Google

The $$\text{Source Text} \to \text{Target Text}$$ format is a standard structure used to frame natural language processing tasks. In this format, an arrow ($$\to$$) separates the input from the expected output. The source text serves as the input to the system, containing a task instruction or description alongside the specific input data. The target text is the desired response generated for that input task. For example, a machine translation task, such as converting Chinese to English, can be framed using this consistent structure.

$$\text{Source Text} \to \text{Target Text}$$ Format

The T5 model utilizes a unified text-to-text format for all training samples, structured as $$\text{Source Text} \to \text{Target Text}$$. Within this structure, the source text provides the system with both a task description or instruction and the specific input data. The target text represents the expected response to the input task. By maintaining this consistent format, the model is able to frame and process many different problems as the same text-to-text task.

T5 Sample Format

A machine learning team is training a single model to handle multiple types of text manipulation tasks. They are finding that the model performs poorly, often generating outputs that do not match the intended task. Below are two examples from their training dataset. After reviewing the examples, identify the fundamental flaw in how the 'Source Text → Target Text' format has been applied and explain why this flaw would lead to the observed poor performance.

Diagnosing a Model Training Issue

Match each Natural Language Processing (NLP) task with its correctly formatted 'Source Text → Target Text' representation.

A training sample for a translation task illustrates the sequence-to-sequence structure by distinguishing the source text from the target text. For instance, in a task translating from Chinese to English, a training sample can be expressed as: $$[\mathrm{CLS}]\ \textrm{Translate from Chinese to English: 你好！} \to \langle s \rangle\ \textrm{Hello!}$$. Here, the arrow ($$\to$$) separates the source text, which encompasses the task instruction and the input, from the target text, representing the system's response. Additionally, $$[\mathrm{CLS}]$$ and $$\langle s \rangle$$ designate the start symbols on the source and target sides, respectively.

Example of a Translation Training Sample

A developer wants to frame a sentiment analysis task for a text-processing system. The goal is to classify the sentence 'The movie was fantastic!' as 'positive'. Based on the standard 'Source Text → Target Text' structure, which of the following options correctly formats this task?

Learn Before

Related