DialectGen: Benchmarking and Improving Dialect Robustness in Multimodal Generation

Yu Zhou*,†,  Sohyun An*,  Haikang Deng*,  Da Yin,  Clark Peng, 
Cho-Jui Hsieh,  Kai-Wei Chang,  Nanyun Peng 
UCLA 
*Core Contributors, †Project Lead
Paper Dataset Code

Multimodal Generative Model Outputs on synonymous prompts that differ only in one synonymous lexical feature from Standard American English (top) or a lower-resource English dialect (bottom).

Abstract

Contact languages like English exhibit rich regional variations in the form of dialects, which are often used by dialect speakers interacting with generative models. However, can multimodal generative models effectively produce content given dialectal textual input? In this work, we study this question by constructing a new large-scale benchmark spanning six common English dialects. We work with dialect speakers to collect and verify over 4200 unique prompts and evaluate on 17 image and video generative models. Our automatic and human evaluation results show that current state-of-the-art multimodal generative models exhibit 32.26% to 48.17% performance degradation when a single dialect word is used in the prompt. Common mitigation methods such as fine-tuning and prompt rewriting can only improve dialect performance by small margins (< 7%), while potentially incurring significant performance degradation in Standard American English (SAE). To this end, we design a general encoder-based mitigation strategy for multimodal generative models. Our method teaches the model to recognize new dialect features while preserving SAE performance. Experiments on models such as Stable Diffusion 1.5 show that our method is able to simultaneously raise performance on five dialects to be on par with SAE (+34.4%), while incurring near zero cost to SAE performance.

DialectGen -- Benchmark


    Paired SAE / Dialect data in DialectGen, including Lexeme, Concise Prompt, and Detailed Prompt.

Mitigation Method

Losses used in our mitigation. Text prompts for Dialect Learning and Polysemy Control come from the DialectGen training set, while image-caption pairs for KL Regularization come from the MSCOCO validation set.


Qualitative Results

Qualitative Comparison of Mitigation Strategies including generation results using Stable Diffusion 1.5 Base Model, Fine-tuning with Diffusion DPO, and Ours (Dialect Learning + Image KL Regularization + Polysemy Control).