Automatic Recognition Method of Machine English Translation Errors Based on Multisignal Feature Fusion

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Machine Translation

Background:

• Current automatic recognition methods for machine English translation errors lack robust semantic analysis, leading to low accuracy.
• Existing models struggle to effectively identify and classify various translation errors.

Purpose of the Study:

• To design an advanced automatic recognition method for machine English translation errors.
• To enhance the accuracy and robustness of machine translation error detection models through multifeature fusion.

Main Methods:

• Manually classifying and summarizing real error sentence pairs and employing data augmentation to create a robust dataset.
• Integrating source text to translation length ratio and language model perplexity (PPL) as input features.
• Utilizing a multifeature fusion approach combining word posterior probability and linguistic features.

Main Results:

• The proposed method significantly reduces the classification error rate compared to existing approaches.
• Incorporating source word features alongside word posterior probability and linguistic features demonstrably improves error detection accuracy.
• The inclusion of source text to translation length ratio and PPL enhances the model's classification accuracy.

Conclusions:

• Multifeature fusion, particularly combining word posterior probability, linguistic features, and source text characteristics, is effective for improving machine English translation error detection.
• The developed error detection scheme can be applied to error correction, user experience enhancement, and machine translation evaluation.