Skip to content

Latest commit

 

History

History
127 lines (72 loc) · 10.6 KB

README.md

File metadata and controls

127 lines (72 loc) · 10.6 KB
Raw

Extract Record Data

The Extract Record Data workflow translates unstructured text into structured records via schema-aligned JSON objects uploaded or defined by the user.

Select the View example outputs tab (in app) or navigate to example_outputs/extract_record_data (on GitHub) for examples.

How it works

  1. [Input] An instance or collection of unstructured text and (optionally) an existing JSON file containing the JSON schema with which to generate output records.
  2. [Process] The user edits the uploaded JSON schema or creates one interactively.
  3. [AI Calls] The system uses generative AI to extract a JSON object from each text following the JSON schema.
  4. [Output] A dataset of structured records following the JSON schema and (optionally) a newly-defined JSON schema.

Input requirements

  • The input schema, if provided, should be a JSON file conforming to the JSON schema standard and following the restrictions of the OpenAI Structured Outputs API.
  • The unstructured text is sent to the AI API for record extraction (either OpenAI or Azure OpenAI). Such data use must comply with all applicable laws, regulations, and policies, including those pertaining to privacy and security.

Use with other workflows

Extract Record Data can be used to create structured data suitable for input to any of the following workflows:

Tutorial

The task for this tutorial is extracting structured data records from transcripts of customer complaint calls (mock data) available for download either:

From the Extract Record Data homepage in a running instance of Intelligence Toolkit, select Prepare data schema.

Modifying the default schema

In the top left, you have the option to upload an existing JSON schema. This is helpful when the data schema represents an existing data standard or has been saved from previous use of the workflow. In this tutorial though, we will begin by defining a new schema using the Edit data schema form. Any changes made using this form will be reflected in the Preview pane to the right.

The initial JSON schema contains some boilerplate metadata fields representing best practices for schema design. The metadata fields are as follows:

  • $schema: Indicates the version of the json-schema standard that the current schema follows. Leave this field as it is.
  • title: A short title indicating the kind of data that the schema represents.
  • description: A longer description of the kind of data that the schema represents.

Try editing some of these metadata fields now, and see them reflected in the Preview of the JSON schema to the right. In particular, set the title field to Customer Complaints.

The schema in progress is validated after every change, with the message Schema is valid confirming that the current schema conforms to the standard specified in the $schema field.

Try downloading an edited schema using the download button, uploading it via the Upload schema control, then continuing as below.

Creating the record collection

Now select the Sample object tab, and notice how none of these fields are contained in the sample object itself. We can understand this by going back to the JSON schema tab and seeing that the schema is of type object and that the properties of the object are currently empty, indicated by the empty braces {}. Whatever we add to the properties of the top-level object in the schema gets added to the Sample object (and indeed to any objects that conform to the schema).

Let's now add some fields to the object using the buttons under Add top-level field in the form to the left.

To create a dataset of records rather than a single object, the schema needs to contain an object array field. Press the obj[] button to add an object array field at the top level (i.e., level 0). The new field will be given a generic name by default: object_array_1. Rename this to complaint_records and see on the right how this creates an array of objects whose properties you can define next.

Note that all new fields have the Required? checkbox checked by default, placing all field names in the required field of the object. This is a requirement for the OpenAI Structured Outputs API, which we'll later use to generate mock data that follows the schema. Similarly, all objects must also have additionalProperties set to false, so the Additional? checkbox is left unchecked by default.

Defining record attributes

Next, we need to add fields to the objects of complaint_records for each attribute of the records we want to create.

Using the controls under Add field to complaint_records, press the str button to add a string (i.e., text) field. This field appears as the level 1 string string_1 (level 1 because the field is nested one level down from the complaint_records array at the top level, i.e., level 0). Edit the text label from string_1 to name.

As further string fields within complaint_records, now add:

  • street and city as string fields
  • age as a numeric field using the num button
  • email as string field
  • price_issue, quality_issue, service_issue, delivery_issue, description_issue as boolean (true/false) fields using the bool button

Next, we want to add a product_code string field, but limit the possible values of the field to a predefined list called an "enumeration". Do this by checking the Enum? checkbox and observing the default values A, B, and C added to the enumeration. These values can be edited, deleted, and expanded as desired. For this tutorial, simply add further enum values alphabetically from D to H.

Note that boolean attributes of the record could also have been created using the str[] button to create a string array, checking Enum?, and specifying price_issue, quality_issue, service_issue, delivery_issue, description_issue as possible values. However, by using independent boolean fields we simplify the overall record structure and avoid the challenges of nested arrays in the final data object.

Finally, add a quarter string indicating in which calendar quarter the complaint was made. In the description, you can add a hint about the field structure and contents, e.g., "The quarter in which the complaint was made (since 2020-Q1)".

The schema is now complete and can be downloaded using the Download complaint_records_[schema].json button below the schema preview. Selecting the Sample object tab shows a minimal JSON object confirming to the schema.

Extracting structured records

Navigate to the Extract structured records tab to use this schema to extract structured records from input text provided.

With Mode set to Extract from single text, you can enter arbitrary text into the Unstructured text input field as desired. For example, try copying and pasting the following mock call transcript:

**Customer Service Representative:** Good afternoon, thank you for calling our customer service hotline. My name is Sarah. How can I assist you today?

**Bob Johnson:** Hi Sarah, my name is Bob Johnson. I'm calling to report an issue with a product I purchased recently.

**Customer Service Representative:** I'm sorry to hear that, Bob. Could you please provide me with some more details about the issue?

**Bob Johnson:** Sure, I live at 123 Maple Street in Springfield, and I recently bought a product from your company. The product code is A, and I received it in the second quarter of 2023.

**Customer Service Representative:** Thank you for that information, Bob. Could you tell me what specific issue you're experiencing with the product?

**Bob Johnson:** Yes, the problem is with the quality of the product. It just doesn't meet the standards I was expecting.

**Customer Service Representative:** I understand how frustrating that can be. Just to confirm, are there any other issues such as price, service, delivery, or description that you're experiencing?

**Bob Johnson:** No, it's just the quality issue. Everything else, like the price and delivery, was fine.

**Customer Service Representative:** Thank you for clarifying that. I see that you're 36 years old, and I have your email as [email protected]. Is that correct?

**Bob Johnson:** Yes, that's correct.

**Customer Service Representative:** Great, Bob. I will escalate this quality issue to our product team for further investigation. We will get back to you via email with a resolution as soon as possible.

**Bob Johnson:** Thank you, Sarah. I appreciate your help.

**Customer Service Representative:** You're welcome, Bob. Thank you for bringing this to our attention, and we apologize for any inconvenience. Have a great day!

**Bob Johnson:** You too, bye.

Next, try the same extraction process over a table containing unstructured text fields.

From the View example outputs tab, select the customer_complaints example and navigate to the Unstructured texts tab. Download the CSV file, then return to the Extract structured records tab.

Set the Mode to Extract from rows of CSV file and upload the customer_complaints_texts.csv file just downloaded.

Press Extract record data to see the Extracted records generated on the right.