This study addresses the problem of brand entity linking for e-commerce search queries. Entity linking is achieved through either i) a two-step process of entity mention detection and entity disambiguation, or ii) an end-to-end linking approach that directly retrieves target entities given input text. Unique challenges arise from short queries (average 2.4 words), a lack of natural language structure, and the handling of vast unique brand spaces. This study presents a novel end-to-end solution using a two-step approach combining named entity recognition and matching, along with extreme multi-class classification. The solution is validated through offline benchmarks and online A/B testing.
