AttnInput: Advancing Context-Aware Pinyin Input with Efficient Language Model Integration

ACL ARR 2025 February Submission5940 Authors

Abstract: Pinyin input methods are essential for typing Chinese characters, yet existing approaches struggle to balance accuracy with computational efficiency when integrating large language models (LLMs). This paper introduces AttnInput, a novel framework that enhances pinyin input performance through lightweight language model adaptation. By integrating pinyin features directly into the model’s inference process via a parameter-efficient side network, AttnInput eliminates the need for costly full-model fine-tuning while significantly improving prediction accuracy. The method employs constrained training and inference strategies to enforce phonetic alignment, reducing ambiguity in pinyin sequences. Experiments demonstrate state-of-the-art results across varying context and pinyin lengths, with a 20–34\% accuracy improvement over existing methods on long sequences. AttnInput achieves these gains while maintaining linear computational complexity, enabling stable latency and low resource consumption even for extended contexts. The framework reduces training costs by over 50\% compared to conventional fine-tuning approaches, showcasing practical advantages for edge-side deployment and scalable language model integration.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: NLP Applications, Efficient/Low-Resource Methods for NLP

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: Chinese

Submission Number: 5940