When humans read a sentence like "The meeting is at two," we effortlessly resolve the ambiguity based on whether it is currently morning or afternoon, or if we are scheduling a future date. For a machine, however, mapping natural language expressions of time (known as temporal expressions or timexes) to an absolute timeline is a profound challenge. Unlike static entities such as names or places, time is inherently relative, shifting, and deeply dependent on context.
To build systems that can reason about timelines, researchers have moved beyond simple pattern matching. The field has evolved from identifying what a temporal expression is to understanding why it means what it means in a specific conversational or narrative moment.
The Dimensions of Temporal Underspecification
Why is time so difficult for algorithms to pin down? The literature identifies three distinct layers of linguistic and pragmatic ambiguity:
1. Deictic and Indexical Shifts
Many temporal expressions are anchored entirely to the moment they are spoken or written (the utterance time). As shown in Robust Temporal Processing of News, Inderjeet Mani et al. point out that a word like "today" can refer to a specific calendar date in a news article, or it can be used generically to describe a broad historical era or social condition. Similarly, exophoric references like "this year" require a dedicated temporal reasoning framework to anchor the text to external metadata rather than relying solely on textual analysis, a challenge highlighted in Resolving Direct and Indirect Anaphora for Japanese Definite Noun Phrases by Naoya Inoue et al.
2. Dynamic Discourse Anchoring
When a narrative progresses, the "reference time" moves. If a text says, "He graduated in 2010. Two months later, he started a job," the phrase "Two months later" anchors to 2010, not the document's creation date. Tracking this shifting anchor requires modeling how context flows across sentences. In Temporal Centering, Megumi Kameyama et al. formalized this process, showing that discourse uses defeasible logic to update the current temporal anchor.
When local cues are missing, models must look globally. For instance, in One Tense per Scene: Predicting Tense in Chinese Conversations, Tao Ge et al. demonstrated that conversational tense in languages like Chinese (which lacks explicit verb conjugations) can be resolved by enforcing consistency across an entire "scene" rather than evaluating sentences in isolation.
3. Polysemy of Connectives and Signals
Even explicit temporal markers like "before", "after", or "while" are highly ambiguous. In Inferring Sentence-internal Temporal Relations, Alex Lascarides notes that conjunctions frequently pull double duty, acting as both temporal connectors and logical operators (such as contrast). Furthermore, Temporal Signals Help Label Temporal Relations by Leon Derczynski et al. shows that a marker like "before" can indicate a simple chronological sequence or define the precise termination boundary of a time interval.
"While she was coding, he cooked dinner." --> Temporal Overlap
"While she loves coding, he prefers math." --> Logical Contrast (Non-Temporal)
The Solution Architecture: Decoupling Recognition from Interpretation
Early systems attempted to use massive, brittle sets of regular expressions to simultaneously find and calculate dates. However, this monolithic approach fails when faced with the pragmatic messiness of human speech.
The breakthrough insight in modern temporal resolution is a two-stage cascaded architecture. As pioneered in An Intermediate Representation for the Interpretation of Temporal Expressions by Pawel P. Mazur et al. and expanded in A Cascaded Machine Learning Approach to Interpreting Temporal Expressions by David Ahn et al., the problem is split into two distinct tasks:
- Detection (Linguistic): Identifying the span of text that constitutes a temporal expression (e.g., recognizing that "the week before last" is a timex).
- Normalization (Pragmatic): Computing the actual, standardized value (such as an ISO-8601 date like
2023-10-02) by combining the expression's literal components with the relevant reference time.
By introducing an intermediate semantic representation, systems can parse the literal structure of a phrase without needing to immediately know the current date. This separates the grammar of language from the arithmetic of timelines.
From Words to Math: Compositional Parsing
How does a machine actually compute a date from an intermediate representation? It treats language like math.
Expressions like "the day before yesterday" have a compositional structure. In A Synchronous Context Free Grammar for Time Normalization, Steven Bethard demonstrated that this can be modeled using a Synchronous Context-Free Grammar (SCFG). The SCFG maps natural language directly to a formal algebra of temporal operators:
$$\text{"the day before yesterday"} \longrightarrow \text{MINUS}(\text{MINUS}(\text{ReferenceDate}, 1\text{ day}), 1\text{ day})$$
To handle the inherent ambiguity of which operators to apply, Parsing Time: Learning to Interpret Time Expressions by Christopher D Manning et al. introduced a probabilistic approach that treats the underlying structure of a time expression as a latent parse tree. By using an expectation-maximization (EM) bootstrapping algorithm, the system learns the statistical priors of how phrases map to temporal calculations. This allows it to resolve ambiguous terms like "next weekend" based on the most likely speaker intent given the specific day of the week the statement was made.
Leveraging Structural and Global Context
When explicit markers are completely absent, systems must rely on the underlying syntax and discourse structure to infer timelines. In Extracting Temporal Event Relation with Syntax-guided Graph Transformer, SHUAICHENG ZHANG et al. addressed this by using a syntax-guided attention mechanism to explicitly query the dependency paths between events, filtering out noisy context to capture subtle temporal cues.
When structural cues conflict or remain vague, incorporating fuzzy temporal commonsense reasoning, as explored in Self-Supervised Logic Induction for Explainable Fuzzy Temporal Commonsense Reasoning by Bibo Cai, helps models resolve underspecified terms (like distinguishing morning vs. evening contexts) by aligning them with human-like expectations of how events typically unfold in the real world.
Key Foundational Work
Uber Zeitkeferenz Und Tempus (1969) — contributed 1 concept elements to this direction.
Chronometric Studies of Lexical Ambiguity Resolution (1980) — contributed 1 concept elements to this direction.