Improving Event Duration Question Answering by Leveraging Existing Temporal Information Extraction Data

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4451–4457 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4451 Improving Event Duration Question Answering by Leveraging Existing Temporal Information Extraction Data Felix Giovanni Virgo, Fei Cheng, Sadao Kurohashi Graduate School of Informatics, Kyoto University Kyoto, Japan {felix,feicheng, kuro}@nlp.ist.i.kyoto-u.ac.jp Abstract Understanding event duration is essential for understanding natural language. However, the amount of training data for tasks like duration question answering, i.e., McTACO, is very limited, suggesting a need for external duration information to improve this task. The duration information can be obtained from existing temporal information extraction tasks, such as UDS-T and TimeBank, where more duration data is available. A straightforward two-stage fine-tuning approach might be less likely to succeed given the discrepancy between the target duration question answering task and the intermediary duration classification task. This paper resolves this discrepancy by automatically recasting an existing event duration classification task from UDS-T to a question answering task similar to the target McTACO. We investigate the transferability of duration information by comparing whether the original UDS-T duration classification or the recast UDS-T duration question answering can be transferred to the target task. Our proposed model achieves a 13% Exact Match score improvement over the baseline on the McTACO duration question answering task, showing that the two-stage fine-tuning approach succeeds when the discrepancy between the target and intermediary tasks are resolved. Keywords: event duration, temporal common sense, question answering, data recasting 1. Introduction Understanding how long an event typically lasts is es- sential in natural language processing. Many NLP tasks, such as narrative understanding, event timeline construction, question answering, and natural language inference (Nakhimovsky, 1987; Ning et al., 2018; Zhou et al., 2019; Leeuwenberg and Moens, 2019; Vashishtha et al., 2020), require knowledge about the typical duration of events. However, it is still challeng- ing for machines to comprehend the duration of various events available. An event verb can have different du- rations depending on its context. For example, “take a vacation” takes longer than “take a shower.” While taking a shower typically takes a few minutes, a va- cation can last for days or even weeks. Acquiring the duration of various events by hand is also costly and time-consuming. McTACO (Zhou et al., 2019) is a temporal common- sense question answering dataset that consists of ques- tions from 5 temporal phenomena including event du- ration. The event duration questions are the focus of our paper. This work shows that the performance of modern pre-trained NLP models for this task is still far behind humans. Since the amount of training data only covers a limited number of events and their at- tributes, incorporating external event duration informa- tion is necessary to improve this task. Leveraging relevant intermediary tasks has shown to be beneficial for improving target tasks with limited data (Phang et al., 2018; Liu et al., 2019). For ex- ample, Liu et al. (2019) shows that it is beneficial to fine-tune the …

Content truncated to 500 words