Abstract
Temporal scope adds a time dimension to facts in Knowledge Bases (KBs). These time scopes specify the time periods when a given fact was valid in real life. Without temporal scope, many facts are underspecified, reducing the usefulness of the data for upper-level applications such as Question Answering. Existing methods for temporal scope inference and extraction still suffer from low accuracy. In this paper, we present a new method that leverages temporal profiles augmented with context—Contextual Temporal Profiles (CTPs) of entities. Through change patterns in an entity’s CTP, we model the entity’s state change brought about by real-world events that happen to the entity (e.g., hired, fired, divorced, etc.). This leads to a new formulation of the temporal scoping problem as a state change detection problem. Our experiments show that this formulation of the problem and the resulting solution are highly effective for inferring the temporal scope of facts.
Introduction
Knowledge Bases (KBs) such as YAGO or Freebase are invaluable resources for structured data, yet they are often "frozen" in time. They may store a fact like "Barack Obama is the President of the USA" without specifying when that fact was true. This lack of temporal scoping limits the ability of applications to answer time-sensitive queries or track historical changes.
Existing approaches to temporal scoping generally fall into two categories: direct extraction from specific sentences (which is often inaccurate due to linguistic complexity) or aggregate mention counting. The latter, exemplified by the previous state-of-the-art system CoTS, relies on the frequency of entity mentions. However, mention counts are noisy and often require manually crafted constraints—such as "a country can only have one president at a time"—to produce reliable results.
Methodology: Contextual Temporal Profiles (CTPs)
This research proposes a shift from counting mentions to analyzing the context surrounding an entity. The core contribution is the Contextual Temporal Profile (CTP), which treats temporal scoping as a state-change detection problem.
1. Building CTPs
For a specific entity (e.g., "Richard Nixon"), the system collects the unigrams and bigrams that appear in its proximity within a time-stamped corpus. To eliminate noise, tf-idf is employed to filter out common words, retaining only the top 100 most descriptive context units for each year.
2. State Change Detection
Rather than searching for a simple spike in mentions, the model looks for a "linguistic signature" of a state change. For instance, when an individual transitions to the presidency, the surrounding vocabulary shifts predictably from terms like "candidate" or "senator" to "elect," then "administration," and finally "president."
3. Learning and Inference
The system learns from a small set of "seed" examples—entities known to have undergone a specific state change. It calculates an Aggregate State Vector (the average context during the state) and an Aggregate Change Vector (the average difference in context from the year prior to the change). To find the start date for a new entity, the system compares its yearly CTP and change vectors against these learned aggregates using cosine similarity. The year with the highest similarity is predicted as the "begin" time.
Experimental Setup
The researchers evaluated the CTP method using two major time-stamped corpora:
- Google Books Ngram (1960–2008): Used to model the US Administration domain (Presidents, Vice Presidents, and Secretaries of State).
- English Gigaword (1994–2008): Used for the Academy Awards domain (Best Director, Best Picture).
Performance was measured against a human-annotated "gold standard" of true begin times.
Key Results and Performance
The CTP approach achieved high precision across multiple domains. In the US Administration domain, the true begin time was consistently found within the top 5 results (Precision @ 5).
Crucially, CTPs demonstrated significant improvements in domains where mention frequency is a poor indicator of state. In the Academy Awards domain, where a win might not cause a massive spike in mentions compared to a general scandal or a major film release, the context-driven approach proved far more robust than the frequency-based CoTS system.
As shown in the comparison, CTPs achieved F1 scores that were comparable to or better than state-of-the-art methods without requiring the manual constraints that systems like CoTS rely upon. This suggests that the linguistic transition itself is a reliable and purely data-driven proxy for real-world events.
Limitations and Future Directions
A primary challenge identified by the authors is the detection of "end" times. While CTPs are highly effective at spotting the beginning of a state, they suffer from "linguistic inertia." Even after a state ends (e.g., a president leaves office), people continue to refer to the entity using their former title in text. This prevents the context from "resetting" as cleanly as it begins.
Future work will focus on identifying specific signals for state termination and exploring finer-grained temporal resolutions, such as months or days, by utilizing daily newsfeeds rather than yearly aggregates.
Summary updated