Improve retrieval of regional titles in streaming services with dense retrieval

Customers search for movie and series titles released across the world on streaming services like primevideo.com (PV), netflix.com (Netflix). In non-English speaking countries like India, Nepal and many others, the regional titles are transliterated from native language to English and are being searched in English. Given that there can be multiple transliterations possible for almost all the titles, searching for a regional title can be a very frustrating customer experience if these nuances are not handled correctly by the search system. Typing errors make the problem even more challenging. Streaming services uses spell correction and auto-suggestions/autocomplete features to address this issue up to certain extent. Autosuggest fails when user searches keywords not in scope of the auto-suggest. Spell correction is effective at correcting common typing errors but as these titles doesn’t follow strict grammar rules and new titles constantly added to the catalog, spell correction have limited success.

With recent progress in deep learning (DL), embedding vectors based dense retrieval is being used extensively to retrieve semantically relevant documents for a given query. In this work, we have used dense retrieval to address the noise introduced by transliteration variations and typing errors to improve retrieval of regional media titles. In the absent of any relevant dataset to test our hypothesis, we created a new dataset of 40K query title pairs from PV search logs. We also created a baseline by bench-marking PV’s performance on test data. We present an extensive study on the impact of 1. pre-training, 2. data augmentation, 3. positive to negative sample ratio, and 4. choice of loss function on retrieval performance. Our best model has shown 51.24% improvement in Recall@16 over PV baseline.