Automatic tense and aspect translation between Chinese and English.

摘要

The current dissertation studies the problem of translation of tense and aspect between Chinese and English, which represent two typologically different languages with diverse tense and aspect strategies. The dissertation first collects human an notations in order to obtain inter-annotator agreement rate. A series of statistical analyses are then performed on the annotation results, revealing the significance of different linguistic factors in human annotations, which motivates feature selection in automatic tense and aspect translation.; The dissertation then formulates the tasks of tense and aspect translation as classification problems. The experimental results show that fully automated classifiers using Conditional Random Fields and automatically extracted features can improve upon the tense and aspect marker generation of state-of-the-art Machine Translation systems. The automatic systems are then augmented with deeper features---in particular, lexical aspectual properties that are possessed by humans but not immediately available for computer systems; punctuality and telicity features demonstrate high utility in the augmented system. This adds to our knowledge as to how to narrow the gap between an automatic tense classifier and human tense translation. Lexical aspectual features, however, do not have as strong an impact on aspect marker classification in the opposite scenario. Additionally, the impact of different feature groups on tense and aspect classifications is reported.; The dissertation is the very first comprehensive investigation of cross-linguistic tense and aspect translation via machine learning techniques. It advances our understanding of the impact of different feature groups in tense and aspect translation across the language pair of Chinese and English. Additionally, it identifies parallels between human and machine tense and aspect translation, which relates to a viable research topic in a broader view, i.e., to look for factors accounting for the gap between human and machine language processing performance.