Time is a fundamental dimension of human language, yet it is notoriously difficult for machines to master. Traditional natural language processing (NLP) relies on expensive, manually labeled datasets to teach models how to recognize dates or understand event durations. However, a growing body of research suggests that we don't need human annotators to "explain" time to AI. Instead, we can treat the temporal structure of the world—the sequence of medical records, the timestamps of knowledge base updates, and the linguistic patterns in news—as a form of "free" or weak supervision.
By leveraging these inherent signals, researchers are building systems that ground messy natural language in precise temporal logic without the bottleneck of manual labeling.
The Future as a Label
One of the most intuitive ways to supervise a model without manual effort is to look at what happens later. In longitudinal data, such as Electronic Health Records (EHR), the importance of a past note is often revealed by a future outcome.
Borna Dabiri, Byron C Wallace, and their colleagues demonstrated this in Query-Focused EHR Summarization to Aid Imaging Diagnosis. Rather than asking doctors to highlight relevant snippets in a patient's history, they used "future" diagnosis codes as distant supervision. If a patient is eventually diagnosed with a specific condition, the model learns that the historical snippets leading up to that diagnosis were likely the most relevant ones.
Similarly, Xiangan Liu and Eric P. Xing used the chronological nature of medical documentation to generate pseudo-labels for summarization. They treated subsequent documents as natural summaries of earlier ones, exploiting the fact that critical information is intrinsically repeated and refined over time.
Distant Supervision from Knowledge Bases
The "birth" and "death" of information in structured databases provides another powerful signal for training models on real-time data like social media.
Satoshi Akasaki and Naoki Yoshinaga 0001 have applied this to the problem of "emerging" and "disappearing" entities. In Fine-grained Typing of Emerging Entities in Microblogs, they identified the exact date an entity was first added to a knowledge base and used that timestamp to automatically pull relevant "initial mentions" from social media. This allowed them to train a classifier on the specific linguistic context of an entity’s debut.
They later inverted this logic in Early Discovery of Disappearing Entities in Microblogs, using historical records of when entities (like restaurants or services) were marked as closed to train a model that recognizes the "terminal context" of a topic. By contrasting posts from an entity's active period with those from its disappearance, the model learns to spot endings weeks before they are officially recorded.
Learning Logic from Denotations
Temporal expressions like "the week before last" are complex because they are compositional—their meaning depends on a reference date and a series of functional operations. Traditionally, this required hand-written rules or semantic parses.
However, Christopher D Manning and Gabor Angeli pioneered an approach that treats the internal structure of time as a latent process. In Parsing Time: Learning to Interpret Time Expressions, they showed that if you provide a model with a phrase and its final grounded date (the "denotation"), the model can autonomously discover the underlying logic required to get there. Angeli and Jakob Uszkoreit further refined this in Language-Independent Discriminative Parsing of Temporal Expressions, proving that mapping natural language directly to target dates is a viable alternative to labor-intensive manual tree-banking.
Mining Temporal Common Sense
Finally, the vast corpus of existing text contains implicit "common sense" about how time works. We know a "concert" usually lasts hours, not seconds or years, because of how people write about them.
Zonglin Yang and Alexander M Rush demonstrated in Improving Event Duration Prediction via Time-aware Pre-training that simple linguistic patterns (like "lasted for" or "spent X hours") can be used to harvest hundreds of thousands of event-duration pairs from news articles. This "cheap" supervision is sufficient to teach models the typical lengths of real-world events.
Ben Zhou and Dan Roth 0001 expanded this in Temporal Common Sense Acquisition with Minimal Supervision, using semantic role labeling to extract event-time pairs at scale. Their model, TACOLM, learns the ordinal nature of time—that "seconds" are shorter than "minutes"—not through a dictionary, but by observing the statistical distribution of these terms in relation to events.
Through these diverse methods, the research direction moves away from "teaching" time through manual labels and toward "discovering" time through the natural structure of data.