Upload 115 files

.gitattributes

@@ -56,4 +56,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+dataset/pdfs/01030000000113.pdf filter=lfs diff=lfs merge=lfs -text
+dataset/pdfs/01030000000163.pdf filter=lfs diff=lfs merge=lfs -text
 dataset/pdfs/* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,202 @@
+# DP-Bench: Document Parsing Benchmark
+
+<div align="center">
+  <img src="https://github.com/user-attachments/assets/54f8201b-5d3f-4673-b97d-848d8fcd10ff" width="800px">
+</div>
+
+Document parsing refers to the process of converting complex documents, such as PDFs and scanned images, into structured text formats like HTML and Markdown.
+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.
+
+While various parsers are available on the market, there is currently no standard evaluation metric to assess their performance.
+To address this gap, we propose a set of new evaluation metrics along with a benchmark dataset designed to measure parser performance.
+
+## Metrics
+We propose assessing the performance of parsers using three key metrics: 
+NID for element detection and serialization, TEDS and TEDS-S for table structure recognition.
+
+### Element detection and serialization
+
+**NID (Normalized Indel Distance).**
+NID evaluates how well a parser detects and serializes document elements according to human reading order.
+NID is similar to the [normalized edit distance](https://en.wikipedia.org/wiki/Levenshtein_distance) metric but excludes substitutions during evaluation, making it more sensitive to length differences between strings.
+
+The NID metric is computed as follows:
+
+$$
+NID = 1 - \frac{\text{distance}}{\text{len(reference)} + \text{len(prediction)}}
+$$
+
+The distance measures the similarity between the reference and predicted text, with values ranging from 0 to 1, where 0 represents perfect alignment and 1 denotes complete dissimilarity. 
+Here, the predicted text is compared against the reference text to determine how many character-level insertions and deletions are needed to match it.
+A higher NID score reflects better performance in both recognizing and ordering the text within the document's detected layout regions.
+
+### Table structure recognition
+
+Tables are one of the most complex elements in documents, often presenting both structural and content-related challenges. 
+Yet, during NID evaluation, table elements (as well as figures and charts) are excluded, allowing the metric to focus on text elements such as paragraphs, headings, indexes, and footnotes. To specifically evaluate table structure and content extraction, we use the [TEDS](https://arxiv.org/abs/1911.10683) and [TEDS-S](https://arxiv.org/abs/1911.10683) metrics.
+
+The [traditional metric](https://ieeexplore.ieee.org/document/1227792) fails to account for the hierarchical nature of tables (rows, columns, cells), but TEDS/TEDS-S measures the similarity between the predicted and ground-truth tables by comparing both structural layout and content, offering a more comprehensive evaluation. 
+
+**TEDS (Tree Edit Distance-based Similarity).**
+The TEDS metric is computed as follows:
+
+$$
+TEDS(T_a, T_b) = 1 - \frac{EditDist(T_a, T_b)}{\max(|T_a|, |T_b|)}
+$$
+
+The equation evaluates the similarity between two tables by modeling them as tree structures ($T_a$ and $T_b$).
+This metric evaluates how accurately the table structure is predicted, including the content of each cell.
+A higher TEDS score indicates better overall performance in capturing both the table layout and the textual content.
+
+**TEDS-S (Tree Edit Distance-based Similarity-Struct).**
+TEDS-S stands for Tree Edit Distance-based Similarity-Struct, measuring the structural similarity between the predicted and reference tables.
+
+## Benchmark dataset
+
+### Document sources
+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.
+
+| Sources   | Count |
+|------------|-------|
+| Library of Congress        | 90   |
+| Open educational resources | 90   |
+| Upstage                    | 20   |
+
+### Layout elements
+
+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. 
+
+This benchmark dataset was created by extracting pages with various layout elements from multiple types of documents. The layout elements consist of 12 element types: **Table, Paragraph, Figure, Chart, Header, Footer, Caption, Equation, Heading1, List, Index, Footnote**. This diverse set of layout elements ensures that our evaluation covers a wide range of document structures and complexities, providing a comprehensive assessment of document parsing capabilities.
+
+Note that only Heading1 is included among various heading sizes because it represents the main structural divisions in most documents, serving as the primary section title. 
+This high-level segmentation is sufficient for assessing the core structure without adding unnecessary complexity. 
+Detailed heading levels like Heading2 and Heading3 are omitted to keep the evaluation focused and manageable.
+
+| Category   | Count |
+|------------|-------|
+| Paragraph  | 804   |
+| Heading1   | 194   |
+| Footer     | 168   |
+| Caption    | 154   |
+| Header     | 101   |
+| List       | 91    |
+| Chart      | 67    |
+| Footnote   | 63    |
+| Equation   | 58    |
+| Figure     | 57    |
+| Table      | 55    |
+| Index      | 10    |
+
+### Dataset format
+
+The dataset is in JSON format, representing elements extracted from a PDF file, with each element defined by its position, layout class, and content. 
+The **category** field represents various layout classes, including but not limited to text regions, headings, footers, captions, tables, and more.
+The **content** field has three options: the **text** field contains text-based content, **html** represents layout regions where equations are in LaTeX and tables in HTML, and **markdown** distinguishes between regions like Heading1 and other text-based regions such as paragraphs, captions, and footers.
+Each element includes coordinates (x, y), a unique ID, and the page number it appears on. The dataset’s structure supports flexible representation of layout classes and content formats for document parsing.
+
+```
+{
+    "01030000000001.pdf": {
+        "elements": [
+            {
+                "coordinates": [
+                    {
+                        "x": 170.9176246670229,
+                        "y": 102.3493458064781
+                    },
+                    {
+                        "x": 208.5023846755278,
+                        "y": 102.3493458064781
+                    },
+                    {
+                        "x": 208.5023846755278,
+                        "y": 120.6598699131856
+                    },
+                    {
+                        "x": 170.9176246670229,
+                        "y": 120.6598699131856
+                    }
+                ],
+                "category": "Header",
+                "id": 0,
+                "page": 1,
+                "content": {
+                    "text": "314",
+                    "html": "",
+                    "markdown": ""
+                }
+            },
+            ...
+    ...
+```
+
+### Document domains
+| Domain                              | Subdomain               | Count |
+|-------------------------------------|-------------------------|-------|
+| Social Sciences                     | Economics               | 26    |
+|                                     | Political Science       | 18    |
+|                                     | Sociology               | 16    |
+|                                     | Law                     | 12    |
+|                                     | Cultural Anthropology   | 11    |
+|                                     | Education               | 8     |
+|                                     | Psychology              | 4     |
+| Natural Sciences                    | Environmental Science   | 26    |
+|                                     | Biology                 | 10    |
+|                                     | Astronomy               | 4     |
+| Technology                          | Technology              | 33    |
+| Mathematics and Information Sciences| Mathematics             | 13    |
+|                                     | Informatics             | 9     |
+|                                     | Computer Science        | 8     |
+|                                     | Statistics              | 2     |
+
+## Usage
+
+### Setup
+
+To setup the environment, clone the repository and install the required dependencies by running the following commands:
+
+```
+$ git clone https://github.com/UpstageAI/dp-bench.git
+$ cd dp-bench
+$ pip install -r requirements.txt
+```
+The repository includes necessary scripts for performing inference and evaluation of document parsers. 
+
+### Dataset
+The benchmark dataset can be found in the `dataset` folder. 
+It contains wide range of document layouts, from text-heavy pages to complex tables, enabling a thorough evaluation of the parser’s performance. 
+The dataset comes with annotations for layout elements such as paragraphs, headings, and tables.
+
+
+### Element detection and serialization evaluation
+This evaluation will compute the NID metric to assess how accurately the text in the document is recognized considering the structure and order of the document layout.
+To evaluate the document layout results, run the following command:
+
+```
+$ python evaluate.py \
+  --label_path <path to the reference json file> \
+  --pred_path <path to the predicted json file> \
+  --mode layout
+```
+
+
+### Table structure recognition evaluation
+This will compute TEDS-S (structural accuracy) and TEDS (structural and textual accuracy).
+To evaluate table recognition performance, use the following command:
+
+```
+$ python evaluate.py \
+  --label_path <path to the reference json file> \
+  --pred_path <path to the predicted json file> \
+  --mode table
+```
+
+# Leaderboard
+| **Source**           | **Date**  | **TEDS**  | **TEDS-S** | **NID**   | **Avg Time (sec.)** |
+|----------------------|-----------|------------|-----------|---------------------|
+| aws                  | 24.09.26  | 90.22      | 95.94     | 14.47               |
+| llamaparse           | 24.09.26  | 70.86      | 90.92     | 4.14                |
+| unstructured         | 24.09.26  | 69.9       | 90.42     | 13.14               |
+| google               | 24.09.26  | 68.75      | 90.09     | 5.85                |
+| microsoft            | 24.09.26  | 89.07      | 87.03     | 4.44                |
+| upstage              | 24.09.26  | **93.47**  | **96.27** | **3.79**            |

dataset/pdfs/01030000000004.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c00e243f924c0b33ccb0aede1a67b433927a1519cb33bfd704e23f6724dd560
+size 56400

dataset/pdfs/01030000000005.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daa5167329cda28a8a3841b6d5b8a37ee30273285f174af98ce5c0e5c76c0976
+size 213760

dataset/pdfs/01030000000009.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e22636736c43765307cbc5e876b0a18e340b52c909fc36043cefb38a7a11638
+size 220047

dataset/pdfs/01030000000013.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a266dd4ddaff135dc5b1ce3d960d4b29cfc96bee5f30f3fb879a0b730c57e6f7
+size 305072

dataset/pdfs/01030000000017.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bd9bf7f193c5670d79bea2d81c123c7dc3094467ec9616245f81a148b063c77
+size 132199

dataset/pdfs/01030000000018.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:321a42833cde040a5e4525c5f84ca2e7df18ee57649c9b47a0e5472e1357abb6
+size 51050

dataset/pdfs/01030000000019.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43725982fe76bf6e308a001ad81cab4781e803e1d8574e9896be7cfdf2fa8732
+size 44497

dataset/pdfs/01030000000021.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cfcbba6c953185e59b1edf5dfab0d36df31ef74a32548308deca21a8a96e615
+size 44252

dataset/pdfs/01030000000022.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:348bff737da27234885d263b9983064513fc33b1ff2ce858171e58f88c954c1a
+size 23578

dataset/pdfs/01030000000024.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da3f562f120acab3005c9735a195e3c57322c6b1439f22c72d7951eaa2544985
+size 36175

dataset/pdfs/01030000000025.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c00f679ff2c9acbd9ae3904fd350d0f24bea93fb90f44128c7c29a00483b6611
+size 35752

dataset/pdfs/01030000000027.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93ce1e59ab01c6352a98eca26b4a356b1e46e923cd60ea7145a458a3fbb5fd41
+size 69392

dataset/pdfs/01030000000031.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5b975710668800f4d4712155bc597a028f0ffbfe162efef92ded23cb2c23b9f
+size 92678

dataset/pdfs/01030000000032.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65be3f78649d4ee59fa39b9d2b94a44570e52ddbf31cb7213f92853bc23ceb16
+size 58311

dataset/pdfs/01030000000034.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1a31e05b65ada20646a1136bdc3293f68e3641893a48674c791bc2b8fadffa9
+size 73436

dataset/pdfs/01030000000039.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a5c949c602fb587c39618443a00915582b2d20a32f959e5d3f4f3fc99ab9db5
+size 72301

dataset/pdfs/01030000000042.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a14ebf0077dbbb76dd91f3d7d0b66b66b4ab62260846cc75fb466b85a4a762b4
+size 92394

dataset/pdfs/01030000000044.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29dfa90e45920fd79e68e82f0f171119fc1b4d494aac703edf21c3ea74dd6fca
+size 81994

dataset/pdfs/01030000000045.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00f0adaaa8358a28b4b4e83bc97dcd83a01f7283605b140c2be8e8d47bba8b6b
+size 113003

dataset/pdfs/01030000000048.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31e7fbfc5b038d471b2fc82b248c4807152284bc35d6d5bd036c0aaa42504f17
+size 27665

dataset/pdfs/01030000000049.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c33c2993fe1e5c2945bb894f1a8e54c756753c344bd66f11024715e6375f07d1
+size 28618

dataset/pdfs/01030000000051.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4682941b1a9a3ec96599d8188b673e8c6d1c4f2a555b5dc0d739c6c706815553
+size 41942

dataset/pdfs/01030000000054.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d09e9a559e05904e7a73e547f179ed617b016a94065e045b9f46bbe0e2ae8528
+size 143372

dataset/pdfs/01030000000056.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5699486e354a93df094404bae0cd17381bfab180f9ab2b5c9f55428c07a106c5
+size 156235

dataset/pdfs/01030000000057.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c8f779a89fd348e058a875f9182d6b3070199531adaa187bc3a602b5f746df3
+size 149578

dataset/pdfs/01030000000058.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0ab823f66709631e6226a937a8d68b52bde1e7ff6ef28086563c0646867c769
+size 156167

dataset/pdfs/01030000000059.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0687b8e7d87922833102a8e792d999184f73fd8ccbc147bba73ab036a28ca776
+size 189330

dataset/pdfs/01030000000063.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9aae5aa80d0ada0a9a0c1306993e96dcbdbec4cdf48766f56955cf8d3d8c46a4
+size 230042

dataset/pdfs/01030000000066.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9c49f563c6f47d9c28f9caf13455ef61345a0c9be4bba251d064958f546c6f5
+size 159229

dataset/pdfs/01030000000067.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30628cf0d5ec4d1d9dd94853d0711fdf496cbaa199fcab19f6fd4a4ddd34965d
+size 49682

dataset/pdfs/01030000000069.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42a416ed4de970abf9cc8ed74dbc543df8d7cc6ed24547189c0e8c77e5fed1d5
+size 121828

dataset/pdfs/01030000000073.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f99de206dc21ba8ea57548b1c0c007eb736da5f26c925dc99a4027806297bd8f
+size 103853

dataset/pdfs/01030000000074.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f974132511fc3655d9602cf317b6ae2ed040787671bb21e387322961b5ac30f1
+size 86417

dataset/pdfs/01030000000075.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a24a2d4ed5bc5e9aa585d4d29ea5b66737ea96bcaa84756d1b183e7c0ad77e91
+size 71165

dataset/pdfs/01030000000080.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2b60a018b0503378c804fbf3f75b8ff956d30b3fcecdc30c37781b7dbe0c1c4
+size 100546

dataset/pdfs/01030000000081.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24ab9416df440de9d792a4034e177eb71933a6eed72d4cf1ead17d22a787702f
+size 117413

dataset/pdfs/01030000000082.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ec2813813e8e7496d97aa4fa15dd414aa66cae2d6f62db818d30d9856ff0439
+size 83331

dataset/pdfs/01030000000083.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:970804c04c81dd9cc3ba80f0063b9551d290e83934c4e2c258e376845b3d180f
+size 87427

dataset/pdfs/01030000000084.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39a0d95abd6dc977dfc7986b57fbbd9a8331af6b5b603e023f41dfcbc67003c7
+size 62263

dataset/pdfs/01030000000085.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df678b583abe6afa6971fac9efa23189bd235105d697ae1b2b12d25e3b9fb8b4
+size 58137

dataset/pdfs/01030000000088.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1039bce8b948fac48b0b9a42f8266700a104aca355c3681a544d65e6b136b8d5
+size 62718

dataset/pdfs/01030000000091.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:475c037beeba4ba75b3e749b51f3781a67e8e8f9db6a31772f7bf2ed2bd05d00
+size 96408

dataset/pdfs/01030000000092.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cf2e856478f66d96690a306b55ff40c01ab24cb26d048fe248eac8e3d66e65c
+size 91072

dataset/pdfs/01030000000093.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9877e33f447b370ad098531c25e340b9e786d1e84f4ac7365bab9c6746b57247
+size 73894

dataset/pdfs/01030000000096.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b966d5aee5a17b377107ade67238d7215faef6303ac980080fd68534ab3ce5ae
+size 235201

dataset/pdfs/01030000000099.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a59c54dbc6c578c176e9c6627245740637da434d62a4460d865528ac3a2eecd4
+size 149455

dataset/pdfs/01030000000104.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f01f786e76db9a4e8e321342660e2a539d4b060e123b42a58ec84df128455d59
+size 123531