Название: Information Retrieval: Advanced Topics and Techniques
Автор: Omar Alonso, Ricardo Baeza-Yates
Издательство: ACM Books
Год: 2025
Страниц: 838
Язык: английский
Формат: pdf (true), epub
Размер: 41.0 MB

In the last decade, Deep Learning and word embeddings have made significant impacts on information retrieval (IR) by adding techniques based in neural networks and language models. At the same time, certain search modalities such as neural IR and conversational search have become more popular. This book, written by international academic and industry experts, brings the field up to date with detailed discussions of these new approaches and techniques. The book is organized in three sections: Foundations, Adaptations and Concerns, and Verticals.

Under Foundations, we address topics that form the basic structure of any modern IR system, including recommender systems. These new techniques are developed to augment indexing, retrieval, and ranking. Neural IR, recommender systems, evaluation, query-driven functionality, and knowledge graphs are covered in this section.

IR systems need to adapt to specific user characteristics and preferences, and techniques that were considered too niche a few years ago are now a matter of system design consideration. The Adaptations and Concerns section covers the following topics: conversational search, cross-language retrieval, temporal extraction and retrieval, bias in retrieval systems, and privacy in search.

The development of recommendation algorithms has naturally mirrored the evolution of the task definition, hand in hand with the design of evaluation procedures and metrics suited to the task. Recommendation can be addressed, in essence, as a supervised learning problem: given examples of observed user choices, we aim to predict present or future (yet unobserved) user interests. Variations in the task formulation give rise to different algorithmic approaches—and different metrics are appropriate to evaluate for different tasks. As a Machine Learning problem, recommendation is quite unique. What makes recommendation singular in this field is, in essence, the human factor at the core of recommendation tasks. In these tasks, both the input signal and the prediction target consist of or involve user behavior at their core. This brings about a specific level of complexity compared to, for instance, recognizing shapes in an image or diagnosing a medical condition from medical tests. Furthermore, recommendation is often not just about predicting people’s actions but about enhancing (and hence changing) such actions by bringing awareness about potentially better choices.

While web search engines are the most popular information access point, there are cases where specific verticals provide a better experience in terms of content and relevance. The Verticals section describes eCommerce, professional search, personal collections, music retrieval, and biomedicine as examples.