title |
---|
Document elements and metadata |
When Unstructured partitions a file, the result is a list of document elements, sometimes referred to simply as elements. These elements represent different components of the source file.
Here's an example of what an element might look like:
{
"type": "NarrativeText",
"element_id": "5ef1d1117721f0472c1ad825991d7d37",
"text": "The Unstructured documentation covers the following services:",
"metadata": {
"last_modified": "2024-05-01T14:15:22",
"page_number": 1,
"languages": ["eng"],
"parent_id": "56f24319ae258b735cac3ec2a271b1d9",
"file_directory": "/content",
"filename": "Unstructured documentation.html",
"filetype": "text/html"
}
}
Every element has a type; an element_id; the extracted text
; and some metadata which might
vary depending on the element type, file structure, and some additional settings that are applied during
partitioning, chunking, and enriching. Optionally, the element can also have an
embeddings derived from the text
; the length of embeddings
depends on the embedding model that is used.
Instead of treating all files as strings of plain text, Unstructured preserves the semantic structure of the files. This gives you more control and flexibility over how you further use the processed files and allows you to take their structure into consideration. At the same time, normalizing data from various file formats to the Unstructured element type scheme lets you treat all files the same in your downstream processing, regardless of source format. For example, if you plan to summarize a file, you might only be interested in the narrative of the file and not care about its headers and footers. You can easily filter out the elements you don't need by specifying their type.
Here are some examples of the element types your file might contain:
Element type | Description |
---|---|
Address |
A text element for capturing physical addresses. |
CodeSnippet |
A text element for capturing code snippets. |
EmailAddress |
A text element for capturing email addresses. |
FigureCaption |
An element for capturing text associated with figure captions. |
Footer |
An element for capturing document footers. |
FormKeysValues |
An element for capturing key-value pairs in a form. |
Formula |
An element containing formulas in a file. |
Header |
An element for capturing document headers. |
Image |
A text element for capturing image metadata. |
ListItem |
ListItem is a NarrativeText element that is part of a list. |
NarrativeText |
NarrativeText is an element consisting of multiple, well-formulated sentences. This excludes elements such titles, headers, footers, and captions. |
PageBreak |
An element for capturing page breaks. |
PageNumber |
An element for capturing page numbers. |
Table |
An element for capturing tables. |
Title |
A text element for capturing titles. |
UncategorizedText |
Base element for capturing free text from within files. Applies to extracted text not associated with bounding boxes if the input is a PDF file. |
If you apply chunking, you will also see the CompositeElement
type.
CompositeElement
is a chunk formed from text (non-Table
) elements.
A composite element might be formed by combining one or more sequential elements produced by partitioning. For example,
several individual list items might be combined into a single chunk.
For the most up-to-date list of available element types, see the TYPE_TO_TEXT_ELEMENT_MAP
type-annotated mapping definition and the
ElementType
class definition in the elements.py
file, located in the Unstructured-IO/unstructured
repository in GitHub.
By default, the element ID is a SHA-256 hash of the element's text, its position on the page, the page number it's on, and the name of the related file. This is to ensure that the ID is deterministic and unique at the file level.
Unstructured tracks a variety of metadata about the elements extracted from files. Metadata is tracked at the element level within metadata
.
Element metadata enables you to do things such as:
- Filter file elements based on an element's metadata value. For instance, you might want to limit your scope to elements from a certain page, or you might want to use only elements that have an email matching a regular expression in their metadata.
- Map an element to the page where it occurred so that the original page can be retrieved when that element matches search criteria.
For the most up-to-date list of all available metadata fields, see the
ElementMetadata
class definition in the elements.py
file, located in the Unstructured-IO/unstructured
repository in GitHub.
All file types return the following metadata
fields when the information is available from the source file:
Metadata field name | Description |
---|---|
category_depth |
The depth of the element relative to other elements of the same category. Category depth is the depth of an element relative to other elements of the same category. It is set by a file partitioner and enables the document hierarchy after processing to compute more accurate hierarchies. Category depth might be set using native document hierarchies, for example reflecting <H1> or <H2> tags within an HTML file or the indentation level of a bulleted list item in a Word document. |
coordinates |
Any X-Y bounding box coordinates. |
detection_class_prob |
The detection model class probabilities. Applies only to Unstructured inference using the High Res strategy. |
emphasized_text_contents |
The related emphasized text (bold or italic) in the original file. |
emphasized_text_tags |
Any tags on the text that are emphasized in the original file. |
file_directory |
The related file's directory. |
filename |
The related file's filename. |
filetype |
The related file's type. |
is_continuation |
True if the element is a continuation of a previous element. Only relevant for chunking, if an element was divided into two due to Max Characters. |
languages |
Document languages at the file or element level. The list is ordered by probability of being the primary language of the text. |
last_modified |
The related file's last modified date. |
parent_id |
The ID of the element's parent element. parent_id might be used to infer where an element resides within the overall document hierarchy. For instance, a NarrativeText element might have a Title element as a parent (a “subtitle”), which in turn might have another Title element as its parent (a "title"). |
text_as_html |
The HTML representation of the related extracted table. Only applicable to table elements. |
Notes on common metadata fields:
parent_id
and category_depth
enhance hierarchy detection to identify the document
structure in various file formats by measuring relative depth of an element within its category. This is especially
useful in files with native hierarchies like HTML or Word files, where elements like headings or list items inherently define structure.
Some file types support location data for the elements, usually in the form of bounding boxes.
The coordinates
metadata field contains:
points
: These specify the corners of the bounding box starting from the top left corner and proceeding counter-clockwise. The points represent pixels, the origin is in the top left and they
coordinate increases in the downward direction.system
: The points have an associated coordinate system. A typical example of a coordinate system isPixelSpace
, which is used for representing the coordinates of images. The coordinate system has a name, orientation, layout width, and layout height.
Field name | Applicable file types | Description |
---|---|---|
page_number |
DOCX, PDF, PPT, XLSX | The related file's page number. |
page_name |
XLSX | The related sheet's name in an Excel file. |
sent_from |
EML | The related email sender. |
sent_to |
EML | The related email recipient. |
subject |
EML | The related email subject. |
attached_to_filename |
MSG | The name of the file that the attached file is attached to. |
header_footer_type |
Word Doc | The pages that a header or footer applies to in a Word document: primary , even_only , and first_page . |
link_urls |
HTML | The URL that is associated with a link in a document. |
link_texts |
HTML | The text that is associated with a link in a document. |
section |
EPUB | The book section title corresponding to a table of contents. |
Here are some notes on additional metadata fields by file type:
Emails will include sent_from
, sent_to
, and subject
metadata. sent_from
is a list of strings because
the RFC 822 spec for emails allows for multiple sent from email addresses.
For Excel files, metadata will contain a page_name
element, which corresponds to the sheet name in the Excel
file.
Headers and footers in Word files include a header_footer_type
indicating which page a header or footer applies to.
Valid values are "primary"
, "even_only"
, and "first_page"
.
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 text_as_html
.
Unstructured will automatically extract all tables for all doc types if you check the Infer Table Structure in the ConnectorSettings area of the Transform section of a workflow.
Here's an example of a table element. The text
of the element will look like this (line breaks are added here for readability):
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 for the same element will look like this (line breaks are added here for readability):
<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>
Documents can include additional file metadata, based on the specified source connector.
date_created
date_modified
date_processed
permissions_data
record_locator
url
version
Source connector | Additional metadata |
---|---|
Azure | protocol , remote_file_path |
Elasticsearch | document_id , index_name , url |
Google Drive | drive_id , file_id |
OneDrive | server_relative_path , user_pname |
S3 | protocol , remote_file_path |
SharePoint | server_path , site_url |
The vision language model (VLM) partitioner also generates an HTML representation of the Unstructured elements that are produced.
Unstructured has developed an element ontology that assigns incoming Unstructured elements to these various defined element ontology types.
These element ontology types are used to generate standard HTML elements with the element ontology type as class attributes on
those HTML elements. The generated HTML elements
are output as text_as_html
along with their parent_id
in metadata
, to allow for easier HTML reconstruction of the entire document as needed.
For example, given the following table element produced with the VLM partitioner, the text_as_html
field is an HTML representation of the
derived table, and parent_id
is the element_id
of the Unstructured element for the page that contains this table. (Line breaks are added here to the
text
and text_as_html
fields for readability.)
{
"type": "Table",
"element_id": "c60aea37616e3db75660918c6d657c38",
"text": "ITEM QUANTITY PRICE TOTAL
Office Desk (Oak wood, 140x70 cm) 2 $249 $498
Ergonomic Chair (Adjustable height & lumbar support) 3 $189 $567
Whiteboard Set (Magnetic, 90x60 cm + 4 markers) 2 $59 $118
SUBTOTAL $1,183
VAT (19%) $224.77
TOTAL $1,407.77",
"metadata": {
"category_depth": 1,
"page_number": 1,
"parent_id": "8cc3b39afcd948d49d85084eaae80ff8",
"text_as_html":
"<table class=\"Table\" id=\"958308a90ccd4fcb825cb12eed20d103\">
<thead>
<tr>
<th>ITEM</th>
<th>QUANTITY</th>
<th>PRICE</th>
<th>TOTAL</th>
</tr>
</thead>
<tbody>
<tr>
<td>Office Desk (Oak wood, 140x70 cm)</td>
<td>2</td>
<td>$249</td>
<td>$498</td>
</tr>
<tr>
<td>Ergonomic Chair (Adjustable height & lumbar support)</td>
<td>3</td>
<td>$189</td>
<td>$567</td>
</tr>
<tr>
<td>Whiteboard Set (Magnetic, 90x60 cm + 4 markers)</td>
<td>2</td>
<td>$59</td>
<td>$118</td>
</tr>
<tr>
<td colspan=\"3\">SUBTOTAL</td>
<td>$1,183</td>
</tr>
<tr>
<td colspan=\"3\">VAT (19%)</td>
<td>$224.77</td>
</tr>
<tr>
<td colspan=\"3\">TOTAL</td>
<td>$1,407.77</td>
</tr>
</tbody>
</table>",
"languages": [
"eng"
],
"filetype": "application/pdf",
"partitioner_type": "vlm_partition",
"filename": "invoice.pdf"
}
}
For the most up-to-date list of available element ontology types, see the ontology.py file, located in the Unstructured-IO/unstructured repository in GitHub.
For the most up-to-date list of mappings between element ontology types and Unstructured element types, see the mappings.py file, located in the Unstructured-IO/unstructured repository in GitHub.