Google

When using a Large Language Model for classification by reframing the task as text completion, there is no guarantee that the model will generate one of the desired label words (e.g., 'positive', 'negative'). To solve this, a technique can be used to constrain the model's predictions to a predefined set of labels. This method restricts the model's vocabulary to only the valid label words for the task. The final output is then determined by selecting the label from this set that has the highest assigned probability.

Constraining LLM Predictions to a Predefined Label Set

In classification tasks where the goal is to select a single label word, such as filling in a blank, the chosen label is the one that maximizes the conditional probability given the input context $$\mathbf{x}$$. This selection process is formalized by the equation: $$\text{label} = \underset{y \in Y}{\arg\max} \, \text{Pr}(y|\mathbf{x})$$ In this formula, $$y$$ represents a candidate label word, and $$Y$$ is the predefined set of all possible label words. For example, in a polarity classification task, the set of labels could be $$Y = \{\text{positive, negative, neutral}\}$$.

Formula for Label Selection via Probability Maximization

A prompt for polarity classification can be structured by combining a direct instruction, the input text, and a concluding question. This format clearly defines the task, provides the content for analysis, and guides the model toward a specific answer. For example:

'Analyze the polarity of the following text and classify it as positive, negative, or neutral.

The service at the restaurant was slower than expected, which was a bit frustrating.

What is the polarity of the text?'

Example of a Prompt for Polarity Classification (Negative Sentiment)

A simple prompt for polarity classification can be constructed by pairing a direct instruction with the text to be analyzed. The following is a complete example of such a prompt:

'Analyze the polarity of the following text and classify it as positive, negative, or neutral.

The service at the restaurant was slower than expected, which was a bit frustrating.'

Example of a Simple Prompt for Polarity Classification

Describe a technique that could be applied directly to the language model to ensure its output is always one of the three required labels ('Urgent', 'High Priority', 'Standard'). Explain the core mechanism of this technique and why it is more reliable than interpreting the model's descriptive sentences.

Automating Support Ticket Classification

A language model is prompted to classify a movie review as 'Positive' or 'Negative'. In an unconstrained test, the model generates the word 'Fantastic' with a very high probability. If the model's output is then constrained to only the words 'Positive' and 'Negative', explain the specific mechanism the model uses to decide which of these two labels to output.

Mechanism of Constrained Prediction

A developer is using a large language model to classify customer feedback into one of three categories: 'Positive', 'Negative', or 'Neutral'. The model correctly identifies the sentiment but often generates free-form text like 'The customer seems unhappy' instead of the specific label 'Negative'. This inconsistency is causing problems for a data analysis pipeline that expects one of the three exact labels. Which of the following approaches would be the most direct and reliable way to ensure the 

Learn Before

Related