Good Night at 4 pm?! Time Expressions in Different Cultures

Findings of the Association for Computational Linguistics: ACL 2022, pages 2842 - 2853 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Good Night at 4 pm?! Time Expressions in Different Cultures Vered Shwartz Department of Computer Science, University of British Columbia vshwartz@cs.ubc.ca Abstract We propose the task of culture-specific time ex- pression grounding, i.e. mapping from expres- sions such as “morning” in English or “manhã” in Portuguese to specific hours in the day. We propose 3 language-agnostic methods, one of which achieves promising results on gold stan- dard annotations that we collected for a small number of languages. We then apply this method to 27 languages and analyze the sim- ilarities across languages in the grounding of time expressions. 1 Introduction Natural language understanding requires the ability to map language such as color descriptions (McMa- han and Stone, 2015), spatial instructions (Chen et al., 2019), and gradable adjectives (Shivade et al., 2016) to real-world physical properties. This pa- per focuses on temporal grounding, particularly mapping time expressions such as “morning” and “evening” to hours in the day. Temporal common- sense reasoning has been gaining traction lately (Zhou et al., 2019; Qin et al., 2021), and this impor- tant capability can benefit various temporal tasks such as event ordering and duration prediction. One of the challenges in grounding time expres- sions to standard times is that such expressions may be interpreted with some variation by differ- ent people. Reiter and Sripada (2002) found that human-written weather forecasts exhibited signifi- cant individual differences between forecasters in the interpretation of time expressions. One factor for this variation is cultural differences. Vilares and Gómez-Rodríguez (2018) analyzed the time of day in which people from 53 countries posted time-specific greetings such as “good morning” and “good evening” on Twitter. They showed variation in greeting times across languages and cultures, which they connected to known facts and published statistics about cultural differences, such as differ- ences in average wake and sleep times. We propose to re-frame the research question posed by Vilares and Gómez-Rodríguez (2018) as a task of time expression grounding: given a time expression, the goal is to map it to a range of hours during the day. For example, what is the range of hours an Italian speaker refers to when saying pomeriggio (afternoon)? Such a grounding model can provide cultural context to machine translation systems (de Medeiros Caseli et al., 2010), language learning apps (Teske, 2017), and user-centered dia- logue systems (Miehle et al., 2016). We collected gold standard interpretations from four countries, which indeed exhibited some varia- tion. We then proposed 3 language-agnostic meth- ods based on either a corpus or a language model (LM). The corpus-based method performed well across languages, outperforming the method pro- posed by Vilares and Gómez-Rodríguez (2018) on 3 out of 4 languages. Encouraged by the perfor- mance on the labelled languages, we applied the method to additional 23 unlabelled languages, and analyzed the differences predicted by the models. In the future, we plan to incorporate …

Content truncated to 500 words