A Mechanism and Optimization Study on the Impact of Information Density on User-Generated Content Named Entity Recognition

Named Entity Recognition (NER) models trained on clean, high-resource corpora exhibit catastrophic performance collapse when deployed on noisy, sparse User-Generated Content (UGC), such as social media. Prior research has predominantly focused on point-wise symptom remediation -- employing customized fine-tuning to address issues like neologisms, alias drift, non-standard orthography, long-tail entities, and class imbalance. However, these improvements often fail to generalize because they overlook the structural sparsity inherent in UGC. This study reveals that surface-level noise symptoms share a unified root cause: low Information Density (ID). Through hierarchical confounding-controlled resampling experiments (specifically controlling for entity rarity and annotation consistency), this paper identifies ID as an independent key factor. We introduce Attention Spectrum Analysis (ASA) to quantify how reduced ID causally leads to ``attention blunting,'' ultimately degrading NER performance. Informed by these mechanistic insights, we propose the Window-Aware Optimization Module (WOM), an LLM-empowered, model-agnostic framework. WOM identifies information-sparse regions and utilizes selective back-translation to directionally enhance semantic density without altering model architecture. Deployed atop mainstream architectures on standard UGC datasets (WNUT2017, Twitter-NER, WNUT2016), WOM yields up to 4.5\% absolute F1 improvement, demonstrating robustness and achieving new state-of-the-art (SOTA) results on WNUT2017.

Understanding the Mechanisms Behind Structural Influences on Link Prediction: A Case Study on FB15k-237

FB15k-237 mitigates the data leakage issue by excluding inverse and symmetric relationship triples, however, this has led to substantial performance degradation and slow improvement progress. Traditional approaches demonstrate limited effectiveness on FB15k-237, primarily because the underlying mechanism by which structural features of the dataset influence model performance remains unexplored. To bridge this gap, we systematically investigate the impact mechanism of dataset structural features on link prediction performance. Firstly, we design a structured subgraph sampling strategy that ensures connectivity while constructing subgraphs with distinct structural features. Then, through correlation and sensitivity analyses conducted across several mainstream models, we observe that the distribution of relationship categories within subgraphs significantly affects performance, followed by the size of strongly connected components. Further exploration using the LIME model clarifies the intrinsic mechanism by which relationship categories influence link prediction performance, revealing that relationship categories primarily modulate the relative importance between entity embeddings and relationship embeddings and relationship embeddings, thereby affecting link prediction outcomes. These findings provide theoretical insights for addressing performance bottlenecks on FB15k-237, while the proposed analytical framework also offers methodological guidance for future studies dealing with structurally constrained datasets.