Processing unstructured data to structured data

Text Analysis with large language models

Emi Tanaka

4th June 2025

These slides are best viewed on a modern browser like Google Chrome on a desktop or laptop. You can navigate the slides using the left and right arrows.

Unstructured data

Unstructured data lacks a predefined schema and can be highly variable in form.
Some examples of unstructured data include:
- Texts (e.g., emails, articles, social media posts, support tickets)
- Images and videos (e.g., photos, video recordings)
- Audio files (e.g., podcasts, voice recordings)
- Sensor data (e.g., IoT devices)

Processing to structured data

Unstructured data is hard to query, analyse or automate without first processing into structured data.
Processing unstructured data to structured data can be done using various methods, including:
- Text extraction and parsing:
  - Rule-based system (e.g. regular expression)
  - Manual labelling
- Image recognition and classification
- Audio transcription

LLMs as text extractor

Trained on a diverse corpora and has a general “understanding” of natural language

library(ellmer)
chat <- chat_openai(model = "gpt-4.1",
                    system_prompt = '
Extract the following fields from the text:
- Customer name
- Complaint category
- Date
- Product

Text: "Hi, I\'m Alex. My AC Model X-200 failed yesterday."

Output:
{ "name": "Alex", "category": "product failure", "date": "yesterday", "product": "AC Model X-200" }')

From one shot training above, we have

chat$chat("Hey there, I'm Emi and my apple watch malfunctioned today.")

{ "name": "Emi", "category": "product malfunction", "date": "today", "product":
"apple watch" }

Structured chat with LLM

Prompt engineering for structured data, however, doesn’t guarantee the adherance to the desired schema.
Certain LLM has been trained specifically for structured chat output:
- AI models starting with GPT-4o allows this by specifying the JSON schema
- Ollama also started to support this from 6th December 2024 but not as good as the Open AI models.
In the first instance, you have to define the JSON schema for the output.

JSON schema

JSON is one of the most popular format for exchanging data

{
    "name": "user_data",
    "strict": true,
    "schema": {
        "type": "object",
        "properties": {
            "name": {
                "type": "string",
                "description": "The name of the user"
            },
            "username": {
                "type": "string",
                "description": "The username of the user. Must start with @"
            },
            "email": {
                "type": "string",
                "description": "The email of the user"
            }
        },
        "additionalProperties": false,
        "required": [
            "name", "username", "email"
        ]
    }
}

JSON type	R class
“string”	“character”
“number”	“numeric”
“boolean”	“logical”
“integer”	“integer”
“enum”	“factor”
“object”	“list”
“array”	“vector” or “data.frame”

JSON schema with `ellmer`

In R, you can write the schema as a list (it will convert to JSON object for you).
The list has to have a particular structure.
In the top level, you need to create the list using type_object().
In the schema, you’ll need to specify:
- the names of the elements,
- their types (e.g. type_number for a numerical output),
- whether its’ strictly required or not, and
- (optionally) the descriptions of the elements.

Extracting structured output with `ellmer`

schema <- type_object(name = type_string("Customer name"),
                      category = type_enum("Complaint category",
                                           values = c("product failure", "other")),
                      date = type_string("Date of the issue"),
                      product = type_string("Product name"))

chat <- chat_openai(model = "gpt-4.1")
chat$chat_structured("Hey there, I'm Emi and my apple watch malfunctioned today.", 
                     type = schema)

$name
[1] "Emi"

$category
[1] "product failure"

$date
[1] "2024-06-08"

$product
[1] "Apple Watch"

Exercise 5

Go to Exercise 5

Next slide

Back to start