Labels

parent
summarization

Related:


Overview:

There are many ways to get the OpenAI API to generate labels. It can get some out of the bag with no prompts, others it can get with a bit of prompting. Some of the current difficulties with keywords involve named entity recognition and terms that the API has not seen before (e.g. recent news events). Providing context in the prompt can improve performance.

[EOP] indicates the end of prompt

Zero Shot Labels

List of known zero shot labels:

- Keywords
- Theme
- State

Keywords Label

Just appending "Keywords:" after a sentence will result in keywords.

Sentence: "The rain in Spain falls mainly on the plain"
Keywords:[EOP] "rain" and "Spain"

Theme Label

Just appending "Theme:" after a sentence will result in a theme.

Sentence: "The rain in Spain falls mainly on the plain"
Theme:[EOP] The weather in Spain


Few Shot: More complicated labeling

Places

Let's say we want to label places instead. Just trying 0 shot won't work. For instance, trying "places" gets the whole sentence back.

Sentence: "The rain in Spain falls mainly on the plain"
Places:[EOP] "The rain in Spain falls mainly on the plain"

Giving an example can fix that.

Sentence: "The king of France lives in Paris"
Places: "France", "Paris"
Sentence: "The rain in Spain falls mainly on the plain"
Places:[EOP] "Spain", "plain"

Multiple Labels

OpenAI API can be primed to apply multiple labels to text. For instance, it can be trained to provide both people and places.

Sentence: "Humpty Dumpty had a great fall"
People: Mr. Dumpty; Actions: Fall

Sentence: "The itsy bitsy spider crawled up the water spout"
People:[EOP] Spider; Actions: Crawl

Contextual Information

Let's say there's novel or ambiguous information the model doesn't have. For instance, suppose we have a jaguar car but OpenAI doesn't know.

Sentence: "I love Taco Bell"
Category: Food

Sentence: "I love my jaguar"
Category:[EOP] Animals

Adding additional prompts with that information can prime the API to return the correct category. For instance, we can include relevant information that we replaced our Toyota Corolla.

Sentence: "I love Taco Bell"
Category: Food

Sentence: "I upgraded my old Toyota last week"
Category:Shopping

Sentence: "I love my jaguar"
Category:[EOP] Cars

You can use the API's Search functionality to automate finding contextual primes.

Additional Resources:

There's probably other resources here, maybe can make children or something.

Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-ShareAlike 3.0 License