DSpace collection: 博碩士論文

以連結廣度和群聚性探究大學課程網路的關聯結構演化;以連結廣度和群聚性探究大學課程網路的關聯結構演化

title: 以連結廣度和群聚性探究大學課程網路的關聯結構演化;以連結廣度和群聚性探究大學課程網路的關聯結構演化 abstract: 本研究以國立中央大學 93 至 111 學年度學士班學生選課資料為基礎，建構課程共修網路（Course Co-enrollment Network），探討大學課程於課程體系中的結構角色及長期演化。相較於過去研究多著重課程先修關係或單一時間點分析，本研究結合課程角色分類、純選修與全課程網路比較，以及長期時序分析，呈現課程結構之動態變化。研究方法以度數中心性（Degree Centrality， DC）衡量課程共修連結廣度，以群聚係數（Clustering Coefficient， CC）衡量修課組合集中程度，並以兩項指標之年度中位數作為門檻，建立 DC×CC 四象限課程角色分類，將課程區分為熱門樞紐（Q1）、跨域橋樑（Q2）、社群內核心（Q3）及專向發展（Q4）。研究結果顯示，111 學年度純選修課程網路共包含 1，250 門課程，DC 與 CC 呈顯著負相關（?

基於黑箱網站探索之端對端測試自動化生成：路徑導向任務萃取與證據導向完成性判定方法;Automated End-to-End Test Generation via Black-Box Web Application Exploration: Path-Based Task Extraction and Evidence-Based Completion Verification

title: 基於黑箱網站探索之端對端測試自動化生成：路徑導向任務萃取與證據導向完成性判定方法;Automated End-to-End Test Generation via Black-Box Web Application Exploration: Path-Based Task Extraction and Evidence-Based Completion Verification abstract: 近年來，大型語言模型與 Web Agent 技術逐漸被應用於端對端（end-to-end， E2E）網站測試，使自動化流程由傳統腳本撰寫進一步延伸至網站功能推論、測試案例生成與自然語言任務執行。若希望將自動化範圍進一步延伸至測試案例建立之前的網站理解階段，則不能假設完整需求文件、既有測試案例、內部程式資訊或人工預先定義的 E2E 任務一定會作為必要測試輸入。在此設定下仍存在三項前置問題：黑箱探索結果多以頁面狀態、事件轉移與操作路徑呈現，尚未直接形成使用者層級的 E2E 任務； Web Agent 測試通常假設自然語言任務已經存在；既有 LLM-based E2E test generation 則主要關注功能推論與測試案例生成，而非確認任務在目前網站版本中是否實際達到完成狀態。因此，本研究提出一套從黑箱網站自動盤點 E2E 任務，並判定其前端可觀察完成狀態的方法。網站入口 URL 僅作為啟動黑箱探索的定位資訊，並可選擇提供測試帳號或既有登入狀態；後續任務建立所需的功能資訊，則由系統在實際探索受測網站時動態取得。系統首先使用 Playwright 蒐集 URL ／ DOM 狀態、可互動元素、表單 affordance、操作前後變化與探索路徑；接著以探索路徑為分析單位，透過大型語言模型逐步抽取包含使用者目標與必要步驟的候選 E2E 任務，並進行跨路徑合併與去重。任務建立後，系統依任務步驟與既有探索證據的對齊程度進行分流：證據充分者直接進入完成性判定，證據不足者才交由 Web Agent 補充執行。最後，系統整合爬蟲與 Web Agent 的結構化互動證據，將任務判定為 pass、partial、fail 或 blocked，並輸出受影響或受阻步驟、失敗原因與可追溯 evidence pack。系統亦可依任務步驟產生初步 Playwright pytest 腳本骨架，但該腳本僅作為人工後續修訂的輔助輸出。本研究以 PrestaShop、Spring PetClinic 與 Conduit 三個開源網站進行評估。任務盤點實驗中，本研究方法的 macro-average Precision、Recall 與 F1 分別為 0.8646、0.8618 與 0.8543，高於 AutoE2E 基準方法的 0.6952、0.6306 與 0.6519。正常版本完成性判定 i 實驗共評估 144 個 E2E 任務，其中 111 個任務被判定為 pass，Pass Rate 為 77%；透過證據分流，僅 64 個任務需要 Web Agent 補驗，相較於所有任務皆交由 Agent 執行，可減少 55.6% 的 Web Agent 執行需求。故障注入實驗共建立十二個前端可觀察故障案例，系統在其中八個案例找到受影響任務，並於五個案例正確回報 fail。結果顯示，本研究可在設定的黑箱探索範圍內，將前端探索證據整理為可追溯的 E2E 任務集合，並提供任務層級的前端可觀察完成性判定；其限制主要在於未被探索的功能與無法由前端觀察之後端或純視覺正確性。;Recent advances in large language models and Web Agents have extended end-to-end (E2E) web testing automation beyond conventional script authoring toward feature inference， test-case generation， and natural-language task execution. To further automate the application- understanding stage that precedes test-case authoring， a testing workflow cannot assume that complete requirements， existing test cases， internal implementation information， or manually predefined E2E tasks are always available as mandatory inputs. Under this setting， three is- sues remain. First， black-box exploration primarily produces page states， event transitions， and navigation paths rather than user-level E2E tasks. Second， Web-Agent-based testing generally assumes that a natural-language task or test case has already been defined. Third， existing LLM- based E2E test generation focuses primarily on feature inference and test generation rather than verifying whether a generated task can actually reach a completion state in the current applica- tion version. This study therefore proposes a method for automatically inventorying E2E tasks from a black-box web application and determining their frontend-observable completion states. The website entry URL serves only as the locator that initiates exploration， with an optional test account or authenticated browser state; the functional information required for subsequent task construction is acquired dynamically by interacting with the system under test. Playwright is first used to collect URL/DOM states， interactive elements， form affordances， before/after interaction changes， and navigation paths. An LLM then progressively extracts candidate E2E tasks from each exploration path， representing each task with a user goal and essential steps， followed by cross-path semantic merging and deduplication. After task construction， the system triages tasks according to the alignment between task steps and existing exploration evidence. Tasks with sufficient evidence proceed directly to completion verification， whereas only evidence- insufficient tasks are assigned to a Web Agent for supplementary execution. Finally， the system integrates structured crawler and Web Agent evidence to classify each task as pass， partial， fail， or blocked， and reports affected or blocked steps， failure reasons， and a traceable evidence pack. Preliminary Playwright pytest script skeletons are also derived from task steps， but they are treated only as auxiliary outputs that require subsequent manual review and refinement. The proposed method is evaluated on three open-source web applications: PrestaShop， Spring PetClinic， and Conduit. For E2E task inventory， the proposed method achieves macro- average Precision， Recall， and F1 scores of 0.8646， 0.8618， and 0.8543， respectively， compared with 0.6952， 0.6306， and 0.6519 for the AutoE2E baseline under the same task-level alignment protocol. In the normal-version completion-verification experiment， 111 of 144 tasks are clas- sified as pass， corresponding to a 77% Pass Rate. Evidence-based triage sends only 64 of the 144 tasks to the Web Agent， reducing Web Agent execution demand by 55.6% relative to exe- cuting the Agent for every task. In twelve controlled frontend-observable fault-injection cases， the system discovers affected tasks in eight cases and correctly reports fail in five cases. These results show that， within the configured black-box exploration scope， the proposed method can transform frontend exploration evidence into a traceable E2E task inventory and provide task- level frontend-observable completion verdicts. Its main limitations are functions that are not reached during exploration and correctness properties that cannot be established from frontend- observable evidence alone.

Data-Purity-Aware Contrastive Gradient Sensitivity for Efficient Medical Domain Adaptation of Large Language Models;Data-Purity-Aware Contrastive Gradient Sensitivity for Efficient Medical Domain Adaptation of Large Language Models

title: Data-Purity-Aware Contrastive Gradient Sensitivity for Efficient Medical Domain Adaptation of Large Language Models;Data-Purity-Aware Contrastive Gradient Sensitivity for Efficient Medical Domain Adaptation of Large Language Models abstract: 將大型語言模型（LLMs）適應於醫療領域對於臨床決策支援至關重要。然而，傳統持續預訓練（CPT）面臨高昂計算成本與災難性遺忘，後者對必須處理多樣指令的臨床安全構成威脅。為此，我們提出一個高效的領域適應框架，包含兩項核心貢獻：(1) 對比式梯度敏感度（Contrastive Gradient Sensitivity， CGS），透過比較醫療與通用語料的梯度響應，自動更新最具領域鑑別力的模組，以在保護通用能力的同時最大化醫療知識吸收；(2) 困惑度導向的資料純度過濾，以精煉訓練語料。我們在 Llama-3.1-8B、Qwen2.5-7B 與 Mistral-7B-v0.3 上進行評估。在僅更新 8% 參數的條件下，CGS 在 MedQA 上均達到與 LISA 基線相當或略優的表現。在 Llama-3.1-8B 上，CGS 取得 63.50% 的 MedQA 準確率，達到與全參數微調的 Meditron3-8B（63.39%）相當的水準，且 MMLU 通用能力幾乎無損（68.47% vs. 原始 68.54%）。自動評估框架與人類一致性達 Cohen’s $\kappa = 1.00$ ($n = 200$)。消融實驗支持了 CGS 定向更新與資料過濾在平衡領域適應與臨床安全上的貢獻。;Adapting large language models (LLMs) to the medical domain is crucial for clinical decision support. However， standard continual pre-training (CPT) to inject clinical knowledge often suffers from prohibitive computational costs and catastrophic forgetting of general capabilities—a critical safety risk for clinical LLMs that must comprehend diverse， general instructions. To address this， we propose an efficient medical domain adaptation framework with two core contributions:(1) Contrastive Gradient Sensitivity (CGS)， a gradient-informed parameter selection strategy that compares gradient responses across medical and general corpora to identify and update the most domain-discriminative modules， protecting general capabilities while maximizing medical knowledge absorption; and(2) Perplexity-guided data purity filtering to refine the training corpus.We evaluate our framework on Llama-3.1-8B， Qwen2.5-7B， and Mistral-7B-v0.3 across major medical benchmarks (MedQA， MedMCQA， PubMedQA) and general tasks (MMLU). Updating only 8% of parameters， CGS achieves comparable or slightly higher MedQA accuracy than the LISA baseline across all models. On Llama-3.1-8B， CGS achieves 63.50% on MedQA—achieving comparable performance to the fully pre-trained Meditron3-8B (63.39%)—while preserving near-intact general capabilities on MMLU (68.47% vs. 68.54% base). Our automatic evaluation pipeline is validated by a human agreement of Cohen’s $\kappa = 1.00$ ($n = 200$). Ablation studies support the importance of CGS’s targeted parameter updates and perplexity-based filtering in balancing domain knowledge absorption and safety-critical capability preservation.

Low-Latency Web Visualization of Georeferenced Wind Fields and Radar Cross-Sections with Viewport-Indexed Loading;Low-Latency Web Visualization of Georeferenced Wind Fields and Radar Cross-Sections with Viewport-Indexed Loading

title: Low-Latency Web Visualization of Georeferenced Wind Fields and Radar Cross-Sections with Viewport-Indexed Loading;Low-Latency Web Visualization of Georeferenced Wind Fields and Radar Cross-Sections with Viewport-Indexed Loading abstract: 在網頁瀏覽器中互動探索三維大氣資料，必須在有限的運算資源下同時協調資料檢索、網路傳輸、瀏覽器記憶體、幾何生成與渲染，因此仍具挑戰性。本研究與國家災害防救科技中心合作，提出一套瀏覽器式、具地理參考的風場與雷達資料視覺化系統。系統在共同的地理座標框架中，整合三維風場流線與由使用者指定之雷達反射率垂直剖面，並不要求瀏覽器直接載入或渲染完整雷達體積。為降低不必要的資料存取，本研究將大氣資料切分為空間資料區塊，並組織成具有空間、時間與解析度中繼資料的多解析度索引紀錄。系統依據目前地圖視埠所涵蓋的範圍，在網路傳輸前選取所需的空間與時間紀錄，再由瀏覽器完成資料解碼、流線生成、雷達剖面幾何建構與地理對位渲染。此外，本研究在受控的瀏覽器工作負載下評估一組預先定義的視覺化與取樣參數，並以最低量測互動延遲選取保留的操作設定，同時報告其瀏覽器記憶體使用量。在本研究的實驗環境中，所提出的流程將完整資料範圍基準的互動延遲由 3552 毫秒降低至 9.8 毫秒，渲染迴圈速率由 11.5 FPS 提升至 134.0 FPS，JavaScript heap 使用量則由 552 MB 降低至 56 MB。代表性的視埠查詢僅傳輸 2 個風場資料區塊、共 1.05 MiB，相較完整 LOD0 風場資料的 48 個區塊與 24.20 MiB，資料量約降低 95.67%。在保留設定 C12 的 30 次大範圍平移測試中，視埠更新平均延遲為 71.6 毫秒，P95 為 79.1 毫秒，最大值為 82.1 毫秒；上述數值不包含地圖移動結束後的 120 毫秒 debounce。結果顯示，以視埠為索引的資料存取與地理對位渲染，可在受測條件下支援區域性大氣資料的低延遲網頁互動探索。;Interactive exploration of three-dimensional atmospheric data in web browsers is challenging because data retrieval， network transfer， browser memory， geometry generation， and rendering must be coordinated under limited computational resources. We present a browser-based visualization system developed in collaboration with Taiwan’s National Science and Technology Center for Disaster Reduction (NCDR) to address requirements for geographically contextualized exploration of wind and radar data. The system provides a complementary three-dimensional view that co-registers wind streamlines and user-defined radar-reflectivity vertical cross-sections within a common geographic reference frame， without requiring direct full-volume rendering in the browser. Atmospheric datasets are partitioned into spatial chunks and organized as metadata-indexed multiresolution records. A viewport-driven region-of-interest strategy selects the spatial and temporal records required for the current map view before network transmission. We further evaluate a predefined set of visualization and sampling configurations and retain the configuration with the lowest recorded interaction latency while reporting its browser-memory usage. In the evaluated environment， the proposed pipeline reduces the recorded interaction latency of the full-domain baseline from 3552 ms to 9.8 ms. The rendering-loop rate increases from 11.5 to 134.0 frames per second， while JavaScript heap consumption decreases from 552 MB to 56 MB. A representative viewport request transfers 1.05 MiB from two wind chunks， compared with 24.20 MiB from 48 chunks for the complete LOD0 wind domain. For the retained configuration， the measured viewport-update latency has a mean of 71.6 ms， a P95 of 79.1 ms， and a maximum of 82.1 ms over 30 large-pan trials; these values exclude the 120 ms debounce interval. These results demonstrate the technical feasibility of viewport-indexed data access and geographically co-registered rendering for interactive， browser-based exploration of regional atmospheric datasets.