On the Language Encoder of Contrastive Cross-modal Models

TL;DR

This paper investigates how different sentence embedding training methods impact the quality of language encoders in contrastive cross-modal models like CLIP and CLAP, revealing benefits in vision-language tasks but limited effects in audio-language tasks.

Contribution

It provides a comprehensive evaluation of sentence embedding training effects on language encoder quality and cross-modal task performance, with insights into their differing impacts on vision and audio modalities.

Findings

Sentence embedding training improves VL model performance.

AL pretraining benefits less from sentence embedding training.

Sentence embedding training enhances text-space uniformity but reduces cross-modal alignment.

Abstract

Contrastive cross-modal models such as CLIP and CLAP aid various vision-language (VL) and audio-language (AL) tasks. However, there has been limited investigation of and improvement in their language encoder, which is the central component of encoding natural language descriptions of image/audio into vector representations. We extensively evaluate how unsupervised and supervised sentence embedding training affect language encoder quality and cross-modal task performance. In VL pretraining, we found that sentence embedding training language encoder quality and aids in cross-modal tasks, improving contrastive VL models such as CyCLIP. In contrast, AL pretraining benefits less from sentence embedding training, which may result from the limited amount of pretraining data. We analyze the representation spaces to understand the strengths of sentence embedding training, and find that it improves text-space uniformity, at the cost of decreased cross-modal alignment.