Improving the Performance of Pre-trained Systems in Sentence Retrieval

Information Retrieval is the task of satisfying an information need by retrieving relevant information from large collections. Recently, contextualised pre-trained systems have surpassed traditional systems on benchmarks that provided vast amounts of context. In contrast, pre-trained systems underperform in environments where context is insufficient. Hence, this work aims to improve the performance of pre-trained systems under such resource constraints. To this aim, we turned to context augmentation. Since sentences typically consist of few terms, we recognised sentence retrieval as a context-constrained environment. However, most context augmentation methods were either ineffective in sentence retrieval or computationally expensive, while those methods that were effective in sentence retrieval and computationally cheap were unable to capture contextual information. Hence, we developed a novel context augmentation method for sentence retrieval that preserves contextual information while maintaining low latency. The proposed method allowed the pre-trained system to achieve a 19% performance gain in a context-constrained environment. Its effectiveness was also validated by participation in the EPIC-QA challenge. Specifically, it allowed the pre-trained system to rank third among 55 competing systems, with less than a 1% performance gap to first place.