Local or Global? A Comparative Study on Applications of Embedding Models for Information Retrieval
Dwaipayan Roy, Mandar Mitra, Philipp Mayr, Amritap Chowdhury · 2022
Application of embedding techniques for improving the performance of different text processing tasks, including information retrieval (IR) has been a much sought after research area for the last few years with decisive improvements over state-of-the-art. Apart from applying embedding techniques for different tasks, significant works have been conducted on setting the best practice. However, an optimal setting to apply embeddings is still an ongoing research area in the text processing community. In this study, we have compared ways of applying embeddings for downstream query expansion task for IR and have discussed some effective variations of locally selecting embedded vectors. Experimentation on TREC news and web corpora empirically validates the ascendancy of the variation over the benchmark settings.