Context-aware ranking refinement with attentive semi-supervised autoencoders

Learning to rank methods aim to learn a refined ranking model from labeled data for desired ranking performance. However, the learned model may not improve the performance on each individual query because the distributions of relevant documents among queries are diversified in document feature space...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Soft computing (Berlin, Germany) Jg. 26; H. 24; S. 13941 - 13952
Hauptverfasser: Xu, Bo, Lin, Hongfei, Lin, Yuan, Xu, Kan
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Berlin/Heidelberg Springer Berlin Heidelberg 01.12.2022
Schlagworte:
Application of Soft Computing
Artificial Intelligence
Computational Intelligence
Control
Engineering
Mathematical Logic and Foundations
Mechatronics
Robotics
Pseudo-relevance feedback
Information retrieval
Learning to rank
Ranking refinement
Machine learning
ISSN:1432-7643, 1433-7479
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Learning to rank methods aim to learn a refined ranking model from labeled data for desired ranking performance. However, the learned model may not improve the performance on each individual query because the distributions of relevant documents among queries are diversified in document feature space. The performance of learned ranking models may be largely affected by the usefulness of document features. To generate high-quality document ranking features, we capture the local context information of individual queries from the top-ranked documents of an initial retrieval using pseudo-relevance feedback. Based on the top-ranked feedback documents, we propose an attentive semi-supervised autoencoder to refine the ranked results using an optimized ranking-oriented reconstruction loss. Furthermore, we devise the hybrid listwise query constraints to capture the characteristics of relevant documents for different queries. We evaluate the proposed ranking model on LETOR collections including OHSUMED, MQ2007 and MQ2008. Our model produces better experimental results and consistent improvements of ranking performance over baseline methods.
ISSN:1432-7643
1433-7479
DOI:10.1007/s00500-022-07433-w