# BMX: Entropy-weighted Similarity and Semantic-enhanced Lexical Search

**Authors:** Xianming Li, Julius Lipp, Aamir Shakir, Rui Huang, Jing Li  
**arXiv:** [2408.06643](https://arxiv.org/abs/2408.06643)  
**Date:** 2024-08-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.06643v2/

## Key points

- Authors introduce BMX as an extension to the widely-used BM25 algorithm
- BMX incorporates entropy-weighted similarity metrics and semantic enhancement techniques
- Aim of BMX is to address limitations of BM25 in neglecting query-document similarity and lacking semantic understanding
- Integration of features results in a more robust lexical search algorithm with improved retrieval of relevant documents
- Introduction of weighted query augmentation technique enhances semantic understanding within lexical search
- Approach bridges gap between traditional lexical matching and modern semantic comprehension
- Study shows that BMX outperformed BM25 and PLM/LLM-based dense retrieval models in long-context and real-world retrieval benchmarks
- Potential highlighted for improving lexical search performance by integrating entropy-weighted similarity metrics and semantic enhancement techniques
- Authors introduced Baguetter evaluation framework for information retrieval with reference implementation of BMX
- Normalization technique presented for both BMX and BM25 scores to enhance utility in information retrieval tasks

## Summary

In this paper, the authors revisit the widely-used BM25 lexical search algorithm and introduce a novel extension called BMX. By incorporating entropy-weighted similarity metrics and semantic enhancement techniques, BMX aims to address the limitations of BM25 in neglecting query-document similarity and lacking semantic understanding. The integration of these features results in a more robust lexical search algorithm that demonstrates improved retrieval of relevant documents. One key aspect of BMX is the introduction of a weighted query augmentation technique that enhances semantic understanding within lexical search, further enhancing its effectiveness. This approach bridges the gap between traditional lexical matching and modern semantic comprehension, offering a promising direction for future information retrieval research. The study conducted extensive experiments to evaluate the performance of BMX compared to traditional BM25 and PLM/LLM-based dense retrieval models. The results consistently showed that BMX outperformed BM25 and surpassed PLM/LLM-based approaches in long-context and real-world retrieval benchmarks. This highlights the potential of integrating entropy-weighted similarity metrics and semantic enhancement techniques in improving lexical search performance. Additionally, the authors introduced Baguetter, an evaluation framework for information retrieval with a reference implementation of BMX. They also presented a normalization technique for both BMX and BM25 scores to further enhance their utility in information retrieval tasks. Overall, this study contributes to advancing information retrieval research by introducing an innovative approach that combines classical lexical search with modern semantic approaches. The findings suggest that incorporating entropy-weighted similarity metrics and semantic enhancement techniques can significantly improve the effectiveness of lexical search algorithms like BM25.