arXMLiv 08.2017 - Word Embeddings; Token Model

Part of the arXMLiv project at the KWARC research group

Author

Deyan Ginev

Release

This page documents: 08.2017
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.

A 5 billion token model for the arXMLiv 08.2017 dataset
- token_model.zip
300 dimensional GloVe word embeddings for the arXMLiv 08.2017 dataset
- glove.arxmliv.5B.300d.zip and vocab.arxmliv.zip
300d GloVe word embeddings for individual subsets
- glove.subsets.zip
the main arXMLiv dataset is available separately here

Token Model Statistics

subset	documents	paragraphs	sentences
no_problem	112,088	3,760,015	17,684,762
warning	574,638	35,215,866	144,166,524
error	401,644	28,555,173	111,798,273
complete	1,088,370	67,531,054	273,649,559

subset	words	formulas	inline cite	numeric literals
no_problem	355,253,671	17,020,161	2,991,053	9,913,009
warning	2,514,340,590	219,167,820	20,163,304	65,294,846
error	1,946,207,151	169,247,016	14,458,082	51,730,645
complete	4,815,801,412	405,434,997	37,612,439	126,938,500

GloVe Model Statistics

subset	tokens	unique words	unique words (freq 5+ )
no_problem	384,951,086	490,134	170,615
warning	2,817,734,902	1,200,887	422,524
error	2,180,119,361	1,889,392	518,609
complete	5,382,805,349	2,573,974	746,673

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 v0.1, corpus_token_model example used for token model extraction
- used llamapun math-aware sentence and word tokenization
- processed logical paragraphs only (excluded non-textual modalities, e.g. tables, figures and their captions, bibliographies)
- marked up formulas replaced with mathformula token
- marked up inline citations replaced with citationelement token
- numeric literals replaced with NUM token
- ignored sentences with unnaturally long words (>30 characters) - almost always due to latexml conversion errors
GloVe repository at sha 76507
- 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: 26.49% (3665/13833)
- Highest score: “gram3-comparative.txt”, 75.83% (1010/1332)
warning
- Total accuracy: 31.16% (4989/16013)
- Highest score: “gram3-comparative.txt”, 75.45% (1005/1332)
error
- Total accuracy: 29.63% (4997/16867)
- Highest score: “gram3-comparative.txt”, 76.58% (1020/1332)
complete
- Total accuracy: 32.86% (5770/17562)
- Highest score: “gram3-comparative.txt”, 78.53% (1046/1332)
demo baseline: text8 demo (first 100M characters of Wikipedia)
- Total accuracy: 23.91% (4262/17827)
- Highest score: “capital-common-countries.txt”, 62.65% (317/506)

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:

python eval/python/word_analogy.py --vocab_file vocab.arxmliv.txt --vectors_file glove.arxmliv.5B.300d.txt

abelian is to group as disjoint is to ?
- Top hit: union, cosine distance 0.644784
convex is to concave as positive is to ?
- Top hit: negative, cosine distance 0.802866
finite is to infinite as abelian is to ?
- Top hit: nonabelian, cosine distance 0.664235
quantum is to classical as bottom is to ?
- Top hit: top, cosine distance 0.719843
eq is to proves as figure is to ?
- Top hit: shows, cosine distance 0.674743

Nearest word vectors