File structure

Here is an example of file structure of DocGenome dataset for discipline math.GM.

math.GM
├── 0906.1099
│   ├── layout_annotation.json
│   ├── order_annotation.json
│   ├── page_xxxx.jpg
│   ├── quality_report.json
│   └── reading_annotation.json
└── 2103.02443
    ├── layout_annotation.json
    ├── order_annotation.json
    ├── page_xxxx.jpg
    ├── quality_report.json
    └── reading_annotation.json

Each paper folder, for example, math.GM/2103.02443 contains five parts:

• 1. page_xxxx.jpg, each image represents each page of the corresponding paper, the page index is contained in the filename. Note that this might be different from the original paper.
• 2. layout_annotation.json, this json file contains the layout annotation bounding box of each category region using the COCO format.
• 3. reading_annotation.json, this json file contains LaTex source code for each block (except for the Figure category). Note that the latex source code may contain macros.
• 4. order_annotation.json, this json file contains the relationship between different blocks, where the key of orders consists of triplets. Each triplet represents the relation type and specifies the source block and the destination block.
• 5. quality_report.json, this json file contains the quality computing result for each page and the whole paper for further use.

Layout annotation task

Category definition

Index	Category	Notes
0	Algorithm
1	Caption	Titles of Images, Tables, and Algorithms
2	Equation
3	Figure
4	Footnote
5	List
7	Table
8	Text
9	Text-EQ	Text block with inline equations
10	Title	Section titles
12	PaperTitle
13	Code
14	Abstract

Known Issues

1. The IoU of Bounding boxes are too large, this happens when the paper template is too complex.
2. The category of the bounding boxes are not correct. This happens when user-defined macros are used. For example, some authors may use \newcommand{\beq}{\begin{equation}}, \newcommand{\eeq}{\end{equation}}, in this case, the equation may be detected as Text class.
3. Bounding box is missing, this happens due to rare packages are used. Some rare packages may not identified by our rule-based methods.
4. Bounding boxes are correct, but overlaps with other adjacent bounding boxe slightly, this happens due to layout adjustments, for example vspace, input commands.

Order annotation category

Category Definition

Category	Description	Example
identical	two blocks corresponding to the same latex code chunk	paragraphs that cross columns or pages
peer	two blocks are both belongs to Title	\section{introduction}, \section{method}
sub	one block is a child of another block logically	\section{introduction} and the first paragraph in Introduction section
adj	two adjacent Text blocks	Paragraph1 and Paragraph2
explicit-cite	one block cites another block with ref	As shown in \ref{Fig: 5}.
implicit-cite	The caption block and the corresponding float environment	\begin{table}\caption{A}\begin{tabular}B\end{tabular}\end{table}, then A implicit-cite B

Order annotation representation

Each order_annotation.json contains two keys:

• 1. annotations: contains the block information for each block, the block_id of each block is used to represent the relationship.
• 2. orders: contains a list of triples, the meaning of each triple is:
  • 1. type, represents the category of the current relationship, see table above for details.
  • 2. from, represents the block_id of the starting block of the relationship
  • 3. to, represents the block_id of the ending block of the relationship

Known issues

1. reading_annotation.json file of some papers may not contain the field annotations for unknown reason.
2. reading_annotation.json doesn't contain the implicit-cite relationship, the implicit-cite relationship is used in test-dataset for efficiency consideration.
3. explicit-cite only supports Equation, the support for Table, Figrue is developed after the training dataset is complete.

Quality report

This file contains the rule-based quality check for further use. Explanation is as follows:

• 1. num_pages: the number of pages of the corresponding paper.
• 2. num_columns: 1 (single column) or 2 (two column), depends on the last page of the paper
• 3. category_quality: we record the number rendered latex code chunks for each category reading_count, and the number of detected bounding boxes geometry_count, then missing_rate is computed as (reading_count - geometry_count)/reading_count. Finally, the Total category is the summary of all other categories.
• 4. page_quality containing IoU information of each page and the whole paper:
  • 1. page: page index
  • 2. num_blocks: how many bounding boxes in this page
  • 3. area: sum of area of all blocks, $\sum_i \text{area}(\text{bbox}_i)$
  • 4. overlap: sum of intersection area of all blocks, $\sum_i\sum_{j>i} \text{area}(\text{bbox}_i\cap bbox_j)$
  • 5. ratio the ratio between overlap and area. Note that this ratio may be very large if there is template issue.