Document Elements

Introduction

The unstructured library aims to simplify and streamline the preprocessing of structured and unstructured documents for downstream tasks. And what that means is no matter where your data is and no matter what format that data is in, Unstructured’s toolkit will transform and preprocess that data into an easily digestable and usable format.

Document elements

When we partition a document, the output is a list of document Element objects. These element objects represent different components of the source document. Currently, the unstructured library supports the following element types:

  • type

    • FigureCaption

    • NarrativeText

    • ListItem

    • Title

    • Address

    • Table

    • PageBreak

    • Header

    • Footer

    • UncategorizedText

    • Image

    • Formula

  • element_id

  • metadata - see: Metadata page

  • text

Other element types that we will add in the future include tables and figures. Different partitioning functions use different methods for determining the element type and extracting the associated content. Document elements have a str representation. You can print them using the snippet below.

elements = partition(filename="example-10k.html")

for element in elements[:5]:
    print(element)
    print("\n")

One helpful aspect of document elements is that they allow you to cut a document down to the elements that you need for your particular use case. For example, if you’re training a summarization model you may only want to include narrative text for model training. You’ll notice that the output above includes a lot of titles and other content that may not be suitable for a summarization model. The following code shows how you can limit your output to only narrative text with at least two sentences. As you can see, the output now only contains narrative text.

from unstructured.documents.elements import NarrativeText
from unstructured.partition.text_type import sentence_count

for element in elements[:100]:
    if isinstance(element, NarrativeText) and sentence_count(element.text) > 2:
        print(element)
        print("\n")

Tables

For Table elements, the raw text of the table will be stored in the text attribute for the Element, and HTML representation of the table will be available in the element metadata under element.metadata.text_as_html. For most documents where table extraction is available, the partition function will extract tables automatically if they are present. For PDFs and images, table extraction requires a relatively expensive call to a table recognition model, and so for those document types table extraction is an option you need to enable. If you would like to extract tables for PDFs or images, pass in infer_table_structure=True. Here is an example (Note: this example requires the pdf extra. This can be installed with pip install "unstructured[pdf]"):

from unstructured.partition.pdf import partition_pdf

filename = "example-docs/layout-parser-paper.pdf"

elements = partition_pdf(filename=filename, infer_table_structure=True)
tables = [el for el in elements if el.category == "Table"]

print(tables[0].text)
print(tables[0].metadata.text_as_html)

The text will look like this:

Dataset Base Model1 Large Model Notes PubLayNet [38] F / M M Layouts of modern scientific documents PRImA [3] M - Layouts of scanned modern magazines and scientific reports Newspaper [17] F - Layouts of scanned US newspapers from the 20th century TableBank [18] F F Table region on modern scientific and business document HJDataset [31] F / M - Layouts of history Japanese documents

And the text_as_html metadata will look like this:

<table><thead><th>Dataset</th><th>| Base Model’</th><th>| Notes</th></thead><tr><td>PubLayNet</td><td>[38] F/M</td><td>Layouts of modern scientific documents</td></tr><tr><td>PRImA [3]</td><td>M</td><td>Layouts of scanned modern magazines and scientific reports</td></tr><tr><td>Newspaper</td><td>F</td><td>Layouts of scanned US newspapers from the 20th century</td></tr><tr><td>TableBank</td><td>F</td><td>Table region on modern scientific and business document</td></tr><tr><td>HJDataset [31]</td><td>F/M</td><td>Layouts of history Japanese documents</td></tr></table>

Converting Elements to Dictionary or JSON

The final step in the process for most users is to convert the output to JSON. You can convert a list of document elements to a list of dictionaries using the convert_to_dict function. The workflow for using convert_to_dict appears below.

from unstructured.staging.base import convert_to_dict

convert_to_dict(elements)

The unstructured library also includes utilities for saving a list of elements to JSON and reading a list of elements from JSON, as seen in the snippet below

from unstructured.staging.base import elements_to_json, elements_from_json


filename = "outputs.json"
elements_to_json(elements, filename=filename)
elements = elements_from_json(filename=filename)

Unique Element IDs

By default, the element ID is a SHA-256 hash of the element text. This is to ensure that the ID is deterministic. One downside is that the ID is not guaranteed to be unique. Different elements with the same text will have the same ID, and there could also be hash collisions. To use UUIDs in the output instead, you can pass unique_element_ids=True into any of the partition functions. This can be helpful if you’d like to use the IDs as a primary key in a database, for example.

from unstructured.partition.text import partition_text

elements = partition_text(text="Here is some example text.", unique_element_ids=True)
elements[0].id

Wrapping it all up

To conclude, the basic workflow for reading in a document and converting it to a JSON in unstructured looks like the following:

from unstructured.partition.auto import partition
from unstructured.staging.base import elements_to_json

input_filename = "example-docs/example-10k.html"
output_filename = "outputs.json"

elements = partition(filename=input_filename)
elements_to_json(elements, filename=output_filename)