IME-Spell: Chinese Spelling Check based on Input Method

Abstract

Intended for reducing manual inspection costs and semantic misunderstandings, Chinese Spelling Check (CSC) has been investigated extensively in natural language processing. However, little work has yet been done on input-method-based CSC in which CSC can make use of Pinyin information to improve spelling correction efficiency. This paper proposes a novel CSC architecture, IME-Spell, based on pre-trained context vectors for input methods, which consists of two parts as follows. The Chinese spelling detection part of the architecture adopts the fusion vectors of character-based pre-trained context vectors and Pinyin vectors, and uses the method of sequence labeling to detect the error characters. The Chinese spelling correction part of the architecture adopts Masked Language Model (MLM) to generate a candidate set of erroneous characters, and uses XGBoost and Pinyin-to-Character conversion models to filter correct characters and correct the error characters for users. IME-Spell has a significant improvement over the benchmark models on the SIGHAN dataset, whose maximum difference of F1 in the spelling detection and correction subtasks reach 48.9% and 27.8%, respectively.