Multi-word Term Embeddings Improve Lexical Product Retrieval

TL;DR

Addressing the unique demands of product search, the paper introduces H1, a multi-word term embedding within a dual-encoder framework that treats selected brand terms as single tokens in a hybrid lexical-semantic index. It trains with a hinge-like objective $L_{H1} = [ gamma - s(q_{1:m}, p_{1:n}^{+}) + s(q_{1:m}, p_{1:n}^{-}) ]_{+0}$ and uses a token-wise similarity mechanism inspired by ColBERT. The authors augment the tokenizer with semantically rich terms, notably brand names, and demonstrate state-of-the-art offline index-based retrieval on WANDS with $mAP@12 = 56.1%$ and $R@1k = 86.6%$, outperforming ColBERT and other baselines. The results show that multi-word embeddings and hybrid search markedly improve precision without reducing recall, offering practical benefits for large-scale e-commerce search systems. They also outline plans for broader benchmarks and architecture exploration to establish objective evaluation frameworks for semantic product-search models.

Abstract

Product search is uniquely different from search for documents, Internet resources or vacancies, therefore it requires the development of specialized search systems. The present work describes the H1 embdedding model, designed for an offline term indexing of product descriptions at e-commerce platforms. The model is compared to other state-of-the-art (SoTA) embedding models within a framework of hybrid product search system that incorporates the advantages of lexical methods for product retrieval and semantic embedding-based methods. We propose an approach to building semantically rich term vocabularies for search indexes. Compared to other production semantic models, H1 paired with the proposed approach stands out due to its ability to process multi-word product terms as one token. As an example, for search queries "new balance shoes", "gloria jeans kids wear" brand entity will be represented as one token - "new balance", "gloria jeans". This results in an increased precision of the system without affecting the recall. The hybrid search system with proposed model scores mAP@12 = 56.1% and R@1k = 86.6% on the WANDS public dataset, beating other SoTA analogues.

Figures (4)

• Figure 1: The indexes and a tokenization model built during offline information extraction are used in online setting to respond to queries with low latency. The quality of built index is detrimental to the performance of the search system.
• Figure 2: Token handling principle of H1 (a), compared to that of a FastText (b). H1 attributes scores to each pair of query and document tokens, while the FastText-based system compares the mean vector representations.
• Figure 3: Ablation study results over tokenization methods and model architectures.
• Figure 4: An illustrative one-query example of how Precision decreases and Recall increases for different semantic retrieval models with respect to cut-off threshold $k$.