Google

Few-shot learning is a versatile technique applied to many Natural Language Processing (NLP) tasks. By providing task-formatted demonstrations within a prompt, models can perform tasks like sentiment sentence classification or phrase translation, such as translating phrases from Chinese to English.

Examples of Few-Shot Learning Applications in NLP

An example of an instruction for a few-shot learning prompt is to direct the model to perform sentiment classification on text snippets, specifying the possible labels such as 'Positive', 'Negative', or 'Neutral'.

Example of a Sentiment Classification Task Instruction

A few-shot learning prompt can utilize a simple pattern to guide a model in a translation task. For instance, to translate Chinese words into English, several demonstrations are provided to establish the input-output format, followed by a new user query for the model to process. An example of such a prompt is:

**DEMO**
现在 → now
来 → come
去 → go
男孩 → boy

**USER**
女孩 →

Example of a Few-Shot Prompt for Chinese-to-English Translation

A specific instance of a text snippet that expresses a positive sentiment is 'I had an amazing day at the park!'. This type of snippet is used as an input for sentiment classification tasks.

Example of a Positive Sentiment Text Snippet for Classification

A developer needs to use a language model to extract the main company name mentioned in a business news headline. They want the model to learn this task from just a few examples. Based on the provided examples, construct a complete text prompt that first demonstrates the task and then asks the model to process a new, unseen headline.

Designing a Prompt for Information Extraction

A language model is given the following text as a single input. Analyze the pattern established by the examples and determine what the model's most likely output will be for the final line. Explain your reasoning.

`apple -> red`
`banana -> yellow`
`lime -> green`
`blueberry ->`

Predicting Model Output from a Pattern

A user provides a language model with the text below. By analyzing the structure and the relationship between the inputs and outputs, determine the specific task the user is instructing the model to perform.

`Review: This movie was a masterpiece. The acting was superb.`
`Sentiment: Positive`

`Review: I was really disappointed with the plot.`
`Sentiment: Negative`

`Review: The film was okay, but not memorable.`
`Sentiment: Neutral`

Learn Before

Related