Abstract
Abstract
Efficient retrieval of relevant documents from massive
collections remains an essential challenge in Information Retrieval
(IR). Modern search engines face immense computational demands,
requiring novel approaches that reduce resource usage without
sacrificing retrieval effectiveness. This thesis makes significant
contributions to improving search efficiency through three distinct
methods. First, we introduce innovative document reordering
techniques specifically optimized for dynamic pruning algorithms.
Our proposed methods achieve up to 1.33x speed-up in query
processing, accompanied by negligible increases in index size and
minimal impact on retrieval quality. Second, we present novel sparse
centroid retrieval strategies tailored to the ColBERT neural
retrieval model. These techniques accelerate ColBERT-based retrieval
by up to 4.6x while maintaining high effectiveness and minimal
additional indexing overhead. Lastly, we propose novel static
pruning methods for ColBERT document embeddings that eliminate
approximately one-third of the tokens from indexed documents without
any loss in retrieval effectiveness. Critically, our pruning methods
require no separate training stages, ensuring ease of integration
into existing retrieval systems. Collectively, these contributions
offer substantial advancements in retrieval efficiency, making
large-scale IR systems faster, more scalable, and economically
sustainable.