Abstract
We address the problem of automatically inferring the tense of events in Chinese text. We use a new corpus annotated with Chinese semantic tense information and other implicit Chinese linguistic information using a “distant annotation” method. We propose three improvements over a relatively strong baseline method – a statistical learning method with extensive feature engineering. First, we add two sources of implicit linguistic information as features – eventuality type and modality of an event, which are also inferred automatically. Second, we perform joint learning on semantic tense, eventuality type, and modality of an event. Third, we train artificial neural network models for this problem and compare its performance with feature-based approaches. Experimental results show considerable improvements on Chinese tense inference. Our best performance reaches 68.6% in accuracy, outperforming a strong baseline method.