MSc Thesis - Geographical Question Answering Leveraging Neural Language Models for Passage Retrieval
This work dealt with geographical passage retrieval. Summary: - Geoparsed large dataset (MS-MARCO). - Studied state-of-the art approaches for neural passage retrieval (cross-encoders and bi-encoders). - Proposed an hard negative sampling technique based on distance between geographic entities within text. - Used a pre-trained T5 model for query generation. - Proposed a cross-architecture knowledge distillation leveraging fully differentable soft-ranking functions.
This project aimed to study novel approaches for mobile application search and recommendation. My tasks: - Scrapped a novel dataset containing metadata on mobile applications. - Conducted data cleaning and analysis. - Built a semantic search engine based on neural language models, leveraging a self-supervised training technique. - Studied the impact of said model for More Like This recommendation. - Created and deployed APIs to serve both the search engine and the recommendation system. - Conducted two rounds of user-centered tests.