An Attention-Enhanced Text Error Detection Model based on Potential Miswritten Term Pre-check Feature in Manufacturing

Manufacturing specifications contain a large number of terms, which makes general domain text error detection methods ineffective in manufacturing. Existing text error detection methods have a limited utilization of term dictionary, and the challenge lies in the inability to determine whether a word not in the term dictionary is an out-of-vocabulary word or a miswritten word. To tackle this challenge, we introduce a Miswritten Term Pre-check feature Matrix (MTP-Matrix), which leverages n-gram similarity matching method to pre-check potential miswritten terms. Besides, we capture the information about error positions and types of miswritten terms through adding Attention-enhanced Network with Local Feature (ANLF), which enhances the representation of miswritten terms. Experimental results demonstrate that the proposed MTP-Matrix and ANLF can improve the F1 values of all baselines. The highest F1 value reaches 76.84%, and the F1 value of term error detection increases by 6.09%.