In Defense of Character-Level Answer Generation Methods for Text-based Visual Question Answering

Character-based language models have traditionally received limited attention due to their perceived shortcomings in capturing the linguistic features of the target language. However, this study demonstrates that, for current Text-based Visual Question Answering (VQA) datasets, deep learning models that generate answers at the character level can achieve performance that is comparable to—or even surpasses—that of models using subword units, all while requiring significantly fewer tokens. Experimental results on multiple Text-based VQA benchmarks in both English and Vietnamese confirm the effectiveness of character-based approaches and showcase their resilience in dealing with rare and out-of-vocabulary (OOV) words.