Utilizing Embeddings for Ad-hoc Retrieval by Document-to-document Similarity

TL;DR

This paper introduces a novel document-to-document similarity method using embeddings for ad-hoc retrieval, addressing limitations of query-based similarity by leveraging content-based comparisons, and demonstrates improved performance on TREC datasets.

Contribution

The paper proposes a new semantic relevance scoring method based on document-to-document similarity of embeddings, enhancing retrieval accuracy over traditional query-based approaches.

Findings

Outperforms strong baselines on TREC test collections

Addresses the problem of multiple degrees of similarity in embeddings

Demonstrates the effectiveness of content-based similarity in IR

Abstract

Latent semantic representations of words or paragraphs, namely the embeddings, have been widely applied to information retrieval (IR). One of the common approaches of utilizing embeddings for IR is to estimate the document-to-query (D2Q) similarity in their embeddings. As words with similar syntactic usage are usually very close to each other in the embeddings space, although they are not semantically similar, the D2Q similarity approach may suffer from the problem of "multiple degrees of similarity". To this end, this paper proposes a novel approach that estimates a semantic relevance score (SEM) based on document-to-document (D2D) similarity of embeddings. As Word or Para2Vec generates embeddings by the context of words/paragraphs, the D2D similarity approach turns the task of document ranking into the estimation of similarity between content within different documents. Experimental results on standard TREC test collections show that our proposed approach outperforms strong baselines.