How Samsung Uses Two AI Models to Improve Sentence Generation

A method for training an AI to write better sentences by using a second model that reads or generates text in a different order to grade the first model's work.

What it covers

This patent describes a way to train a machine learning model to generate text more accurately. It uses two 'decoding models'—essentially the parts of an AI that pick the next word in a sentence. The first model generates a target sentence, while the second model acts as a critic by generating words in a different order (such as backward) to calculate 'reward information.' This reward information acts like a score, telling the first model how good its sentence was, which is then used to update the first model's internal weights to improve future performance.

What it doesn't cover

• —Does not cover simple text generation that lacks a secondary, order-reversed validation model.
• —Does not cover training methods that rely solely on human-provided labels or static datasets.
• —Does not cover architectures that do not use neural network weight resetting as the primary training mechanism.

The clever bit

By forcing the second model to process or generate words in a different order (like backward), the system gains a structural check on the first model's output, helping to catch errors that a forward-only model might consistently repeat.

Why it matters

As AI-driven text generation becomes standard in consumer electronics, improving the quality of machine-generated language without massive amounts of human-labeled data is a major technical hurdle. This approach helps Samsung refine its virtual assistants and automated translation tools by creating a self-improving loop between two different AI perspectives.

Real-world examples

• 1.Samsung Bixby voice assistant responses
• 2.Automated machine translation software
• 3.Predictive text and smart reply features in mobile operating systems