Table 1.
Samples of various question types.
Table 2.
Statistical information of FQSD dataset.
Table 3.
Descriptive metrics for FQSD dataset.
Table 4.
Pearson correlation among annotators.
Table 5.
Interpretation ranges for Fleiss’s Kappa.
Fig 1.
Top 30 TF-IDF scores of nouns and noun phrases in FQSD.
Fig 2.
Top 20 TF-IDF Scores of Adverbs/Adjectives in FQSD Categorized by Subjectivity Comparison-Form Classes: a) CO b) CS c) SO d) SS.
Fig 3.
Histograms of TF-IDF Adverbs/Adjectives Scores in FQSD Categorized by Subjectivity Comparison-Form Classes: a) CO b) CS c) SO d) SS.
Table 6.
Statistical information of Yu et al. [7] dataset.
Table 7.
Statistical information of ConvEx-DS dataset [9].
Table 8.
Statistical information of SubjQA dataset [8].
Fig 4.
Distribution of total question count and multi-sentence question count across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the size and structural analysis of each dataset.
Fig 5.
Distribution of the total word count and unique word count across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the lexical richness of each dataset.
Fig 6.
Distribution of the average words per sentence, average sentence length, average word length, and average syllables per word across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the linguistic complexity of each dataset.
Fig 7.
Visualizing the average parse tree depth across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the syntactic complexity of each dataset.
Fig 8.
Visualizing the Mean Dependency Distance (MDD) across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the dependency analysis of each dataset.
Fig 9.
Visualizing the Root Type-Token Ratio (RTTR) and the Corrected Type-Token Ratio (CTTR) across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the lexical diversity of each dataset.
Fig 10.
Visualizing the sparsity degree and total question count across the FQSD, ConvEx-DS, Yu et al., 2012, and SubjQA datasets, showcasing the data sparsity of each dataset.
Table 9.
RoBERTa’s five-fold cross-validation evaluation on FQSD.
Fig 11.
LIME visualizations of word influence in model’s predictions for Instances 1 (Fig 11a), and 2 (Fig 11b) on the Yu et al. [7] dataset.
Table 10.
Model performance across different dataset sizes (averaged over 5 runs using stratified 5-fold cross-validation).
Table 11.
Analysis of the proposed subjectivity classification model over five separate runs on the SUBJQA dataset [8].
Table 12.
Analysis of the proposed subjectivity-comparison form classification model over five separate runs on the ConvEx-DS dataset [9].
Table 13.
Evaluation of the proposed model (Trained on FQSD and tested on Yu et al. [7] dataset) vs. Yu et al. [7] model.
Table 14.
Comparative analysis of transformer models’ performance (LR = 1e-5) on the FQSC task across the FQSD dataset over five independent runs.
Table 15.
Comparative analysis of transformer models’ performance (LR = 1e-5) on the subjectivity-comparison form classification task across the ConvEx-DS dataset [9] over five independent runs.
Table 16.
Comparative analysis of transformer models’ performance (LR = 3e-5) on the subjectivity classification task across the SubjQA dataset [8] over five independent runs.
Table 17.
Comparative analysis of transformer models’ performance on fine-grained subjectivity tasks across multiple datasets over five independent runs.