Automatic temporal expression normalization with reference time dynamic-choosing

Temporal expressions in texts contain significant temporal information. Understanding temporal information is very useful in many NLP applications, such as information extraction, documents summarization and question answering. Therefore, the temporal expression normalization which is used for transforming temporal expressions to temporal information has absorbed many researchers' attentions. But previous works, whatever the hand-crafted rules-based or the machine-learnt rules-based, all can not address the actual problem about temporal reference in real texts effectively. More specifically, the reference time choosing mechanism employed by these works is not adaptable to the universal implicit times in normalization. Aiming at this issue, we introduce a new reference time choosing mechanism for temporal expression normalization, called reference time dynamic-choosing, which assigns the appropriate reference times to different classes of implicit temporal expressions dynamically when normalizing. And then, the solution to temporal expression defuzzification by scenario dependences among temporal expressions is discussed. Finally, we evaluate the system on a substantial corpus collected by Chinese news articles and obtained more promising results than compared methods.