Factorized Learning Assisted with Large Language Model for Gloss-free Sign Language Translation

TL;DR

This paper introduces FLa-LLM, a two-stage training approach for gloss-free sign language translation that effectively leverages large language models without compromising visual representation learning.

Contribution

It proposes a novel factorized training framework that pre-trains visual encoders separately and then fine-tunes LLMs, improving gloss-free SLT performance.

Findings

Significant improvements on three SLT datasets.

Effective separation of visual and language learning stages.

Enhanced translation accuracy without gloss annotations.

Abstract

Previous Sign Language Translation (SLT) methods achieve superior performance by relying on gloss annotations. However, labeling high-quality glosses is a labor-intensive task, which limits the further development of SLT. Although some approaches work towards gloss-free SLT through jointly training the visual encoder and translation network, these efforts still suffer from poor performance and inefficient use of the powerful Large Language Model (LLM). Most seriously, we find that directly introducing LLM into SLT will lead to insufficient learning of visual representations as LLM dominates the learning curve. To address these problems, we propose Factorized Learning assisted with Large Language Model (FLa-LLM) for gloss-free SLT. Concretely, we factorize the training process into two stages. In the visual initialing stage, we employ a lightweight translation model after the visual encoder to pre-train the visual encoder. In the LLM fine-tuning stage, we freeze the acquired knowledge in the visual encoder and integrate it with a pre-trained LLM to inspire the LLM's translation potential. This factorized training strategy proves to be highly effective as evidenced by significant improvements achieved across three SLT datasets which are all conducted under the gloss-free setting.