How And Where Do People Fail With Time: Temporal Reference Mapping Annotation By Chinese And English Bilinguals

This work reports on three human tense annotation experiments for Chinese verbs in Chinese-to-English translation scenarios. The results show that inter-annotator agreement increases as the context of the verb under the annotation becomes increasingly specified, i.e. as the context moves from the situation in which the target English sentence is unknown to the situation in which the target lexicon and target syntactic structure are fully specified. The annotation scheme with a fully specified syntax and lexicon in the target English sentence yields a satisfactorily high agreement rate. The annotation results were then analyzed via an ANOVA analysis, a logistic regression model and a log-linear model. The analyses reveal that while both the overt and the latent linguistic factors seem to significantly affect annotation agreement under different scenarios, the latent features are the real driving factors of tense annotation disagreement among multiple annotators. The analyses also find the verb telicity feature, aspect marker presence and syntactic embedding structure to be strongly associated with tense, suggesting their utility in the automatic tense classification task.