yxc97 commited on
Commit
fa9ac72
Β·
verified Β·
1 Parent(s): 41705d5

Upload DETAIL_README.md

Browse files
Files changed (1) hide show
  1. DETAIL_README.md +99 -0
DETAIL_README.md ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File structure
2
+
3
+ Here is an example of file structure of DocGenome dataset for discipline `math.GM`.
4
+
5
+ ```bash
6
+ math.GM
7
+ β”œβ”€β”€ 0906.1099
8
+ β”‚ β”œβ”€β”€ layout_annotation.json
9
+ β”‚ β”œβ”€β”€ order_annotation.json
10
+ β”‚ β”œβ”€β”€ page_xxxx.jpg
11
+ β”‚ β”œβ”€β”€ quality_report.json
12
+ β”‚ └── reading_annotation.json
13
+ └── 2103.02443
14
+ β”œβ”€β”€ layout_annotation.json
15
+ β”œβ”€β”€ order_annotation.json
16
+ β”œβ”€β”€ page_xxxx.jpg
17
+ β”œβ”€β”€ quality_report.json
18
+ └── reading_annotation.json
19
+ ```
20
+
21
+ Each paper folder, for example, `math.GM/2103.02443` contains five parts:
22
+
23
+ - 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.
24
+ - 2) `layout_annotation.json`, this json file contains the layout annotation bounding box of each category region using the COCO format.
25
+ - 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.
26
+ - 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.
27
+ - 5) `quality_report.json`, this json file contains the quality computing result for each page and the whole paper for further use.
28
+
29
+ # Layout annotation task
30
+
31
+ ## Category definition
32
+
33
+ | **Index** | **Category** | **Notes** |
34
+ |----------------|-------------------|------------------------------------------|
35
+ | 0 | Algorithm | |
36
+ | 1 | Caption | Titles of Images, Tables, and Algorithms |
37
+ | 2 | Equation | |
38
+ | 3 | Figure | |
39
+ | 4 | Footnote | |
40
+ | 5 | List | |
41
+ | 7 | Table | |
42
+ | 8 | Text | |
43
+ | 9 | Text-EQ | Text block with inline equations |
44
+ | 10 | Title | Section titles |
45
+ | 12 | PaperTitle | |
46
+ | 13 | Code | |
47
+ | 14 | Abstract | |
48
+
49
+ ## Known Issues
50
+
51
+ 1. The IoU of Bounding boxes are too large, this happens when the paper template is too complex.
52
+ 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.
53
+ 3. Bounding box is missing, this happens due to rare packages are used. Some rare packages may not identified by our rule-based methods.
54
+ 4. Bounding boxes are correct, but overlaps with other adjacent bounding boxe slightly, this happens due to layout adjustments, for example `vspace`, `input` commands.
55
+
56
+ # Order annotation category
57
+
58
+ ## Category Definition
59
+
60
+ | **Category** | **Description** | **Example** |
61
+ |----------------|-------------------|------------------------------------------|
62
+ | identical | two blocks corresponding to the same latex code chunk | paragraphs that cross columns or pages |
63
+ | peer | two blocks are both belongs to Title | \section{introduction}, \section{method} |
64
+ | sub | one block is a child of another block logically | \section{introduction} and the first paragraph in Introduction section |
65
+ | adj | two adjacent Text blocks | Paragraph1 and Paragraph2 |
66
+ | explicit-cite | one block cites another block with `ref` | As shown in \ref{Fig: 5}. |
67
+ | 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 |
68
+
69
+ ## Order annotation representation
70
+
71
+ Each `order_annotation.json` contains two keys:
72
+
73
+ - 1) `annotations`: contains the block information for each block, the `block_id` of each block is used to represent the relationship.
74
+ - 2) `orders`: contains a list of triples, the meaning of each triple is:
75
+ - 1. `type`, represents the category of the current relationship, see table above for details.
76
+ - 2. `from`, represents the `block_id` of the starting block of the relationship
77
+ - 3. `to`, represents the `block_id` of the ending block of the relationship
78
+
79
+ ## Known issues
80
+
81
+ 1. `reading_annotation.json` file of some papers may not contain the field `annotations` for unknown reason.
82
+ 2. `reading_annotation.json` doesn't contain the `implicit-cite` relationship, the `implicit-cite` relationship is used in test-dataset for efficiency consideration.
83
+ 3. `explicit-cite` only supports `Equation`, the support for `Table`, `Figrue` is developed after the training dataset is complete.
84
+
85
+
86
+ # Quality report
87
+
88
+ This file contains the rule-based quality check for further use. Explanation is as follows:
89
+
90
+ - 1) `num_pages`: the number of pages of the corresponding paper.
91
+ - 2) `num_columns`: 1 (single column) or 2 (two column), depends on the last page of the paper
92
+ - 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.
93
+ - 4) `page_quality` containing IoU information of each page and the whole paper:
94
+ - 1. `page`: page index
95
+ - 2. `num_blocks`: how many bounding boxes in this page
96
+ - 3. `area`: sum of area of all blocks, $\sum_i \text{area}(\text{bbox}_i)$
97
+ - 4. `overlap`: sum of intersection area of all blocks, $\sum_i\sum_{j>i} \text{area}(\text{bbox}_i\cap bbox_j)$
98
+ - 5. `ratio` the ratio between `overlap` and `area`. Note that this ratio may be very large if there is template issue.
99
+