Dialectic Feature-Based Fuzzy Graph Learning for Label Propagation Assisting Text Classification

The abundant deposits of unstructured and scarcely labeled data over social networks make text classification (TC) vital for structuring and extracting useful information. In addition, ignoring dialectal variations significantly hinders the performance of international English (especially American and British) TC across numerous data domains. To address this multifaceted challenge, a comprehensive and adaptable framework termed dialectic feature-based fuzzy graph learning (DFFGL) is introduced that learns feature vectors by inculcating semantics and dialect variations from the inputted text. DFFGL then proficiently extracts uniquely modified terms frequency-inverse document frequency, parts-of-speech-tagged $\mathcal {N}-\text{grams}$, with dialect-specific dictionary features in the fuzzy feature space to realize a novel language model. Later, these fuzzified features are affined by a novel fuzzy distance measure to construct an interpretable fuzzy graph that is then optimized using a novel elastic net regularizer for characterizing nodal relations, promising efficient classification through effective label propagation. Exhaustive $F1-$score evaluations on 6 English corpora and 17 diverse datasets reveal DFFGL's superiority in consistently registering over 93% and 80% in dialect identification and TC even with just 10 labeled samples. Furthermore, DFFGL offers remarkable $F1-$score improvements of 10.2% and 17.3% over its peers in respective tasks, highlighting its extension to real-world data classification.