Fig 1.
The TSRCDF-SS structure diagram includes a dual encoders, and the improved classifier combines sequential transfer and cross-domain transfer.
Fig 2.
Encoder t-SNE dimensionality reduction projection.
Comparison of sentence embedding performance of different encoders and combined encoders using t-SNE dimensionality reduction projection. This visualization highlights the differences in encoding capabilities of different encoders. The dataset used is TRAINNLI.
Fig 3.
Software requirement pair encoding model based on SBERT and SimCSE.
Use the SBERT model and SimCSE model to encode individual requirements to obtain their respective embeddings. Then these embeddings are fused to finally obtain the six-element concatenated embedding result.
Fig 4.
The pseudocode of fusion algorithm of sequential transfer and cross-domain transfer.
Table 1.
Operating parameters of the model used in the experiment.
Fig 5.
Comparison of encoder combination results.
Precision, recall and F1 of different encoder combination experiments on TRAINNLI (30000) dataset.
Fig 6.
Comparison of results of fully connected layers with different numbers of layers.
Results of comparative experiments on different numbers of FFNN layers on the TRAINNLI (30000) dataset.
Fig 7.
t-SNE visualization of feature embeddings on the CDN dataset at different stages of the FFNN classifier.
(a) input embeddings before the fully connected layers, (b) embeddings after one fully connected layer, and (c) embeddings after two fully connected layers.
Fig 8.
Comparison of results of different loss functions.
The results of the experiment using a two-layer FFNN and different loss functions on the TRAINNLI (30000) dataset.
Table 2.
The improvement effect of different datasets on the improved model is evaluated, and ablation comparison experiments are performed on the TRAINNLI (30000) dataset (Unit:%).
Table 3.
Performance evaluation results on the TRAINNLI (30000) dataset (Unit:%).
Table 4.
Evaluation of cross-domain models trained with different combinations of requirement pair datasets (Unit:%).
Fig 9.
Representative confusion matrices for cross-domain requirement conflict detection.
These two examples correspond to (a) the WorldVista case, where the target domain data distribution is moderately imbalanced, and (b) the OPENCOSS case, where it is extremely imbalanced. These confusion matrices are reported to illustrate typical patterns under cross-domain transfer rather than to exhaustively characterize all source–target configurations.