Improve machine translation in e-commerce multilingual search with contextual signal from search sessions

Over a period of years, search engines have become adept at understanding and providing relevant results for short user generated queries for monolingual search. However, the brevity of search queries can be a limitation for cross-lingual e-commerce search. Previous studies have demonstrated that discourse-level context information can improve machine translation (MT) for document translation but there is no well-defined context regarding MT for query translation. Therefore, in this study, we aim to improve MT for search by incorporating contextual signals from search sessions. Our first step is to explore and categorize two types of contextual queries from search sessions: those with content variations and those with spelling variations. We then propose an innovative approach to derive bilingual training data from search sessions and incorporate the session queries as contextual signals. Using this data, we augment the training data to improve MT. Our initial experimental results demonstrate that augmenting the training data with content variant session queries as context can enhance MT for query translation. Overall, our study provides insights into how contextual information from search sessions can be leveraged to improve machine translation in multilingual e-commerce search.