ASR-enhanced Multimodal Representation Learning for Cross-Domain Product Retrieval

TL;DR

This paper introduces AMPere, a novel multimodal product representation learning method that leverages noisy ASR text and visual data to improve cross-domain product retrieval in e-commerce.

Contribution

It proposes a new approach combining LLM-based ASR text summarization with visual data for enhanced multimodal product embeddings.

Findings

AMPere significantly improves cross-domain product retrieval accuracy.

The method effectively denoises ASR text using LLM summarization.

Experiments on a large-scale dataset validate the approach's superiority.

Abstract

E-commerce is increasingly multimedia-enriched, with products exhibited in a broad-domain manner as images, short videos, or live stream promotions. A unified and vectorized cross-domain production representation is essential. Due to large intra-product variance and high inter-product similarity in the broad-domain scenario, a visual-only representation is inadequate. While Automatic Speech Recognition (ASR) text derived from the short or live-stream videos is readily accessible, how to de-noise the excessively noisy text for multimodal representation learning is mostly untouched. We propose ASR-enhanced Multimodal Product Representation Learning (AMPere). In order to extract product-specific information from the raw ASR text, AMPere uses an easy-to-implement LLM-based ASR text summarizer. The LLM-summarized text, together with visual data, is then fed into a multi-branch network to generate compact multimodal embeddings. Extensive experiments on a large-scale tri-domain dataset verify the effectiveness of AMPere in obtaining a unified multimodal product representation that clearly improves cross-domain product retrieval.