Semantic Networks as Clues: A Theoretical Foundation and Process Optimization for Semantic Network Construction

This paper addresses the challenge of constructing semantic networks (SNs) from textual data, specifically focusing on networks that represent non-propositional knowledge. Unlike Knowledge Graphs (KGs), which often function as surrogates for reality and rely on verifiable propositions, the authors argue that these specific SNs serve as "clues" to reality. Because these clues lack a definitive "gold standard" for verification, the authors reformulate the Semantic Network Construction (SNC) process as a Process Optimization Problem (POP).

The authors break down the SNC process into three stages: Automatic Keyphrase Extraction (AKE), Edge Weighting (EW), and Community Detection (CD). They define specific local evaluation criteria for each stage—hF1 for keyness, a percolation-theory-based RI for interpretability, and a community scoring function for distinctiveness—and integrate these into a global objective function to rank and optimize the resulting networks.

The primary contribution is the ClueNetwork framework, which provides a systematic way to evaluate and rank candidate semantic networks based on their joint confidence in achieving the three local objectives. By moving away from the requirement of an external gold standard, the framework allows researchers to select the most reliable method combinations for a given dataset. The paper also introduces innovative methodological components, such as a veracity pretest for edge weighting and the use of percolation theory to measure network interpretability, providing a more rigorous foundation for exploratory text analysis.

In many fields, researchers use semantic networks to explore large, unstructured textual datasets where the goal is not to verify specific propositions, but to uncover latent structures and topics. By formalizing the SNC process as an optimization problem, this work helps researchers move beyond conventional, ad-hoc pipelines (like simple TF-IDF and co-occurrence frequency) toward more evidence-based and reproducible network construction. This is particularly valuable for exploratory research where the "truth" is not pre-defined, but rather something to be inferred from the data itself.