Traceable Reverse Engineering of UML Artifacts from SRS/SDD Documents Using NLP-Based Text Processing and Siamese BiLSTM: An IdVar4CL Case Study

Software Engineering (SE) is increasingly shifting toward Artificial Intelligence (AI)-driven engineering, emphasizing process automation, quality improvement, and faster development cycles across industry and academia [1]. In Software Requirements Engineering (SRE), advances include the use of Design Thinking for richer elicitation [2], Recommender Systems for predicting or suggesting requirements [3], and systematic reviews on Unified Modeling Language (UML) diagram utilization in SE research [4-9]. In parallel, Natural Language Processing (NLP)—including terminology extraction and ontology construction [10], Exabsum-based summarization for information compression [11], and data augmentation for improving model generalization [12, 13]—has enhanced the accuracy of unstructured text analysis. Nevertheless, integrating cross-text artifact semantics between Software Requirements Specification (SRS) and Software Design Description (SDD) to ensure consistent design representation across Use Case, Sequence, Class, and State Diagrams remains an open and significant research challenge.

Several gaps and challenges remain. First, many AI studies in SE emphasize conceptual frameworks or automation at specific stages (e.g., testing, debugging, predictive maintenance) but lack deep text processing to preserve semantic coherence between documentation artifacts and diagrams [1, 4]. Second, studies on Identification Variable for Causal Loop (IdVar4CL) have demonstrated the extraction of causal variables from system narratives via text mining [8] and shown potential for semantic integration in Use Case Scenarios (UCSs) and FR/NFR artifacts [6, 7]; however, these processes remain manual and have not yet leveraged bidirectional sequential models such as Bidirectional Long Short-Term Memory (BiLSTM) to enhance consistency and accuracy. Third, while BiLSTM has proven effective at capturing bidirectional context in text-based tasks—such as fake review detection using BiLSTM with attention and positional embeddings [14] and code evaluation/improvement—its application to the reverse engineering of SRS/SDD-based UML artifacts remains limited. Furthermore, ensuring trustworthiness and explainability requires a robust pipeline that maintains data quality and transparent process traceability to produce reliable reverse-engineering outcomes [15-17]. Overall, this highlights a critical research gap in the lack of an integrated approach combining advanced text processing and BiLSTM for semantically consistent reverse engineering of UML artifacts from SRS/SDD documents.

Our research is motivated by the practical need in the field that development teams require a measurable and reproducible way to (i) extract information units from SRS/SDD, (ii) map them to UML artifacts consistently (e.g., Use Case/Sequence Diagram), and (iii) translate the results into a Unified Specification Document (USD) that minimizes semantic drift. The IdVar4CL artifact prototype, which produces mappings between Architecturally Significant Requirements (ASR) and Quality Attributes (QA) [5], UCSs Alignment [6] and FR/NFR formation [7], has shown the right direction, but requires improvement through a two-way sequential model and Machine Learning-based validation flow to ensure robustness and traceability in line with the Responsible AI / Trustworthy AI agenda [15-17]. Thus, adopting BiLSTM, which has been proven effective for text/sequence modeling [14, 18-21], is considered appropriate to overcome the limitations of manual approaches and to build an accurate and transparent reverse-engineering pipeline.

In line with the broader roadmap of our previous research [5-7], the main objective of this study is to propose and evaluate a text-processing and Siamese BiLSTM-based reverse engineering pipeline for UML-related artifacts in the IdVar4CL case. The pipeline is designed not only to infer semantic relationships among documentation artifacts but also to generate a complete and traceable USD by constructing explicit FR→UC→SP paths. Specifically, this study performs the following activities:

1. Capturing the bidirectional context of SRS/SDD through BiLSTM.
2. Generating semantically consistent UML artifacts, such as FR/NFR, Use Case Diagrams (UCDs), and UCSs.
3. Compiling an updated IdVar4CL USD that explicitly represents traceable FR→UC, UC→SP, FR→SP, and end-to-end FR→UC→SP paths as an integrated specification-design artifact.
4. Ensuring trustworthiness through data validation and measurable evaluation (accuracy/consistency).

The main contributions of this research are as follows:

1. An integrated reverse-engineering pipeline that converts written requirements into UML-related artifacts and USD. This study presents a pipeline that uses text preprocessing and Siamese BiLSTM-based pair classification. It identifies semantic relationships among Functional Requirements, Use Cases, and Steps Performed.
2. Generation of traceable paths linking Functional Requirements (FR) to Use Cases (UC) and Steps Performed (SP). The approach predicts whether document pairs are linked or unlinked. It then transforms these results into clear traceability paths that connect Functional Requirements, Use Cases, and Steps Performed. These paths form the structural basis of the updated IdVar4CL USD.
3. A complete and traceable IdVar4CL USD. The updated USD preserves the artifact layers: FR, UC, and SP. It shows their directional relationships using FR→UC, UC→SP, and FR→SP mappings. This supports specification-design consistency and reduces semantic drift.
4. Evaluation and validation of artifact linkage inference.  The pipeline is evaluated using stratified k-fold validation. Standard classification metrics, including accuracy, precision, recall, F1-score, and ROC-AUC, measure the reliability of semantic linkage inference.

In terms of scientific implications and impacts, this research addresses the integration gap between NLP and BiLSTM for Reverse Engineering UML from unstructured specification documents. Practically, the proposed pipeline provides a reproducible path for SE teams to align SRS/SDD, UML, and USD with a clear process trail (in line with Responsible/Trustworthy AI good practices) [15-17]. This approach can also be extended to other domains with semantic integration challenges (e.g., Knowledge Graphs/Ontologies [18]) and to large-scale data scenarios (e.g., big data summarization [22-26]).

Furthermore, the organization of this paper includes the following sections: Section 2 reviews related work (SRE, NLP, BiLSTM, Trustworthy/Responsible AI) that underlies this research [1-27]; Section 3 defines the use of datasets derived from UML artifacts called IdVar4CL. Section 4 describes the methods and pipeline (text preprocessing, BiLSTM process, mapping to UML, USD compilation, and evaluation scheme); Section 5 presents the experimental results and analysis; Section 6 summarizes the results, contributions, and future work plans.

Based on these SPs, the resulting documents used as datasets are as follows:

1. The researcher should access the data input form. IdVar4CL Application must display the page. The researcher should upload the document. IdVar4CL. Application must receive the document from the researcher.
2. The researcher should access the preprocessing section. IdVar4CL Application must show the preprocessing interface. The researcher must select the uploaded documents. IdVar4CL Application should verify document availability. The researcher should start preprocessing. IdVar4CL Application must process the selected documents. IdVar4CL Application should apply case folding. IdVar4CL Application should tokenize the text. IdVar4CL Application should remove stopwords. IdVar4CL Application should apply stemming. IdVar4CL Application should save the processed text. IdVar4CL Application must notify preprocessing completion.
3. The researcher should select document pairs for similarity calculation. IdVar4CL Application must verify selected documents are available. The researcher should initiate the similarity calculation. IdVar4CL Application must calculate similarity using TF-IDF and semantic algorithms. The IdVar4CL Application should display similarity results to the researcher. A researcher may analyze the similarity scores.
4. The researcher should select word pairs with similarity scores above 0.50. IdVar4CL Application must verify the selected word pairs. The researcher should initiate the extraction of candidate variables. IdVar4CL Application must extract candidate variables from the selected word pairs. The IdVar4CL Application should display the extracted candidate variables. The researcher may review and analyze the candidates.
5. The Systems Analyst should select extracted variables. IdVar4CL Application must verify that the variables are available. The Systems Analyst should initiate diagram construction. IdVar4CL Application must generate causal loop diagrams from the variables. The IdVar4CL Application should display the constructed diagrams. Systems Analyst may review and modify the diagrams.

By applying a One-to-Many relationship across the 15 documents, a total of 75 variations of inter-document relationships are generated as datasets. The implementation of the One-to-Many relationship in this research is intended solely to maintain consistency in the linkage patterns and directional relationships among research artifacts.

It should be noted that the dataset used in this study is intentionally limited to the IdVar4CL case study and consists of 15 labeled documents, namely five Functional Requirements, five Use Cases, and five Steps Performed. These documents were expanded into 75 inter-document relationship pairs using a one-to-many relationship scheme. Therefore, the dataset represents a controlled case-study setting rather than a large-scale multi-project benchmark. This design allows the proposed pipeline to be evaluated in a focused, traceable manner, but it also limits the extent to which the model's performance can be generalized to other systems, domains, or documentation styles.

Based on the explanations of the three artifacts (FR, UCD, and UCS), the final documents used as datasets are defined as follows:

1. D1 = Researcher must upload storytelling documents into the application interface.
2. D2 = Researcher must initiate text preprocessing on uploaded documents, including case folding, tokenization, stopword removal, and stemming.
3. D3 = Researcher must calculate similarity between document pairs using TF-IDF and semantic similarity algorithms.
4. D4 = Researcher must extract candidate variables from word pairs with semantic similarity scores above 0.50.
5. D5 = Systems Analyst must construct causal loop diagrams using the extracted variables.
6. D6 = Upload Storytelling Document.
7. D7 = Initiate Text Preprocessing.
8. D8 = Calculate Similarity Documents.
9. D9 = Extract Candidate Variables.
10. D10 = Construct Causal Loop Diagram.
11. D11 = The researcher should access the data input form. IdVar4CL Application must display the page. The researcher should upload the document. IdVar4CL Application must receive the document from the researcher.
12. D12 = The researcher should access the preprocessing section. IdVar4CL Application must show the preprocessing interface. The researcher must select the uploaded documents. IdVar4CL Application should verify document availability. The researcher should start preprocessing. IdVar4CL Application must process the selected documents. IdVar4CL Application should apply case folding. IdVar4CL Application should tokenize the text. IdVar4CL Application should remove stopwords. IdVar4CL Application should apply stemming. IdVar4CL Application should save the processed text. IdVar4CL Application must notify preprocessing completion.
13. D13 = The researcher should select document pairs for similarity calculation. IdVar4CL Application must verify selected documents are available. The researcher should initiate the similarity calculation. IdVar4CL Application must calculate similarity using TF-IDF and semantic algorithms. The IdVar4CL Application should display similarity results to the researcher. A researcher may analyze the similarity scores.
14. D14 = Researcher should select word pairs with similarity scores above 0.50. IdVar4CL Application must verify the selected word pairs. The researcher should initiate the extraction of candidate variables. IdVar4CL Application must extract candidate variables from the selected word pairs. The IdVar4CL Application should display the extracted candidate variables. The researcher may review and analyze the candidates.
15. D15 = Systems Analyst should select extracted variables. IdVar4CL Application must verify that the variables are available. The Systems Analyst should initiate diagram construction. IdVar4CL Application must generate causal loop diagrams from the variables. The IdVar4CL Application should display the constructed diagrams. Systems Analyst may review and modify the diagrams.

5.6 False positive analysis and negative sample quality

The validation results show that the model achieved perfect recall across all folds, meaning that no positive linked pairs were missed. However, false positives occurred in Fold 3 and Fold 5. This indicates that the model occasionally classified unlinked pairs as linked. Such errors are important in the context of traceability generation because false positives may introduce incorrect FR→UC, UC→SP, or FR→SP links into the updated USD. If not controlled, these incorrect links may cause over-linking and reduce the precision of the generated FR→UC→SP paths.

The presence of false positives suggests that the current negative sample construction may not be sufficiently challenging. Several IdVar4CL artifacts share similar actors, domain terms, and sentence structures. For example, multiple documents refer to the “Researcher,” “IdVar4CL Application,” “documents,” “similarity,” “variables,” and “diagrams.” Such lexical overlap may cause the Siamese BiLSTM model to assign high similarity to unlinked pairs, especially when the negative examples do not include enough semantically close but incorrect pairs. Therefore, the false positives are likely related not only to model behavior but also to the difficulty level and representativeness of the negative samples.

To improve the model's robustness, future training should incorporate hard negative sampling. In this strategy, negative samples are intentionally selected from unlinked pairs that are lexically or structurally similar to positive pairs. This can include pairs with shared actors, similar verbs, common domain terms, or high textual similarity scores that do not belong to the same traceability path. By exposing the model to more difficult unlinked examples, hard negative sampling is expected to improve precision and reduce the risk of over-linking in the generated USD.

5.7 Process: Impact of evaluation metrics on FR→UC→SP traceability

The main objective of the evaluation is not only to assess whether the Siamese BiLSTM model can correctly classify artifact pairs, but also to determine whether the predicted links can support the construction of a reliable USD traceability matrix. In this study, the traceability matrix is constructed from predicted relationships among FR–UC, UC–SP, and FR–SP pairs. These relationships are then interpreted as directional traceability paths connecting Functional Requirements to Use Cases and Steps Performed. From this perspective, recall is the most critical metric for traceability completeness. If the model produces false negatives, valid links may be omitted from the USD. For example, a missed FR→UC link would prevent a functional requirement from being connected to its corresponding use case, while a missed UC→SP link would prevent the use case from being elaborated into scenario-level steps. Such missing links could break the end-to-end FR→UC→SP path. Therefore, the recall value of 1.0000 across all folds indicates that the model preserved complete traceability coverage for the linked pairs in the IdVar4CL dataset.

Precision is equally important for traceability correctness. Although high recall ensures that valid links are not missed, false positives may introduce incorrect links into the traceability matrix. In the context of USD construction, false positives may result in over-linking, where an FR is connected to an unrelated UC or SP, or a UC is connected to scenario steps from another context. This may reduce the interpretability and reliability of the generated USD. Therefore, the observed false positives in Fold 3 and Fold 5 indicate that negative sample construction and semantic validation should be strengthened to improve precision.

F1-score provides a balanced view of traceability quality by combining precision and recall. A high F1-score indicates that the model can maintain both completeness and correctness of trace links. In this study, the average F1-score of 0.9754 ± 0.0396 suggests that the generated USD traceability matrix achieves a strong balance between preserving valid links and avoiding incorrect links, although remaining false positives still require further improvement.Accuracy provides a general indication of matrix consistency, but it is not sufficient as the sole metric for traceability construction. In traceability tasks, the costs of false negatives and false positives are different.

False negatives may break traceability paths, while false positives may create misleading paths. Therefore, accuracy must be interpreted together with precision, recall, and F1-score. ROC-AUC indicates the model’s ability to distinguish linked and unlinked pairs across different thresholds. This is relevant for future USD construction because traceability matrices can be generated not only as binary matrices but also as probability-based matrices. In such a setting, ROC-AUC can guide threshold calibration to determine which predicted links should be included in the USD. This would allow stakeholders to balance traceability completeness and correctness according to project needs.

In practical terms, the two main error types have different effects on the USD. A false negative removes a valid traceability link. For example, if FR-CL02 is not linked to its corresponding preprocessing use case, the USD would fail to show how the requirement is realized in the design artifact. This reduces traceability completeness. Conversely, a false positive adds an invalid traceability link. For example, if a preprocessing-related FR is incorrectly linked to an unrelated diagram construction scenario, the USD would contain an incorrect FR→UC→SP path. This reduces traceability correctness and may mislead stakeholders during requirements validation or design review. Please see Table 7.

Table 7. Impact of classification errors on Unified Specification Document (USD) traceability

5.8 Output 1: Updated Unified Specification Document

The updated USD is constructed based on the artifact-pair classifier’s evaluation results by transforming predicted semantic links into explicit traceability relations among FR, UC, and SP. True positive links are included as valid traceability mappings, while true negative pairs are excluded. False positives may reduce precision by adding incorrect links, whereas false negatives may reduce traceability completeness by omitting required links. Therefore, recall represents traceability coverage, precision reflects traceability correctness, and the F1-score serves as a combined indicator of overall USD traceability quality. As the final artifact of the proposed reverse-engineering pipeline, the updated USD provides more than a prediction report; it serves as an integrative documentation artifact with complete, directional traceability. It explicitly represents FR→UC, UC→SP, and FR→SP mappings derived from document pairs labeled linked (1), enabling end-to-end FR→UC→SP paths. These paths show how functional requirements are realized through use cases and operationalized into scenario-level steps, supporting both forward and backward traceability, minimizing semantic drift, and strengthening specification-design consistency.

Example traceability path (illustrative): D1 (FR-CL01) → {D6..D10} (UC) → {D11..D15} (SP).

To formally present the mapping, Tables 8–10 show probability-based traceability matrices between artifacts. Each cell represents the predicted probability of a semantic link generated by the Siamese BiLSTM model and is interpreted using thresholds: ✓ (p ≥ 0.70) for accepted links, ? (0.50 ≤ p < 0.70) for borderline links, and × (p < 0.50) for rejected links. This approach reflects the model’s ability to distinguish valid and invalid relationships rather than assuming full connectivity among artifacts.

Table 8. Traceability matrix FR × UC

Table 9. Traceability matrix UC × SP

Table 10. Traceability matrix FR × SP

The evaluation results show that the generated traceability matrices achieve strong within-case coverage, with no false negatives across all folds, indicating that all expected linked relationships are preserved in the IdVar4CL setting. However, false positives suggest that some predicted links still require semantic validation before inclusion in the USD. Thus, the proposed pipeline should be viewed as a semi-automated mechanism for constructing traceability, where model predictions generate candidate links and human or rule-based validation confirms borderline cases. Collectively, these matrices support end-to-end FR→UC→SP path construction, enabling stakeholders to trace how requirements relate to use case behavior and scenario steps while aligning SRS/SDD artifacts with UML-based documentation.

5.9 Output 2: Recommendations for improvement/ development

Recommendations are based on validation results and error analysis. Despite excellent recall, several false positives occurred in Fold 3 and Fold 5, reducing precision. This finding indicates that similarities in sentence structure or terminology between documents can lead the model to treat unrelated pairs as similar. Based on these findings, recommendations include:

1. Standardizing domain terminology and requirement/ scenario writing patterns to reduce ambiguity.
2. Applying hard negative sampling in follow-up experiments by constructing unlinked artifact pairs that are lexically or structurally similar to linked pairs. These hard negatives may include pairs with shared actors, similar domain terms, similar verb-object patterns, or high TF-IDF/semantic similarity scores, yet remain outside the valid FR→UC→SP traceability path. This strategy is expected to reduce false positives and improve precision by forcing the model to distinguish true semantic traceability from superficial textual similarity.
3. Implementing a human-in-the-loop to review pairs with borderline probability as part of an audit trail and trustworthiness assessment.
4. Metric-based USD validation interprets recall as traceability coverage, precision as traceability correctness, F1-score as their balance, and ROC-AUC as the basis for threshold calibration to decide whether FR→UC→SP paths should be accepted, reviewed, or rejected.

5.10 Dataset size, generalization, and threats to validity

Although the proposed Siamese BiLSTM pipeline achieved strong validation results, the dataset size should be considered when interpreting the model performance. The current experiment uses 15 labeled documents from the IdVar4CL case study, consisting of five Functional Requirements, five Use Cases, and five Steps Performed. These documents were transformed into 75 inter-document relationship pairs through a one-to-many relationship scheme. While this setting is sufficient to demonstrate the feasibility of the proposed reverse-engineering pipeline, it remains relatively small for drawing broad generalization claims.

The small dataset size may affect the reported model performance in several ways. First, the high accuracy, recall, F1-score, and ROC-AUC values may partly reflect the controlled structure of the IdVar4CL artifacts, where the documents follow consistent terminology, writing patterns, and artifact relationships. Second, because all data originate from the same system context, the model may learn case-specific semantic patterns rather than generalizable cross-domain relationships. Third, the one-to-many relationship construction increases the number of training pairs, but it does not fully replace the diversity that would be obtained from additional independent projects or systems. Therefore, the current validation results should be interpreted as within-case evidence rather than proof of general applicability.

Another potential impact is the risk of overestimating model robustness. Stratified K-Fold validation helps evaluate performance stability across the available pairs, but the folds are still derived from the same underlying document set. As a result, lexical overlap and repeated domain-specific terms across FR, UC, and SP artifacts may make the classification task easier than in a real cross-project scenario. This may explain why the model achieved perfect recall across all folds and ROC-AUC values of 1.0000. Although this result is beneficial for maintaining traceability coverage, it should be interpreted carefully because external validation has not yet been conducted.

To mitigate this limitation, future work should expand the dataset to include multiple software systems, diverse application domains, and more varied SRS/SDD writing styles. Cross-project validation can be performed by training the model on one or more systems and testing it on another system. Cross-system validation can also be used to evaluate whether the model can infer FR→UC→SP paths in unseen documentation contexts. In addition, future experiments should include more rigorous negative pair construction, hard negative sampling, and probability threshold calibration to reduce false positives and improve generalization. These extensions are necessary before the proposed pipeline can be claimed as broadly applicable beyond the IdVar4CL case study.