Non-negative Elastic Net Decoding for Information Retrieval

The paper presents Non-Negative Elastic Net (NNN) decoding as an innovative approach to information retrieval, addressing the limitations of traditional dense retrieval methods that often yield redundant results. NNN decoding treats document selection as a joint decoding problem, allowing for a more diverse set of retrieved documents by using a non-negative linear combination of embeddings. Experimental results demonstrate that NNN decoding significantly outperforms conventional dense retrieval across various benchmarks, highlighting its potential for enhancing retrieval systems by optimizing embeddings beyond mere inner-product scoring.