File size: 3,052 Bytes
c229954
 
6395107
 
 
 
 
 
62c8b23
 
ef218e0
 
 
 
7791de1
ef218e0
7791de1
ef218e0
7791de1
 
3c5ce62
ef218e0
7791de1
62c8b23
 
 
7791de1
 
 
 
 
 
 
62c8b23
 
 
 
 
7791de1
 
 
 
 
 
62c8b23
7791de1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
---
license: mit
task_categories:
- text-classification
language:
- en
size_categories:
- 100K<n<1M
source_datasets:
- original
configs:
- config_name: default
  data_files:
  - split: train
    path: train.zip
  - split: validation
    path: val.zip
  - split: test
    path: test.zip
pretty_name: MSC
---
## Dataset Description
The text data (title and abstract) of 164,230 arXiv preprints which are associated with at least one [MSC (mathematical subject classification)](https://en.wikipedia.org/wiki/Mathematics_Subject_Classification) code. Predicting 3-character MSC codes based on the cleaned text (processed title+abstarct) amounts to a multi-label classification task.

## Dataset Structure
- The column `cleaned_text` should be used as the input of the text classification task. This is obtained from processing the text data (titles and abstracts) of math-related preprints.
- The last 531 columns are one-hot encoded MSC classes, and should be used as target variables of the multi-label classification task.
- Other columns are auxiliary:
  - `url`) the URL of the preprint (the latest version as of December 2023),
  - `title`) the original title,
  - `abstract`) the original abstract,
  - `primary_category`) the primary [arXiv category](https://arxiv.org/category_taxonomy) (for this data, almost always a category of the math archive, or the mathematical physics archive). 
- **Subtask**) Predicting `primary_category` based on `cleaned_text`, a multi-class text classification task with ~30 distinct labels.   


## Data Splits
Stratified sampling was used for splitting the data so that the proportions of a target variable among the splits are not very different. 

|Dataset  |Description       |Number of instances   |
|---------|------------------|----------------------|
|main.zip |the whole data    |164,230               |
|train.zip|the training set  |104,675               |
|val.zip  |the validation set|18,540                |
|test.zip |the test set      |41,015                |

## Data Collection and Cleaning
The details are outlined in this [notebook](https://github.com/FilomKhash/Math-Preprint-Classifier/blob/main/Scarping%20and%20Cleaning%20the%20Data.ipynb). 
As for the raw data, with the help of the [arxiv package](https://pypi.org/project/arxiv/), we scraped preprints listed, or cross-listed, under the math archive. This raw data was then processed:

- dropping preprints with an abnormally high number of versions,
- keeping only the last arXiv version,
- dropping preprints whose metadata does not include any MSC class,
- dropping entries with pre-2010 mathematics subject classification convention,
- concatenating abstract and title strings and carrying out the following steps to obtain the `cleaned_text` column:
  - removing the LaTeX math environment and URL citations,
  - make the text lower case, normalizing accents and removing special characters,
  - removing English and some corpus-specific stop words,
  - stemming.
 
## Citation
<https://github.com/FilomKhash/Math-Preprint-Classifier>