Deep learning models named transformers achieved state-of-the-art results in a vast majority of NLP tasks at the cost of increased computational complexity and high memory consumption. Using the transformer model in real-time inference becomes a major challenge when implemented in production, because it requires expensive computational resources. The more executions of a transformer are needed the lower the overall throughput is, and switching to the smaller encoders leads to the decrease of accuracy. Our paper is devoted to the problem of how to choose the right architecture for the ranking step of the information retrieval pipeline, so that the number of required calls of transformer encoder is minimal with the maximum achievable quality of ranking. We investigated several late-interaction models such as Colbert and Poly-encoder architectures along with their modifications. Also, we took care of the memory footprint of the search index and tried to apply the learning-to-hash method to binarize the output vectors from the transformer encoders. The results of the evaluation are provided using TREC 2019-2021 and MS Marco dev datasets.
View on arXiv