Entropy-based discrimination between translated Chinese and original Chinese using data mining techniques

Area of Science:

• Computational Linguistics
• Corpus Linguistics
• Translation Studies

Background:

• Distinguishing translated text from original text is crucial for understanding translationese.
• Previous studies have explored linguistic features, but data mining offers a novel approach.
• Monolingual comparable corpora provide a basis for analyzing subtle linguistic differences.

Purpose of the Study:

• To apply data mining techniques to differentiate between translated and original Chinese.
• To investigate the effectiveness of entropy-based metrics in identifying translational features.
• To evaluate machine learning classifiers for classifying translated versus original Chinese text.

Main Methods:

• Operationalized seven entropy-based metrics (character, wordform, and POS trigrams) from comparable corpora.
• Utilized four machine learning techniques: Support Vector Machines (SVMs), Linear Discriminant Analysis (LDA), Random Forest (RF), and Multilayer Perceptron (MLP).
• Trained and tested classifiers on balanced corpora of translated and non-translated Chinese.

Main Results:

• Support Vector Machines (SVMs) demonstrated superior performance, achieving an Area Under the Curve (AUC) of 90.5% and an accuracy rate of 84.3%.
• The study confirmed that translational language exhibits distinct characteristics compared to original language.
• Entropy measures combined with machine learning effectively identified features indicative of translation.

Conclusions:

• The hypothesis that translational language is categorically different from original language is affirmed.
• Combining information-theoretic indicators (Shannon's entropy) with machine learning provides a robust method for analyzing translationese.
• This research offers new insights into corpus-based studies of translation phenomena.