shinseung428
commited on
Commit
•
a53e9bc
1
Parent(s):
b837da3
Update README.md
Browse files
README.md
CHANGED
@@ -1,9 +1,10 @@
|
|
1 |
# DP-Bench: Document Parsing Benchmark
|
2 |
|
3 |
<div align="center">
|
4 |
-
<img src="https://
|
5 |
</div>
|
6 |
|
|
|
7 |
Document parsing refers to the process of converting complex documents, such as PDFs and scanned images, into structured text formats like HTML and Markdown.
|
8 |
It is especially useful as a preprocessor for [RAG](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) systems, as it preserves key structural information from visually rich documents.
|
9 |
|
@@ -56,12 +57,16 @@ TEDS-S stands for Tree Edit Distance-based Similarity-Struct, measuring the stru
|
|
56 |
### Document sources
|
57 |
The benchmark dataset is gathered from three sources: 90 samples from the Library of Congress; 90 samples from Open Educational Resources; and 20 samples from Upstage's internal resources, which is composed of company/product materials and [Solar 10.7B](https://arxiv.org/abs/2312.15166) paper. Together, these sources provide a broad and specialized range of information.
|
58 |
|
59 |
-
|
60 |
-
|
|
|
|
|
61 |
| Library of Congress | 90 |
|
62 |
| Open educational resources | 90 |
|
63 |
| Upstage | 20 |
|
64 |
|
|
|
|
|
65 |
### Layout elements
|
66 |
|
67 |
While works like [ReadingBank](https://github.com/doc-analysis/ReadingBank) often focus solely on text conversion in document parsing, we have taken a more detailed approach by dividing the document into specific elements, with a particular emphasis on table performance.
|
@@ -72,8 +77,10 @@ Note that only Heading1 is included among various heading sizes because it repre
|
|
72 |
This high-level segmentation is sufficient for assessing the core structure without adding unnecessary complexity.
|
73 |
Detailed heading levels like Heading2 and Heading3 are omitted to keep the evaluation focused and manageable.
|
74 |
|
|
|
|
|
75 |
| Category | Count |
|
76 |
-
|
77 |
| Paragraph | 804 |
|
78 |
| Heading1 | 194 |
|
79 |
| Footer | 168 |
|
@@ -87,6 +94,8 @@ Detailed heading levels like Heading2 and Heading3 are omitted to keep the evalu
|
|
87 |
| Table | 55 |
|
88 |
| Index | 10 |
|
89 |
|
|
|
|
|
90 |
### Dataset format
|
91 |
|
92 |
The dataset is in JSON format, representing elements extracted from a PDF file, with each element defined by its position, layout class, and content.
|
@@ -130,9 +139,11 @@ Each element includes coordinates (x, y), a unique ID, and the page number it ap
|
|
130 |
...
|
131 |
```
|
132 |
|
|
|
|
|
133 |
### Document domains
|
134 |
| Domain | Subdomain | Count |
|
135 |
-
|
136 |
| Social Sciences | Economics | 26 |
|
137 |
| | Political Science | 18 |
|
138 |
| | Sociology | 16 |
|
@@ -149,6 +160,8 @@ Each element includes coordinates (x, y), a unique ID, and the page number it ap
|
|
149 |
| | Computer Science | 8 |
|
150 |
| | Statistics | 2 |
|
151 |
|
|
|
|
|
152 |
## Usage
|
153 |
|
154 |
### Setup
|
@@ -192,11 +205,15 @@ $ python evaluate.py \
|
|
192 |
```
|
193 |
|
194 |
# Leaderboard
|
195 |
-
|
196 |
-
|
|
|
|
|
197 |
| aws | 24.09.26 | 90.22 | 95.94 | 14.47 |
|
198 |
| llamaparse | 24.09.26 | 70.86 | 90.92 | 4.14 |
|
199 |
| unstructured | 24.09.26 | 69.9 | 90.42 | 13.14 |
|
200 |
| google | 24.09.26 | 68.75 | 90.09 | 5.85 |
|
201 |
| microsoft | 24.09.26 | 89.07 | 87.03 | 4.44 |
|
202 |
-
| upstage | 24.09.26 |
|
|
|
|
|
|
1 |
# DP-Bench: Document Parsing Benchmark
|
2 |
|
3 |
<div align="center">
|
4 |
+
<img src="https://cdn-uploads.huggingface.co/production/uploads/6524ab1e27d1f3d84ad07705/Q7CC2z4CAJzZ4-CGaSnBO.png" width="800px">
|
5 |
</div>
|
6 |
|
7 |
+
|
8 |
Document parsing refers to the process of converting complex documents, such as PDFs and scanned images, into structured text formats like HTML and Markdown.
|
9 |
It is especially useful as a preprocessor for [RAG](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) systems, as it preserves key structural information from visually rich documents.
|
10 |
|
|
|
57 |
### Document sources
|
58 |
The benchmark dataset is gathered from three sources: 90 samples from the Library of Congress; 90 samples from Open Educational Resources; and 20 samples from Upstage's internal resources, which is composed of company/product materials and [Solar 10.7B](https://arxiv.org/abs/2312.15166) paper. Together, these sources provide a broad and specialized range of information.
|
59 |
|
60 |
+
<div style="width: 500px;">
|
61 |
+
|
62 |
+
| Sources | Count|
|
63 |
+
|:--------------------------:|:----:|
|
64 |
| Library of Congress | 90 |
|
65 |
| Open educational resources | 90 |
|
66 |
| Upstage | 20 |
|
67 |
|
68 |
+
</div>
|
69 |
+
|
70 |
### Layout elements
|
71 |
|
72 |
While works like [ReadingBank](https://github.com/doc-analysis/ReadingBank) often focus solely on text conversion in document parsing, we have taken a more detailed approach by dividing the document into specific elements, with a particular emphasis on table performance.
|
|
|
77 |
This high-level segmentation is sufficient for assessing the core structure without adding unnecessary complexity.
|
78 |
Detailed heading levels like Heading2 and Heading3 are omitted to keep the evaluation focused and manageable.
|
79 |
|
80 |
+
<div style="width: 500px;">
|
81 |
+
|
82 |
| Category | Count |
|
83 |
+
|:----------:|:-----:|
|
84 |
| Paragraph | 804 |
|
85 |
| Heading1 | 194 |
|
86 |
| Footer | 168 |
|
|
|
94 |
| Table | 55 |
|
95 |
| Index | 10 |
|
96 |
|
97 |
+
</div>
|
98 |
+
|
99 |
### Dataset format
|
100 |
|
101 |
The dataset is in JSON format, representing elements extracted from a PDF file, with each element defined by its position, layout class, and content.
|
|
|
139 |
...
|
140 |
```
|
141 |
|
142 |
+
<div style="width: 800px;">
|
143 |
+
|
144 |
### Document domains
|
145 |
| Domain | Subdomain | Count |
|
146 |
+
|:-----------------------------------:|:-----------------------:|:-----:|
|
147 |
| Social Sciences | Economics | 26 |
|
148 |
| | Political Science | 18 |
|
149 |
| | Sociology | 16 |
|
|
|
160 |
| | Computer Science | 8 |
|
161 |
| | Statistics | 2 |
|
162 |
|
163 |
+
</div>
|
164 |
+
|
165 |
## Usage
|
166 |
|
167 |
### Setup
|
|
|
205 |
```
|
206 |
|
207 |
# Leaderboard
|
208 |
+
<div style="width: 800px;">
|
209 |
+
|
210 |
+
| Source | Date | TEDS | TEDS-S | NID |
|
211 |
+
|:--------------------:|:---------:|:----------:|:---------:|:-------------------:|
|
212 |
| aws | 24.09.26 | 90.22 | 95.94 | 14.47 |
|
213 |
| llamaparse | 24.09.26 | 70.86 | 90.92 | 4.14 |
|
214 |
| unstructured | 24.09.26 | 69.9 | 90.42 | 13.14 |
|
215 |
| google | 24.09.26 | 68.75 | 90.09 | 5.85 |
|
216 |
| microsoft | 24.09.26 | 89.07 | 87.03 | 4.44 |
|
217 |
+
| upstage | 24.09.26 | 93.47 | 96.27 | 3.79 |
|
218 |
+
|
219 |
+
</div>
|