The paper never reports OOV rates per model, dataset, or checkpoint, nor does it provide a sensitivity analysis (e.g., excluding OOV pairs). Without such evidence, the rankings in Table 1 and the claim that all three architectures are competitive may reflect vocabulary-management artifacts rather than embedding quality. Please report OOV rates and repeat the ranking with OOV pairs excluded or otherwise controlled.
  3. [§4.3, Table 1] The experimental results consist of single-run mean scores with no error bars, confidence intervals, or statistical tests. Since the neural models use negative sampling and randomized initialization, the differences between adjacent rows in Table 1 (e.g., overall means 0.407 vs. 0.406) may be noise. Please report multiple independent runs with variance, or justify why a single run is sufficient, and apply a significance test or a clear threshold for ranking stability.
  4. [§4.3, §5] The conclusion that "all three architectures are competitive" lacks a static baseline. No comparison is made to a non-incremental embedding method (e.g., a standard Word2Vec or GloVe trained on the same corpus). Without such a reference point, the reader cannot judge whether the incremental methods' absolute scores are competitive or merely internally comparable. Please add a static baseline or explicitly restrict the claim to relative performance among the three implemented methods.
minor comments (6)
  1. [§4] The first sentence of Section 4 contains a typo: "bechmarks" should be "benchmarks."
  2. [§2.1] The name "Kaji and Kobadashi" is misspelled; the correct spelling is "Kobayashi" throughout the related-work discussion.
  3. [Algorithm 1] The indentation of line 7 (c += length(batch)) is ambiguous: it should be clarified that the counter is incremented after the periodic evaluation check and that it belongs to the while loop. This will help readers verify the period semantics.
  4. [§4.3] The sentence "we ordered the obtained average, with the lower position indicating the optimal configuration" is inconsistent with Table 1, where position 1 corresponds to the highest overall mean (0.407), not the lowest. Please rephrase to describe the ranking direction correctly.
  5. [§5] The phrase "incremental words embedding" should be "incremental word embeddings."
  6. [§3.2] The paper notes that the OOV averaging "can impact the model's overall performance" and that "results should be interpreted cautiously," but it does not quantify any such impact in the experiments. A brief mention of this limitation in Section 5 or in the experimental discussion would be helpful, beyond the static-gold caveat already present.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RiverText is a systems/resource paper whose empirical benchmark relies on external gold datasets, and the few self-citations are not load-bearing.

full rationale

This paper is a software/resource description rather than a formal derivation, so there is no fitted input that is later renamed as a prediction and no theorem whose conclusion is built into its assumptions. The central contribution is the open-source library itself, which is externally checkable through the released repository. The periodic evaluation in Section 3.2 uses the external Word Embeddings Benchmark library [21] with the independent MEN [10], MTurk [38], and AP [3] datasets, and the Borda ranking in Table 1 is computed from those external judgments, not from quantities fitted to the same evaluation targets. The IWCM implementation is attributed to the authors' own prior work [9], and the paper's claim of being the first standardized library does not depend on that prior algorithm being correct; it is a described adaptation, not a load-bearing self-citation. The admitted limitations—out-of-vocabulary words receiving the average in-vocabulary embedding in Section 3.2 and the assumption in Section 5 that gold relations remain static—are evaluation-design weaknesses that could affect the empirical ranking, but they are not circular because the test data are not manufactured from model outputs or from parameters fitted to the conclusions. The statement that all three architectures are competitive is an empirical interpretation of Table 1, not a derivation forced by the evaluation definition. No circular step can be exhibited by quoting an equation in which the output is equivalent to the input, so the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The library claim and benchmark rest on standard stream-learning requirements, hand-selected hyperparameters, and domain assumptions about intrinsic evaluation. No new scientific entity such as a particle, force, or dimension is introduced.

free parameters (6)
  • embedding_size = Best: 100 for ICBOW, 100 for ISG, 100 for IWCM
    One of three values (100, 200, 300) selected by Borda ranking on the benchmark; the best-settings claim depends on this choice.
  • window_size = Best: 3 for ICBOW, 1 for ISG, 3 for IWCM
    Grid-searched over 1, 2, 3 and selected on evaluation data.
  • num_negative_samples = Best: 6 for ICBOW, 8 for ISG, not used for IWCM
    Grid-searched over 6, 8, 10 for the neural models; affects ISG and ICBOW quality.
  • context_size = Best: 1000 for IWCM
    Grid over 500, 750, 1000; needed for word-context matrix capacity.
  • period_p = 320,000 per Section 4.2.1; 3,200,000 per Figure 1 caption
    Controls when the evaluator runs; inconsistent across the manuscript, and the performance curves depend on it.
  • unigram_smoother_alpha = not reported
    Parameter of Algorithm 3 used to update the adaptive unigram table; no value is given, so the neural sampling distribution is under-specified.
assumptions (6)
  • domain assumption Words appearing in similar contexts have similar meanings (distributional hypothesis).
    Invoked in Section 2.1 to justify all three incremental embedding models.
  • domain assumption Intrinsic tasks such as word similarity and categorization with static human-annotated gold data are meaningful proxies for embedding quality in a stream.
    Section 3.2 uses MEN, MTurk, and AP as periodic evaluators; the authors acknowledge in Section 5 that gold relations are assumed static.
  • ad hoc to paper Assigning an out-of-vocabulary test word the average of in-vocabulary embeddings does not systematically bias evaluation.
    Section 3.2 states OOV words get the average embedding and warns that results should be interpreted cautiously; the benchmark rankings depend on this step.
  • standard math The Misra-Gries sketch keeps a bounded vocabulary while preserving enough frequency information for PPMI and negative sampling.
    Section 3 relies on Misra-Gries to bound vocabulary and context size in constant memory.
  • ad hoc to paper Algorithm 3 correctly implements the Kaji-Kobayashi adaptive unigram table.
    The pseudocode in Section 3.3.2 is the basis for negative sampling; it is under-specified, so correctness is assumed.
  • domain assumption Incremental PCA can reduce sparse high-dimensional word-context vectors without a full data pass.
    Section 3.3.1 applies incremental PCA for dimensionality reduction in IWCM; no error analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RiverText: A Python Library for Training and Evaluating Incremental Word Embeddings from Text Data Streams." pith.science (2026). https://pith.science/paper/ZRO36426

@misc{pith2026250623192,
  author       = {Pith},
  title        = {Pith review of: RiverText: A Python Library for Training and Evaluating Incremental Word Embeddings from Text Data Streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZRO36426}},
  note         = {Machine review of arXiv:2506.23192}
}
read the original abstract

Word embeddings have become essential components in various information retrieval and natural language processing tasks, such as ranking, document classification, and question answering. However, despite their widespread use, traditional word embedding models present a limitation in their static nature, which hampers their ability to adapt to the constantly evolving language patterns that emerge in sources such as social media and the web (e.g., new hashtags or brand names). To overcome this problem, incremental word embedding algorithms are introduced, capable of dynamically updating word representations in response to new language patterns and processing continuous data streams. This paper presents RiverText, a Python library for training and evaluating incremental word embeddings from text data streams. Our tool is a resource for the information retrieval and natural language processing communities that work with word embeddings in streaming scenarios, such as analyzing social media. The library implements different incremental word embedding techniques, such as Skip-gram, Continuous Bag of Words, and Word Context Matrix, in a standardized framework. In addition, it uses PyTorch as its backend for neural network training. We have implemented a module that adapts existing intrinsic static word embedding evaluation tasks for word similarity and word categorization to a streaming setting. Finally, we compare the implemented methods with different hyperparameter settings and discuss the results. Our open-source library is available at https://github.com/dccuchile/rivertext.

Figures

Figures reproduced from arXiv: 2506.23192 by the authors.

Figure 1
Figure 1. Best setting models for MEN, Mturk, and AP datasets. The period [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

  1. [1]

    RR Ade and PR Deshmukh. 2013. Methods for incremental learning: a survey. International Journal of Data Mining & Knowledge Management Process 3, 4 (2013), 119

    work page 2013

  2. [2]

    Charu C Aggarwal. 2007. Data streams: models and algorithms . Vol. 31. Springer

    work page 2007

  3. [3]

    Abdulrahman Almuhareb and Massimo Poesio. 2005. Concept learning and categorization from the web. In proceedings of the annual meeting of the Cognitive Science society, Vol. 27

    work page 2005

  4. [4]

    Matej Artac, Matjaz Jogan, and Ales Leonardis. 2002. Incremental PCA for on- line visual learning and recognition. In 2002 International Conference on Pattern Recognition, Vol. 3. IEEE, 781–784

    work page 2002

  5. [5]

    Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Sel- jebotn, and Kurt Smith. 2010. Cython: The best of both worlds. Computing in Science & Engineering 13, 2 (2010), 31–39

    work page 2010

  6. [6]

    Albert Bifet, Ricard Gavalda, Geoffrey Holmes, and Bernhard Pfahringer. 2018. Machine learning for data streams: with practical examples in MOA . MIT press. SIGIR ’23, July 23–27, 2023, Taipei, Taiwan Iturra-Bocaz & Bravo-Marquez

    work page 2018

  7. [7]

    Albert Bifet, Geoffrey Holmes, and Bernhard Pfahringer. 2011. Moa-tweetreader: real-time analysis in twitter streaming data. In International conference on discov- ery science. Springer, 46–60

    work page 2011

  8. [8]

    Albert Bifet, Geoff Holmes, Bernhard Pfahringer, Philipp Kranen, Hardy Kremer, Timm Jansen, and Thomas Seidl. 2010. Moa: Massive online analysis, a framework for stream classification and clustering. In Proceedings of the first workshop on applications of pattern analysis . PMLR, 44–50

    work page 2010

Show all 46 references
  1. [9]

    Felipe Bravo-Marquez, Arun Khanchandani, and Bernhard Pfahringer. 2022. Incre- mental Word Vectors for Time-Evolving Sentiment Lexicon Induction. Cognitive Computation 14, 1 (2022), 425–441

    2022

  2. [10]

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

    2014

  3. [11]

    Evandro Cunha, Gabriel Magno, Giovanni Comarela, Virgilio Almeida, Mar- cos André Gonçalves, and Fabricio Benevenuto. 2011. Analyzing the dynamic evolution of hashtags on twitter: a language-based approach. In Proceedings of the workshop on language in social media (LSM 2011) . 58–65

    2011

  4. [12]

    Jurafsky Daniel, Martin James H, et al. 2007. Speech and language processing: An introduction to natural language processing, computational linguistics, and speech recognition. prentice hall

    2007

  5. [13]

    Peter Emerson. 2013. The original Borda count and partial voting. Social Choice and Welfare 40, 2 (2013), 353–358

    2013

  6. [14]

    Xin Geng and Kate Smith-Miles. 2009. Incremental Learning

    2009

  7. [15]

    Anna Gladkova and Aleksandr Drozd. 2016. Intrinsic evaluations of word embed- dings: What can we do better?. In Proceedings of the 1st Workshop on Evaluating Vector-Space Representations for NLP. 36–42

    2016

  8. [16]

    Yoav Goldberg. 2017. Neural network methods for natural language processing. Synthesis lectures on human language technologies 10, 1 (2017), 1–309

    2017

  9. [17]

    Amit Goyal, Jagadeesh Jagarlamudi, Hal Daumé III, and Suresh Venkatasubra- manian. 2010. Sketching techniques for large scale NLP. In Proceedings of the NAACL HLT 2010 Sixth Web as Corpus Workshop . 17–25

    2010

  10. [18]

    Max Halford, Geoffrey Bolmier, Raphael Sourty, Robin Vaysse, and Adil Zouitine

  11. [19]

    Zellig S Harris. 1954. Distributional structure. Word 10, 2-3 (1954), 146–162

    1954

  12. [20]

    Sam Henry, Clint Cuffy, and Bridget T McInnes. 2018. Vector representations of multi-word terms for semantic relatedness. Journal of biomedical informatics 77 (2018), 111–119

    2018

  13. [21]

    Stanisław Jastrzebski, Damian Leśniak, and Wojciech Marian Czarnecki. 2017. How to evaluate word embeddings? on importance of data efficiency and simple supervised tasks. arXiv preprint arXiv:1702.02170 (2017)

    2017 arXiv

  14. [22]

    Nobuhiro Kaji and Hayato Kobayashi. 2017. Incremental skip-gram model with negative sampling. arXiv preprint arXiv:1704.03956 (2017)

    2017 arXiv

  15. [23]

    Omer Levy and Yoav Goldberg. 2014. Neural word embedding as implicit matrix factorization. Advances in neural information processing systems 27 (2014)

    2014

  16. [24]

    Christopher D Manning. 2008. Introduction to information retrieval . Syngress Publishing,

    2008

  17. [25]

    James H Martin. 2009. Speech and language processing: An introduction to natural language processing, computational linguistics, and speech recognition . Pearson- /Prentice Hall

    2009

  18. [26]

    Chandler May, Kevin Duh, Benjamin Van Durme, and Ashwin Lall. 2017. Streaming word embeddings with the space-saving algorithm. arXiv preprint arXiv:1704.07463 (2017)

    2017 arXiv

  19. [27]

    Ahmed Metwally, Divyakant Agrawal, and Amr El Abbadi. 2005. Efficient compu- tation of frequent and top-k elements in data streams. In International conference on database theory. Springer, 398–412

    2005

  20. [28]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)

    2013

  21. [29]

    Jayadev Misra and David Gries. 1982. Finding repeated elements. Science of computer programming 2, 2 (1982), 143–152

    1982

  22. [30]

    Jacob Montiel, Max Halford, Saulo Martiello Mastelini, Geoffrey Bolmier, Raphael Sourty, Robin Vaysse, Adil Zouitine, Heitor Murilo Gomes, Jesse Read, Talel Abdessalem, et al. 2021. River: machine learning for streaming data in Python. (2021)

    2021

  23. [31]

    Jacob Montiel, Jesse Read, Albert Bifet, and Talel Abdessalem. 2018. Scikit- multiflow: A multi-output streaming framework.The Journal of Machine Learning Research 19, 1 (2018), 2915–2914

    2018

  24. [32]

    Shanmugavelayutham Muthukrishnan et al. 2005. Data streams: Algorithms and applications. Foundations and Trends® in Theoretical Computer Science 1, 2 (2005), 117–236

    2005

  25. [33]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing system...

    2019

  26. [34]

    Hao Peng, Mengjiao Bao, Jianxin Li, Md Zakirul Alam Bhuiyan, Yaopeng Liu, Yu He, and Erica Yang. 2018. Incremental term representation learning for social network analysis. Future Generation Computer Systems 86 (2018), 1503–1512

    2018

  27. [35]

    Hao Peng, Jianxin Li, Yangqiu Song, and Yaopeng Liu. 2017. Incrementally learn- ing the hierarchical softmax function for neural language models. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 31

    2017

  28. [36]

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) . 1532–1543

    2014

  29. [37]

    Saša Petrović, Miles Osborne, and Victor Lavrenko. 2010. The edinburgh twit- ter corpus. In Proceedings of the NAACL HLT 2010 workshop on computational linguistics in a world of social media . 25–26

    2010

  30. [38]

    Kira Radinsky, Eugene Agichtein, Evgeniy Gabrilovich, and Shaul Markovitch

  31. [39]

    Jesse Read, Albert Bifet, Bernhard Pfahringer, and Geoff Holmes. [n. d.]. Batch- Incremental vs. Instance-Incremental Learning in Dynamic and Evolving Data. ([n. d.])

  32. [40]

    Tobias Schnabel, Igor Labutov, David Mimno, and Thorsten Joachims. 2015. Evaluation methods for unsupervised word embeddings. In Proceedings of the 2015 conference on empirical methods in natural language processing . 298–307

    2015

  33. [41]

    Ian Stewart, Dustin Arendt, Eric Bell, and Svitlana Volkova. 2017. Measuring, predicting and visualizing short-term change in word representation and usage in vkontakte social network. In Eleventh international AAAI conference on web and social media

    2017

  34. [42]

    Peter D Turney and Patrick Pantel. 2010. From frequency to meaning: Vector space models of semantics. Journal of artificial intelligence research 37 (2010), 141–188

    2010

  35. [43]

    Clark Wissler. 1905. The Spearman correlation formula. Science 22, 558 (1905), 309–311

    1905

  36. [44]

    Michael Zhai, Johnny Tan, and Jinho Choi. 2016. Intrinsic and extrinsic evalu- ations of word embeddings. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 30

    2016

  37. [2011]

    In Proceedings of the 20th international conference on World wide web

    A word at a time: computing word relatedness using temporal semantic analysis. In Proceedings of the 20th international conference on World wide web . 337–346

  38. [2019]

    creme, a Python library for online machine learning

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.