Abstract
This paper introduces a framework for identifying and utilizing Semantically Identical Temporally Altering Concepts (SITACs)—entities whose names or titles evolve over time, such as "Ceylon" becoming "Sri Lanka." By combining linguistic dependency analysis with association rule mining, the authors developed a method to automatically translate modern search queries into their historical equivalents. This ensures that archives spanning decades remain searchable despite shifting terminology, bridging the temporal gap in information retrieval.
The Problem of Temporal Terminology Evolution
Standard search engines typically struggle with historical archives because language is dynamic. For instance, a query regarding the independence of "Sri Lanka" might fail to retrieve critical documents from 1948 that refer exclusively to the country as "Ceylon."
This is not a simple matter of synonymy; it is a problem of historical fact and temporal evolution. The authors identify a critical need for "intelligent" systems capable of understanding how concepts—including people, places, and organizations—transform across time. Without such systems, vast quantities of historical data remain effectively hidden from modern researchers.
Methodology
The proposed pipeline for discovering and utilizing SITACs consists of three primary stages:
1. Linguistic Preprocessing
The system utilizes the Minipar parser to analyze text archives, focusing specifically on the relationships between nouns (concepts) and verbs (events). The underlying hypothesis is that if two different nouns are frequently associated with the same types of events or actions, they likely represent the same concept at different points in time.
2. Association Rule Mining
Using the Apriori algorithm, the system generates rules in the format of (Concept1, Time1) => (Concept2, Time2). For example, the system might identify a strong association between the term "Union" in the 1860s and "United States" in the 1900s.
3. Ranking with Jaccard’s Coefficient
To filter out noise and weak associations, the authors apply Jaccard’s coefficient to measure contextual similarity. If "Union" and "USA" share a high percentage of surrounding context words (e.g., "government," "accession," or "treaty"), the rule is ranked higher in the knowledge base.
Evaluation and Key Insights
The primary evaluation was conducted using the Gutenberg corpus, specifically U.S. Presidential speeches. This dataset provides a rich chronological span, allowing researchers to track the evolution of terminology (e.g., "The Union" vs. "The Nation" vs. "The United States") through different political eras.
Key insights from the study include:
- Verbs as the "Glue": Using verbs to identify shared events is a highly effective way to link concepts across time. If "Concept A" and "Concept B" are both subjects of the same action in different years, there is a high probability they are SITACs.
- Beyond Synonymy: SITACs are distinct from synonyms. While a synonym is a linguistic choice, a SITAC represents a historical transformation.
- Simulating Human Association: Association rule mining effectively mimics how humans mentally link evolving concepts based on shared historical context.
Results and Future Work
The researchers successfully generated a knowledge base of temporal rules, such as 1795=Union ==> 1958=United States and 1952=war ==> 1999=terrorist. By ranking these rules, the system creates a "translation engine" that intercepts user queries and expands them to include historically accurate terms.
While the framework demonstrates the feasibility of automated temporal translation, future work is needed to establish large-scale quantitative benchmarks (such as Precision/Recall) and to improve the scalability of the pipeline for multi-terabyte news archives and the broader web.
Summary updated