Existing information retrieval systems mostly optimize for relevance to the question, ignoring diversity. This work proposes a benchmark and task to evaluate the ability of retrieval systems to surface diverse perspectives on complex and contentious questions.
A novel two-tower model with a unified embedding-based product encoder and joint user-query encoder to handle diverse search queries and provide personalized product recommendations on an e-commerce platform.
Using large language models (LLMs) to generate relevance judgments for information retrieval evaluation is problematic, as it limits the ability to measure systems that may outperform the LLM-generated judgments.
Personalization can significantly improve user engagement in short-video search by leveraging user profiles, long-term interests, and real-time behaviors to retrieve and rank relevant content.
Effective retrieval strategies are crucial for Retrieval-Augmented Generation (RAG) systems to provide accurate and relevant responses. This study evaluates the performance of different document splitting methods and retrieval techniques across diverse document types, including textbooks, articles, and novels, to identify optimal approaches for enhancing retrieval accuracy and efficiency.
The proposed counterfactual framework can identify the terms that need to be added to a document to improve its ranking with respect to a specific retrieval model and query.
MessIRve is a large-scale Spanish information retrieval dataset that accounts for the diverse dialects and topics across Spanish-speaking countries, aiming to advance Spanish IR research and improve information access for Spanish speakers.
Rs4rs is a web application that enables semantic search to efficiently find recent, high-quality publications from top conferences and journals related to Recommender Systems.
QueryBuilder, an interactive system, allows novice users to efficiently create fine-grained queries for cross-lingual information retrieval by leveraging an English development corpus and a combination of probabilistic and neural information retrieval models.
Fusion-in-T5 (FiT5) is a unified document ranking model that integrates text matching, ranking features, and global document information through attention fusion, outperforming complex cascade ranking pipelines.