outline: deep

Zero-Shot Classification

Zero-shot classification stands out in natural language processing (NLP) by allowing models to classify text into categories they haven't explicitly been trained on. Unlike traditional text classification, zero-shot classification leverages natural language understanding, enabling models to intuitively categorize texts based on a set of candidate labels provided at runtime.

Task ID

• zero-shot-classification

Default Model

• Xenova/distilbert-base-uncased-mnli

Use Cases

Zero-shot classification's versatility opens up a broad spectrum of applications, including but not limited to:

• Content Categorization: Automatically categorizing articles or posts into themes or genres not seen during training.
• Intent Detection: Identifying the intent behind user queries in chatbots or search engines, even for new, unseen intents.
• Product Tagging: Classifying products or services into categories or tags that were not available at the time of model training.
• Sentiment Analysis: Determining sentiments or opinions in texts towards entities or topics that were not predefined.

Running a Pipeline Session

The zero-shot classification pipeline requires two primary inputs: the text to classify and an array of candidate labels. Here's an example:

use function Codewithkyrian\Transformers\Pipelines\pipeline;

$classifier = pipeline('zero-shot-classification', 'Xenova/mobilebert-uncased-mnli');
$result = $classifier('Who are you voting for in 2020?', ['politics', 'public health', 'economics', 'elections']);

Pipeline Input Options

When running the zero-shot-classification pipeline, you can the following options:

• texts (string)

The piece of text you want to categorize. There's no need to provide it as a named argument. It's always going to be the first input

• candidateLabels (string[])

  An array of strings representing the labels among which the model will classify the text. There's also no need to provide it as a named argument. It's always going to be the second argument, and it's required.

• multiLabel (bool)

  [Optional] Indicates whether the text can belong to multiple labels simultaneously. Defaults to false, where the model assigns the highest score to the most likely label.

  $classifier = pipeline('zero-shot-classification', 'Xenova/mobilebert-uncased-mnli');
  $result = $classifier(
    'I have a problem with my iphone that needs to be resolved asap!',
    ['urgent', 'not urgent', 'phone', 'tablet', 'computer'],
    multiLabel: true
  );
  

• hypothesisTemplate (string)

  A template used to frame the labels for the model before inference, defaulting to "This example is {}.". Adjust this if the default template does not fit your use case.

  $classifier = pipeline('zero-shot-classification', 'Xenova/mobilebert-uncased-mnli');
  $result = $classifier(
    'My favorite fruit is the apple.',
    ['fruit', 'vegetable', 'meat'],
    hypothesisTemplate: "The topic of this text is about {}."
  );
  

Pipeline Outputs

The output includes the original text (sequence), an array of labels, and their corresponding confidence scores (scores). These scores indicate the model's confidence in associating each label with the text, with a range from 0 to 1.

The model's ability to assign multiple labels (when multiLabel is true) significantly enhances its flexibility, allowing for more nuanced text classification. This feature is particularly useful in scenarios where texts inherently belong to multiple categories. Let's compare the outputs of the earlier example

Without multiLabel:

[
  "sequence" => "I have a problem with my iphone that needs to be resolved asap!",
  "labels" => ["urgent", "phone", "computer", "tablet", "not urgent"],
  "scores" => [0.9992283160322, 0.042726408774879,  0.022445628181903, 0.0039340428157268, 0.0016656041952894]
]

With multiLabel:

[
  "sequence" => "I have a problem with my iphone that needs to be resolved asap!",
  "labels" => ["urgent", "phone", "computer", "tablet", "not urgent"],
  "scores" => [0.99588709563603, 0.9923963400697,  0.0023335396113424, 0.0015134149376, 0.0010699384208377]
]

As you can see, it was able to correctly pick up the phone category as well.