DP-Bench: Document Parsing Benchmark

Name: dp-bench
Creator: upstage
License: https://choosealicense.com/licenses/mit/

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 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 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 normalized distance in the equation 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 and TEDS-S metrics.

The traditional metric 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 structure and the content of each cell.

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. While the metric formulation is identical to TEDS, it uses modified tree representations, denoted as $T_{a}^{'}$ and $T_{b}^{'}$ , where the nodes correspond solely to the table structure, omitting the content of each cell. This allows TEDS-S to concentrate on assessing the structural similarity of the tables, such as row and column alignment, without being influenced by the contents within the cells.

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 documents. 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 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
Social Sciences	Political Science	18
Social Sciences	Sociology	16
Social Sciences	Law	12
Social Sciences	Cultural Anthropology	11
Social Sciences	Education	8
Social Sciences	Psychology	4
Natural Sciences	Environmental Science	26
Natural Sciences	Biology	10
Natural Sciences	Astronomy	4
Technology	Technology	33
Mathematics and Information Sciences	Mathematics	13
Mathematics and Information Sciences	Informatics	9
Mathematics and Information Sciences	Computer Science	8
Mathematics and Information Sciences	Statistics	2

Usage

Setup

Before setting up the environment, make sure to install Git LFS, which is required for handling large files. Once installed, you can clone the repository and install the necessary dependencies by running the following commands:

$ git clone https://huggingface.co/datasets/upstage/dp-bench.git
$ cd dp-bench
$ pip install -r requirements.txt

The repository includes necessary scripts for inference and evaluation, as described in the following sections.

Inference

We offer inference scripts that let you request results from various document parsing services. For more details, refer to this README.

Evaluation

The benchmark dataset can be found in the dataset folder. It contains a 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.

The following options are required for evaluation:

--ref_path: Specifies the path to the reference JSON file, predefined as dataset/reference.json for evaluation purposes.
--pred_path: Indicates the path to the predicted JSON file. You can either use a sample result located in the dataset/sample_results folder, or generate your own by using the inference script provided in the scripts folder.

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 \
  --ref_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 \
  --ref_path <path to the reference json file> \
  --pred_path <path to the predicted json file> \
  --mode table

Leaderboard

Source	Request date	TEDS ↑	TEDS-S ↑	NID ↑	Avg. Time (secs) ↓
upstage	2024-10-24	93.48	94.16	97.02	3.79
aws	2024-10-24	88.05	90.79	96.71	14.47
llamaparse	2024-10-24	74.57	76.34	92.82	4.14
unstructured	2024-10-24	65.56	70.00	91.18	13.14
google	2024-10-24	66.13	71.58	90.86	5.85
microsoft	2024-10-24	87.19	89.75	87.69	4.44