The Will to Fight and Win Against Unfamiliar Typos

SSG.COM's search is the core path responsible for customers reaching products. The existing log-based typo correction had high accuracy, but it was a reactive approach that could only collect data after a search failure occurred, which meant it had the limitation of low coverage.

To overcome this, the team reviewed search term embedding, vector search, and the introduction of a translation model, and ultimately selected the typo-to-correct translation model, which had the highest resource efficiency. To secure data, they used a Data Augmentation strategy that directly generated typos from correctly typed text.

The Hallucination problem that arose during the model introduction process was resolved using the following methods.

• Multitask Learning: The relationships between search terms, product names, and categories were trained together so that contextual information would be reflected.
• Edit Distance Filtering: A safeguard was built to reject results when the edit distance between input and output exceeded a certain threshold.