Enhancing Swipe Typing with Gated Linear Transformer
Ni Jia, Ying Zhang, Qing Yu Wu · 2024
Swipe typing has gained popularity as an efficient method for smart devices. However, decoding swipe input presents challenges, including trajectory variability, jitter noise, and deviation tendencies. To address these challenges, we propose a novel approach for sliding input encoding that enhances long-term memory. Our method combines a multi-layer multi-head memory gated linear transformer with Connectionist Temporal Classification (CTC) to effectively preserve dependencies. By incorporating multiple layers and heads, our model captures diverse aspects of the sliding input, enabling comprehensive representation. The integration of CTC ensures accurate alignment between input and output labels. Furthermore, we employ efficient decoding techniques such as many-to-one mapping, lexicon finite-state transducers, beam search, and n-gram ranking, enhancing decoding performance. Extensive experiments demonstrate significantly improved long-term memory retention and accuracy in sliding input encoding tasks.