arXMLiv 08.2018 - Word Embeddings; Token Model

Part of the arXMLiv project at the KWARC research group

Author

Deyan Ginev

Release

This page documents: 08.2018
Latest: 08.2019

Accessibility and License

The content of this Dataset is licensed to SIGMathLing members for research and tool development purposes.

Access is restricted to SIGMathLing members under the SIGMathLing Non-Disclosure-Agreement as for most arXiv articles, the right of distribution was only given (or assumed) to arXiv itself.

An 11.5 billion token model for the arXMLiv 08.2018 dataset, including subformula lexemes
- token_model.zip
300 dimensional GloVe word embeddings for the arXMLiv 08.2018 dataset
- glove.arxmliv.11B.300d.zip and vocab.arxmliv.zip
300d GloVe word embeddings for individual subsets
- glove.subsets.zip
Embeddings and vocabulary with math lexemes omitted
- glove.arxmliv.nomath.11B.300d.zip and vocab.arxmliv.nomath.zip
- added on July 20, 2019
- used as a control when evaluating the contribution of formula lexemes
the main arXMLiv dataset is available separately here

Token Model Statistics

subset	documents	paragraphs	sentences
no_problem	137,864	4,646,203	21,533,963
warning	705,095	45,797,794	183,246,777
error	389,225	26,759,524	99,641,978
complete	1,232,184	77,203,521	304,422,718

subset	words	formulas	inline cite	numeric literals
no_problem	430,217,995	20,910,732	3,709,520	11,177,753
warning	3,175,663,430	281,832,412	25,337,574	83,606,897
error	1,731,971,035	153,186,264	13,145,561	43,399,720
complete	5,337,852,460	455,929,408	42,192,655	138,184,370

GloVe Model Statistics

subset	tokens	unique words	unique words (freq 5+ )
no_problem	622,968,267	715,433	219,304
warning	7,203,536,205	3,478,235	666,317
error	3,691,805,321	2,444,532	574,467
complete	11,518,309,793	5,285,379	1,000,295

Citing this Resource

Please cite the main dataset when using the word embeddings, as they are generated and distributed jointly. Instructions here

Download

Download links. This is a temporary solution as we are in the process of migrating the data.
SIGMathLing members only. Joining is free and mostly a legal checkmark on our end - all researchers welcome!

Generated via

Generation Parameters

token model distributed as 3 subsets - no_problem, warning and error. complete model is derived via:

   cat token_model_no_problem.txt \
       token_model_warning.txt \
       token_model_error.txt > token_model_complete.txt

llamapun v0.2, corpus_token_model example used for token model extraction
- processed logical paragraphs only (excluded non-textual modalities, e.g. tables, figures and their captions, bibliographies, as well as abstracts)
- excluded paragraphs containing latexml errors (marked via a ltx_ERROR HTML class)
- used llamapun math-aware sentence and word tokenization, with subformula math lexemes
- marked up inline citations replaced with citationelement token
- numeric literals replaced with NUM token (both in text and formulas)
- ignored sentences with unnaturally long words (>25 characters) - almost always due to latexml conversion errors
GloVe repository at sha 07d59d
- build/vocab_count -min-count 5
- build/cooccur -memory 32.0 -window-size 15
- build/shuffle -memory 32.0
- build/glove -threads 16 -x-max 100 -iter 25 -vector-size 300 -binary 2

Examples and baselines

GloVe in-built evaluation (non-expert tasks e.g. language, relationships, geography)

no_problem
- Total accuracy: 27.47% (4028/14663)
- Highest score: “gram3-comparative.txt”, 75.30% (1003/1332)
warning
- Total accuracy: 31.97% (5351/16736)
- Highest score: “gram3-comparative.txt”, 77.40% (1031/1332)
error
- Total accuracy: 29.67% (4910/16549)
- Highest score: “gram3-comparative.txt”, 72.75% (969/1332)
complete
- Total accuracy: 35.48% (6298/17750)
- Highest score: “gram3-comparative.txt”, 76.65% (1021/1332)
demo baseline: text8 demo (first 100M characters of Wikipedia)
- Total accuracy: 23.62% (4211/17827)
- Highest score: “gram6-nationality-adjective.txt”, 58.65% (892/1521)

Evaluation note: These in-built evlauation runs are provided as a sanity check that the generated GloVe models pass a basic baseline against the non-expert tasks in the default GloVe suite. One would need a scienctific discourse tailored set of test cases to evaluate the arXiv-based models competitively.

Measuring word analogy

In a cloned GloVe repository, start via:

python2 eval/python/word_analogy.py --vocab_file vocab.arxmliv.txt --vectors_file glove.arxmliv.11B.300d.txt

abelian is to group as disjoint is to ?
- Top hit: union, cosine distance 0.653377
convex is to concave as positive is to ?
- Top hit: negative, cosine distance 0.786877
finite is to infinte as abelian is to ?
- Top hit: nonabelian, cosine distance 0.676042
quantum is to classical as bottom is to ?
- Top hit: top, cosine distance 0.734896
eq is to proves as figure is to ?
- Top hit: showing, cosine distance 0.668502
italic_x is to italic_y as italic_a is to ?
- Top hit: italic_b, cosine distance 0.902608

Nearest word vectors