Extraction from PDFs: Extracting balance sheet amounts

In this tutorial, you will use Snorkel Flow to extract balance sheet amounts from the SEC filings of public companies.

Snorkel Flow supports the extraction of structured information from semi-structured documents such as PDFs, HTML, and docx files. Common use cases extract information from financial filings, insurance claims, and medical reports. These documents are converted to a special format called "Rich Doc" in the platform. This conversion allows you to leverage text, layout, and image modalities to write labeling functions (LFs) and train models.

In this tutorial, you will learn how to:

1. Upload data into Snorkel Flow
2. Create an application
3. Add ground truth
4. Define labeling functions
5. Train a model

Upload data

Snorkel Flow can ingest data from a range of storage options, including cloud storage, databases, or local files. For PDF extraction applications, the input data must include the following fields:

Field name	Data type	Description
uid	int

A unique id mapped to each row.

This is standard across all Snorkel Flow datasources.

| | url | str |

The file path of the original PDF.

This will be used to access and process the PDF.

|

For this tutorial, use the provided dataset from in AWS S3 Cloud Storage. Start by creating your new dataset.

1. Select the Datasets option from the left menu.

2. Select the + Upload new dataset button on the top left corner of your screen.

3. In the New Dataset window, enter the following information:

   • Dataset name: Enter balance-sheets-dataset.

   • Select a data source: Select Cloud Storage as your data source.

   • File path and Split: Add these datasets:

     File path	Split
     s3://snorkel-native-pdf-sample-dataset/splits/train.csv	train
     s3://snorkel-native-pdf-sample-dataset/splits/valid.csv	valid
     s3://snorkel-native-pdf-sample-dataset/splits/test.csv	test

4. Select Verify data source(s) to run data quality checks to ensure data is cleaned and ready to upload.

5. After clicking Verify data source(s), select uid from the UID column dropdown box. The uid is the unique entry ID column in the data sources.

6. Select Add data source(s).

Once the data is ingested from all three data sources, you can create an application with your PDF extraction dataset. For more information about uploading data, see Data upload.

Create an application

There are two categories of PDF documents:

• Native PDFs are documents that were created digitally. These PDFs can be parsed without additional processing.
• Scanned PDFs are created from scans of printed documents. These PDFs don't have the metadata that we need to parse layout information. Scanned PDFs require additional preprocessing with an optical character recognition (OCR) library.

The application creation process is slightly different for these two formats. This tutorial uses native PDFs.

You will extract the numbers from the Consolidated Balance Sheets provided in public company filings, and classify these numbers into the following classes: ASSETS, LIABILITIES and EQUITY.

1. Select the Applications option in the left-side menu and select Create application to create a new application. Enter the values provided in the table:

   Stage	Field	Value
   Start	Application name	line-item-classification
   Data	Dataset	balance-sheets-dataset
   Label schema	Data type	PDF
   Task type	Extraction
   PDF type	Native PDF
   PDF URL field	url

   You selected the dataset created in the previous step, defined the data and task type, and specified where the PDFs can be found (url).

2. Select Generate Preview to generate and display a preview sample on the right from the input documents.

3. Edit the Label schema table:

4. 1. Select Add new label and add ASSETS. Next, add LIABILITIES and EQUITY to the table as new labels. These are the entities to extract.
   2. Edit the NEGATIVE label and rename it to OTHER. Any data point that does not fall into one of the positive classes will be labeled as OTHER.
      Do not edit the UNKNOWN label. This is used for unlabelled data.
   3. Select Next.

5. In the Preprocessors pane, select the following preprocessing operations to perform on the dataset:

   • Split docs into pages: Yes
   • Page split method: PageSplitter
   • PageSplitter window_size: 0
   • Extraction method: NumericSpanExtractor from an existing extractor
   • NumericSpanExtractor field: rich_doc_text

6. Select Commit for the PageSplitter and the NumericSpanExtractor operators.
   Operators perform transformations on your input data. You can add other operators in the background to pre-compute useful features to help define labeling functions. For more information on these operators, see PDF-based operators. The PageSplitter and the NumericSpanExtractor transformations are visualized in the Preview sample pane on the right side of your screen.

   • The PageSplitter operator splits the documents into pages. This split helps us decrease memory usage, retrieve metadata more efficiently, and improve latency.
   • The NumericSpanExtractor operator is used to extract all numeric values as candidate spans from the raw text. The Preview sample pane highlights the numeric values.

7. Click Next to start a job that runs the preprocessing operators on the entire dataset.

8. In Development Settings, select Yes using span_text to Compute embeddings.

9. Select Go To Studio.

To set up a Scanned PDF application, you would select the Scanned PDF, no need to run OCR option for PDF type. For more information about preprocessing scanned PDFs, see Scanned PDF Guide.

Add ground truth

Snorkel Flow allows you to label your data with programmatic labels. First, you must establish ground truth with annotations to validate the performance. When you start a new project, you can annotate your data using our Annotation Studio. For this tutorial, you will annotated data.

1. Select Develop to go to the Overview page.
2. Select View Data Sources to view the data sources that we have added. In this view, you can see the total number of data points that are extracted for each data source.
3. Select Upload GTs.
4. Enter the following values:
   • File path: s3://snorkel-native-pdf-sample-dataset/native_pdf_ground_truth.csv
   • File format: CSV
   • Label column: label
   • UID column: x_uid
5. Select Add.
6. Refresh the page after uploading the ground truth.
   The # of GT labels column has non-zero values.
7. Click Go To Studio to return to the Studio page.

Review data in Studio

Snorkel Flow extracts candidate spans from the raw text of the document using a heuristic. Common span types include numbers, dates, and email addresses. You can define labeling functions to assign labels to the candidate spans in the document. The labels are aggregated and used to train a classification model. The spans are highlighted with bounding boxes and color-coded by label.

1. Use your cursor to hover over any word in the document. You can see the bounding box coordinates of the word as measured in pixels.
2. Select Rich Doc toggle to see the raw text extracted from the document.
3. Select the arrow to expand Highlight regions. This allows you to highlight the bounding boxes of the different regions in the document: row, word, line, paragraph, and area.
   Studio shows the dev split. The dev split is a sample of the train split used for iterating on labeling functions.
4. Select the Dev split dropdown in the top toolbar.
5. Select Resample data.
6. Set the Sample size to 4, and select Resample dev split.

Next, you'll define your labeling functions.

Define labeling functions

A labeling function (LF) is a programmatic rule or heuristic that assigns labels to unlabeled data. LFs are the key data programming abstraction in Snorkel Flow. Each labelling function votes on whether a data point has a particular class label. The Snorkel Flow label model is responsible for estimating LF accuracies and aggregating them into training labels.

For PDF extraction applications, we add a few text fields that are useful for defining LFs using the RichDocSpanRowFeaturesPreprocessor operator. These fields are based on heuristics as defined below:

Field	Description
rich_doc_row_text_inline	text in the same row as the span
rich_doc_row_header	text that is the furthest to the left of the span
rich_doc_inferred_row_headers	text that is above the span and indented to the left
rich_doc_row_text_before	text in the row before the span
rich_doc_row_text_after	text in the row after the span

You can use these columns to define text-based LFs. You can also use the word bounding boxes (top, left, bottom and right) to define location-based LFs.

1. Create new labeling functions.
   1. In Patterns, select your LF template by typing /, followed by the pattern, such as Keyword, Regex, or Numeric.
   2. Enter the requirements for the labeling function.
   3. Select a label from the dropdown menu.
   4. Select Preview LF to see the precision and coverage of your LF.
   5. Select Create LF.
      The new LF shows up in the Labeling Functions pane.

Using the steps above, add these LFs to your application:

LF Template	Settings	Label	Explanation
Keyword	rich_doc_inferred_row_headers [CONTAINS] Assets, Cash, Land	ASSETS	The LF checks if the inferred row header of the span contains the keywords Assets, Cash or Land.
Keyword	rich_doc_row_text_inline [CONTAINS] Liabilities, Payable, Debt	LIABILITIES	The LF checks if the text in the same as the span contain the keywords Liabilities, Payable or Debt.
Regex	rich_doc_row_header matches pattern Total.{1,15}equity	EQUITY	The LF checks if the row header of the span matches the regex pattern.
Numeric	right [<=] 1200	OTHER	The LF checks if the span's right boundary is less than or equal to 1200 pixels i.e. if the span is on left side of the page.

Some LF templates are available for only PDF extraction applications. These LFs allow us to combine the text and layout information to define LFs.

Using the steps above, add these LFs to your application:

LF Template	Settings	Label	Explanation
Span Regex Proximity Builder	If the span is up to [6] [LINE(s)] [AFTER] the regex pattern current assets:	ASSETS	This LF specifies that spans up to 6 lines after the expression “current assets:” will be labeled ASSETS.
Span Regex Row Builder	If the span is [0] rows before and [5] rows after the regex pattern current liabilities:	LIABILITIES	This LF specifies that spans up to 5 rows after the expression “current liabilities:” will be labeled LIABILITIES.
Rich Doc Expression Builder	Evaluate Common stock with SPAN.left > PATTERN1.right and SPAN.top >= PATTERN1.top	EQUITY	This LF specifies that spans to the right and below the expression “Common stock” will be labeled as EQUITY.

For more information on Rich Doc builders, see Rich document LF builders.

You can also encode your custom labeling function logic using the Python SDK. For examples using the RichDocWrapper object, refer to the SDK reference.

Train a model

You can aggregate the newly defined labeling functions to train a model.

1. In the Models pane, select Train a model.
2. For the Model architecture, select Logistic Regression (TFIDF).
3. To include the text fields used to define labeling functions, select Input Fields and select rich_doc_inferred_row_headers, rich_doc_row_text_inline, rich_doc_row_header, and right from the dropdown.
4. Leave the default settings in Train options.
5. Select Train custom Model to generate labels using the LFs and train a logistic regression model using the labels.

Once the model training is completed, you will see your first model in the Models pane.

Snorkel Flow offers default configurations for several commonly used models. For more information on models, see Model training.

Snorkel Flow enables users to iteratively generate better labels and better end models. For information about getting metrics and improving model performance, see Analysis: Rinse and repeat.

Conclusion

You have set up and iterated on a PDF extraction application in Snorkel Flow. To learning more about PDF extraction features in Snorkel, see this documentation:

• Learn how to work with scanned documents with our OCR guide.
• Read more about PDF-based Operators.
• Check out more examples of Rich Document Builders.

The Snorkel Flow Python SDK provides greater flexibility to define custom operations. Select Notebook on the left-side pane to access the SDK. See the SDK reference to learn more about custom operators and custom labeling functions.