結構化與非結構化輸入於大型語言模型中之比較：以異常事件預測任務中的語意與語用預測能力評估為例

Abstract

大型語言模型（LLMs）近年被廣泛應用於時間知識圖譜（Temporal Knowledge Graph, TKG）預測任務中，作為傳統圖神經網路方法的泛化替代方案。特別是在異常事件預測這類仰賴時間推理與語用合理性的任務中，模型需要理解一連串行為在時間軸上的邏輯與變化，因此如何提供具備結構化時序資訊的輸入成為關鍵挑戰。然而，目前仍不清楚 LLMs 在此任務中，究竟是依賴結構化的 TKG 輸入效果較佳，或是非結構化的自然語言描述表現更好。本研究比較兩種輸入格式（TKG → Text 與 Text → Text）對於語意一致性與語用推理能力的影響。

TKG 以主詞－關係－受詞的三元組表示事件，並附有明確的時間區間，其結構化格式能提升資訊密度、降低語言歧義，並幫助模型掌握事件順序與語用邏輯，特別適合異常行為偵測等需要精確推理的任務。相對地，非結構化的自然語言輸入雖然更接近 LLM 的預訓練分布，也能提供豐富語境，但容易引入冗詞與語意聯想雜訊。

本研究實驗採用 UCA 資料集，該資料集包含具時間標註的影片說明文字與異常事件標籤，為評估語意生成與語用異常判斷能力提供了合適的實驗資料集。此研究分為兩部分進行評估。首先，使用 BGE 嵌入模型計算 LLM 預測輸出與真實描述之間的餘弦相似度。結果顯示，非結構化輸入的語意對齊表現略高（平均值 0.5978），而結構化輸入亦表現接近（平均值 0.5718），顯示即便上下文減少，TKG 輸入亦能維持語意理解能力。

第二部分實驗中，此研究將不同輸入格式作為上下文，判斷當前幀是否異常，並以 AUC 作為評估指標。結果顯示：使用 TKG 作為時間上下文可提升異常判斷準確性，AUC 分數優於原始文本與摘要文本輸入，突顯其在語用推理上的穩健性。

整體而言，非結構化輸入有利於語意連貫與表達，結構化輸入則在語用推理上展現出更高穩健性，特別是在異常事件預測這類高度仰賴時間順序與行為合理性的任務中，TKG 輸入可提供更具邏輯性的語用支撐。本研究結果突顯兩者在上下文學習（in-context learning, ICL）任務上的互補潛力，亦為未來混合式提示輸入設計提供了方向，進一步提升模型對時序事件之語用理解能力。

Large language models (LLMs) have recently been applied to temporal knowledge graph (TKG) forecasting tasks, offering a generalizable alternative to traditional graph-based approaches. Particularly in abnormal event forecasting, where temporal reasoning and pragmatic coherence are essential, models must interpret the logical sequence and behavioral dynamics of actions over time. This raises a key challenge: how to provide input representations that encode structured temporal information. However, it remains unclear whether LLMs perform better in such tasks when guided by structured TKG inputs or by unstructured natural language descriptions. This thesis compares the effectiveness of structured (TKG → Text) and unstructured (Text → Text) inputs in forecasting abnormal events, with a focus on both semantic alignment and pragmatic reasoning.

TKGs represent events using subject–relation–object triples with explicit temporal spans. This structured format provides high information density, reduces linguistic ambiguity, and facilitates reasoning over temporal and causal dependencies—features that are particularly beneficial in tasks such as abnormal event detection. In contrast, unstructured textual descriptions offer rich contextual information and better alignment with LLMs’ pretraining data, but may introduce interpretive noise due to overgeneralization.

Experiments were conducted on the UCA (UCF Crime Annotation) dataset, which contains temporally annotated video captions and ground-truth anomaly labels, serving as a suitable benchmark for evaluating both semantic generation and pragmatic anomaly detection. A two-part evaluation was performed. First, cosine similarity between the LLM-generated and ground-truth descriptions was computed using the BGE embedding model. Results show that unstructured inputs yield slightly higher semantic alignment (mean = 0.5978) compared to structured inputs (mean = 0.5718), suggesting that LLMs remain effective even with reduced contextual cues.

In the second experiment, different input formats were used as temporal context to identify abnormal video frames. Using AUC as the evaluation metric, structured TKG inputs outperformed raw and summarized text, demonstrating stronger grounding and robustness in pragmatic reasoning.

Overall, while unstructured inputs promote semantic fluency, structured representations, such as TKGs, enhance logical grounding, particularly in abnormal event forecasting tasks that rely heavily on temporal reasoning and pragmatic coherence. These findings highlight the potential of hybrid input designs—within text-based in-context learning setups—to combine structural clarity with contextual richness, enabling more accurate and pragmatically coherent reasoning in temporal event forecasting.