Update approaches

Author: dfuchss

_approaches/ardocode.md

@@ -14,5 +14,3 @@ In practice, it extracts key terms from the documentation and tries to align the
 
 - Key idea: Apply the SWATTR approach without an explicit SAM by interpreting the codebase as a model. For example, if the doc mentions a component "WebUI" and there is a WebUI package in code, ArDoCode will link them.
 - Effectiveness: Because it skips the formal modeling step, ArDoCode is easier to apply but less precise. In evaluations, ArDoCode achieved a weighted F1 of only ~0.62, substantially lower than the full TransArC method. It serves mainly as a baseline and demonstrates that without structured models, the TLR performance drops.
-
-See our [ICSE 2024 publication page](/c/icse24) for details, links, and resources.

_approaches/inconsistency-detection.md

@@ -18,5 +18,3 @@ The method runs a TLR procedure (namely SWATTR) and then flags any model element
 
 - Detection strategy: Use the TLR results as a bridge. After linking as many sentences to model elements as possible, any "orphan" model nodes or text mentions indicate a consistency gap. For example, if the model has a "Cache" component with no sentence linked, that is an UME; if the doc talks about "Common" but the model lacks it, that is an MME.
 - Results: The approach achieved an excellent F1 (0.81) for the underlying trace recovery. For inconsistency detection, it attained ~93% accuracy in identifying UMEs and ~75% for MMEs, significantly better than naive baselines. These results suggest that using trace links is a promising way to find documentation-model mismatches.
-
-See our [ICSA 2023 publication page](/c/icsa23) for details, links, and resources.

_approaches/lissa.md

@@ -15,5 +15,3 @@ It then queries the LLM with the retrieved context to generate or suggest the mo
 
 - Scope: LiSSA was tested on multiple tasks including requirements→code, documentation→code, and architecture-docs→models. The same RAG process is applied in each case, making it a one-size-fits-many solution.
 - Effectiveness: In experiments, LiSSA significantly outperformed state-of-the-art tools on the code-centric tasks. For example, it showed much higher accuracy when linking requirements to code than prior methods.
-
-LiSSA is primarily associated with our [ICSE 2025 publication page](/c/icse25), but is also related to our [REFSQ 2025 publication page](/c/refsq25). See these pages for details, links, and resources.

_approaches/swattr.md

@@ -16,5 +16,3 @@ Finally, it connects these identified text elements to model elements to form tr
 
 - Pipeline stages: The framework is extendable, meaning you can plug in different strategies at each step. For example, one agent might use term matching to find components in sentences, while another uses more advanced similarity measures. All results are aggregated to produce the final links.
 - Results: SWATTR was evaluated on three case studies and achieved a weighted average F1-score of about 0.72 for trace recovery. This was a strong performance (outperforming simple baselines by ~0.24 F1) and demonstrated the benefit of the multi-stage approach.
-
-See our [ECSA 2021 publication page](/c/ecsa21) for details, links, and resources.

_approaches/transarc.md

@@ -15,5 +15,3 @@ This two-step strategy helps bridge the semantic gap between informal text and c
 
 - How it works: TransArC extracts combines the two link sets of trace links, namely SWATTR and ArCoTL, to produce trace links transitively from documentation to code.
 - Results: In experiments on five systems, TransArC achieved a high average F1 score (~0.82) for recovering documentation-to-code links, significantly outperforming baseline methods. This shows that combining the two specialized steps yields much more accurate links than simpler approaches.
-
-See our [ICSE 2024 publication page](/c/icse24) for details, links, and resources.

_approaches/transarcai.md

@@ -16,5 +16,3 @@ The goal is to bridge the SAD–code gap without manual modeling.
 
 - How it works: Given the software architecture text and the codebase, the system asks the LLM to list likely component names. That list of names forms a "Simple Software Architecture Model" (SSAM). Finally, code elements with matching names or descriptions are linked to the documentation. This pipeline avoids needing an explicit UML model.
 - Effectiveness: TransArC-AI achieved very competitive results. Using GPT-4o, it obtained a weighted F1 of about 0.86, nearly as good as the original TransArC with a hand-made model (F1 0.87). It also substantially outperformed the ArDoCode baseline (which scored ~0.62). This shows that LLMs can automatically infer the key architectural components.
-
-See our [ICSA 2025 publication page](/c/icsa25) for details, links, and resources.