Table 1.
Compares key aspects of the existing works, including model architecture, embedding techniques, datasets, results, and noted strengths – limitations.
Fig 1.
Overall Framework of the Proposed Text Similarity System which begins with text preprocessing and feature extraction, followed till explainability achieved through SHAP and LIME for interpretability..
Table 2.
Hyperparameter Settings for the Proposed Siamese BiLSTM Model.
Fig 2.
Logical flow diagram of the proposed Siamese BiLSTM-based semantic similarity framework, illustrating each processing step from initial token-level text encoding to final similarity score generation, including semantic accumulation, sentence interaction modeling, and nonlinear projection through a regression head.
Table 3.
Experimental Setup and Implementation Details.
Table 4.
Comprehensive results analysis of all applied models for text similarity.
Fig 3.
Error Comparison Between TF-IDF Cosine Baseline and Siamese BiLSTM Model.
Fig 4.
Analysis of BiLSTM Similarity Predictions with Confidence Bands.
Fig 5.
Ground Truth vs. Predicted Similarity: TF-IDF vs. Siamese BiLSTM.
Fig 6.
Smoothed output curves for TF-IDF and Siamese BiLSTM predictions exhibit lower variance and more stable behavior, indicating stronger semantic representation and reduced sensitivity to lexical noise.
Fig 7.
Residuals prediction analysis of the Siamese BiLSTM compared to the SVR baseline.
Fig 8.
The relationship between average sentence length and the corresponding similarity scores for both the predicted values (light blue) and the ground truth annotations (green).
Fig 9.
BiLSTM prediction compared to ground truth, along with sentence lengths and prediction error (Δ).
The near-zero error demonstrates accurate semantic understanding even with differences in sentence structure or length.
Fig 10.
Analysis of absolute error as a function of average sentence length, divided into four interpretative quadrants shows Siamese model performs strongly in both accurate quadrants, showing robustness to sentence length variation.
Fig 11.
The feature-importance plot shows the relative contribution of lexical and embedding-based features.
Table 5.
Statistical numerical results analysis of semantic similarity metrics for individual sentence pairs.
Fig 12.
Distribution Analysis of Feature Metrics shows the skewed patterns, and lexical metrics with multi-modal behavior, reflecting the dataset’s varied semantic relationships.
Fig 13.
Correlation structure analysis among lexical similarity measures and embedding-based distances.
Fig 14.
Global SHAP values showing the importance and direction of each feature in the model.
Fig 15.
SHAP Force Plot for Local Similarity Explanation showing how individual similarity features push the model’s prediction higher (red) or lower (blue) relative to the base value.
Fig 16.
Local SHAP breakdown illustrating how individual features raise or lower the predicted similarity score relative to the model’s performance.
Fig 17.
LIME Local Explanation for proposed model Interpretation showing the contribution of specific feature intervals to the model’s prediction.
Fig 18.
Comparative analysis of similarity score progression of all models against the ground truth, with shaded uncertainty regions.
Fig 19.
Comparison based on per-sample prediction deviation for all models.
The Siamese BiLSTM exhibits the smallest residual fluctuations, indicating higher prediction stability and resilience to lexical inconsistency.
Fig 20.
Sliding-window Pearson correlation trends comparing models across sorted test samples.
The envelope band shows the overall correlation stability range, highlighting the enhanced consistency and semantic robustness of the BiLSTM model.
Fig 21.
Scatterplot showing predicted similarity versus ground-truth similarity for all models.
Table 6.
Comparison analysis of existing approaches performance with proposed study.