Abstract
This paper addresses the challenge of culture-specific temporal grounding—mapping subjective terms like "morning" or "evening" to precise hourly ranges across different languages. We propose three language-agnostic methods, including a corpus-based extractive approach and two language model (LM) based techniques, refined through Integer Linear Programming (ILP). Our results show that the extractive method, which leverages Wikipedia co-occurrence data, significantly outperforms LMs by capturing distinct cultural boundaries that models often over-smooth.
Introduction
Grounding natural language in physical properties is a core challenge in AI. While spatial and color grounding have seen significant progress, temporal grounding remains difficult due to its inherent subjectivity and cultural dependence. A "good evening" in a tropical climate may start hours earlier than in northern latitudes. Without culturally aware temporal grounding, AI systems risk logistical failures and social faux pas.
Methodology
We evaluate three primary methods for grounding five core expressions (morning, noon, afternoon, evening, night):
- Extractive Approach: This method parses Wikipedia for sentences where a time expression co-occurs with a specific timestamp, building a frequency distribution of hours for each term.
- LM-Based Distribution Estimation: Using mBERT, we estimate the probability of specific hours within fixed templates (e.g.,
"It is [MASK] in the morning."). - LM-Based Direct Prediction: We query the model directly for start and end times (e.g.,
"The morning starts at [MASK].").
To ensure logical consistency (e.g., non-overlapping spans in chronological order), we apply Integer Linear Programming (ILP) to the raw distributions. This optimization maximizes the inclusion of observed data points within the predicted spans while enforcing temporal constraints.
Data and Cultural Variation
The study utilizes 349 human annotations for English (US), Hindi (India), Italian (Italy), and Portuguese (Brazil). Wikipedia serves as the cross-lingual corpus.
As illustrated in Figure 2, cultural differences are pronounced. In India, "night" is perceived to start much earlier than in Western cultures, likely due to climate and sunset patterns. In Brazil, the term noite is used broadly for both evening and night, often causing the "evening" category to collapse in predictive models.
Results
The Extractive method proved most robust, achieving high accuracy across English (84.3%), Italian (90.1%), and Hindi (80.4%).
As shown in Figure 6, the extractive approach closely mirrors human "Gold" standards. In contrast, Language Models tend to produce "too smooth" distributions, failing to distinguish between mutually exclusive temporal concepts because their vector representations are too similar. The extractive method also outperformed previous state-of-the-art baselines based on Twitter greetings in most languages.
Discussion and Limitations
While Wikipedia is a powerful cultural proxy, it is subject to "reporting bias," where common knowledge (e.g., "10 am is morning") is rarely explicitly stated. Additionally, using language as a proxy for culture ignores geographic variations within language groups (e.g., Portugal vs. Brazil). Future work should incorporate seasonality, as temporal boundaries shift with the sun throughout the year.
Conclusion
This work demonstrates that cultural nuances in time perception can be automatically extracted from large-scale corpora. By combining extractive methods with ILP, we can build more culturally aware AI systems that better understand the subjective nature of time across the globe.
Summary updated