May 2022 – Jul 2022 · 2 mos · Banglore

• Developed a query relevance model using Cross-lingual Language Bert Model Architecture.
• Employed SCOPE to prepare data for training the models.
• Built a query relevance model using the XLM-R architecture, which predicts the relevance of an advertisement
• concerning the user query.
• Utilised 80% of the training data to build several relevance and coherence models using the Bert architecture
• Applied features from these models on 20% of training data to train the XGBoost ensemble model to predict the
• relevancy of an advertisement.
• Determined best features from the trained models and then further improved the scores by using 100% training
• data for training the selected models.
• Scaled the model to work on global markets consisting of about 40 regions worldwide.
• Achieved remarkable gains up to 1.01% in roc-auc score in Global markets with ensemble modeling.
• This work will further improve the current relevance model in Bing-ads, which will then be used in production.