Search results

Figure 1 Overview of our multi-step RAG ensemble. Three answer candidates are generated via (a) Query2Doc-based with Wikipedia retrieval, (b) Rationale-based with Wikipedia retrieval, and (c) Web Search API applied. The Decision Maker then compares the candidates and selects the final answer pair or generate

Figure 2 Analysis of source preferences across different Decision Maker models. Model configuration for RAG stage: (a) Query2Doc and Rationale steps by GPT-5.2, Web Search by Gemini 3 Pro, (b) all RAG steps by Gemini 3 Pro. Two stacked bar charts showing the proportion of final answers attributed to each

Figure 2 Network Visualization Features. Node types include protein-coding genes (circles) and microRNAs (diamonds). Edge types indicate relationship types; thickness encodes confidence. Nodes use locus IDs as unique identifiers ensuring correct network topology. Screenshot of the GeneReL Network page wit

Figure 3 Database Composition. (a) Distribution of relationship types. (b) Confidence level distribution showing 90.2% high confidence. (c) Evidence type classification. Three charts. (A) Horizontal bar chart of relationship-type counts (n = 13,710), led by Regulates (3,344; 24.4 percent), Interacts (3,05

Figure 4 Comparison of GeneReL interactions with STRING and IntAct databases, restricted to matchable interactions where both genes are resolved to AGI locus IDs. (a) Gene ID resolution overview. Of 13,710 total GeneReL interactions, 6,569 (47.9%) have both genes resolved to AGI locus IDs and are matchable a

Figure 5 Interactive network views from the GeneReL Network page. Node size scales with degree; edge colors denote relationship types (e.g., Activates, Represses, Regulates), and the toolbar allows filtering by relationship type and confidence. (a) One-step view of the canonical flowering-time module seeded

Figure 1 GeneReL Pipeline Architecture and Gene Normalization Workflow. (a) Overview of the data extraction and curation pipeline showing data collection from PMC, tiered AI processing (Haiku screening, Sonnet extraction, Opus verification), five-step gene normalization, and web application interfaces. (b) F

Graphical Abstract Graphical Abstract For image description, please refer to the figure legend and surrounding text.

Figure 2 Taxonomic distribution of bacterial and fungal genomes by phylum. Bar charts show the number of genomes assigned to each phylum on a logarithmic scale. (a) Distribution of the 35 681 bacterial genomes, with Pseudomonadota, Bacillota, and Actinomycetota representing the most abundant phyla. (b) Distr

Graphical Abstract Graphical Abstract For image description, please refer to the figure legend and surrounding text.

Figure 8 UpSet plot analysing the trigger vocabulary composition across five data sources after the LLM enrichment step in dictionary construction. (1) The horizontal bars on the left represent the total vocabulary size of each dataset, comprising both translated vocabulary and generated aliases. (2) Similar

Figure 9 Effect of the similarity threshold on context-aware data augmentation. Bars indicate the number of generated training samples under each threshold after the first round of self-training, while lines show the F1 scores for trigger and argument extraction on the development set. All scores are reporte

Figure 3 Step-by-step workflow of the Search module. The interface allows users to construct complex, multi-layered queries. The numbered steps highlight the logical flow: (1) the user selects the main category. (2) A suitable sub-category is selected based on the intended screening strategy. (3) To deepen t

Figure 1 Qualitative and quantitative analysis of contextual scope in substance use NER. (a) Illustrative examples of heterogeneous contextual dependencies. (b) Aggregated token-level attention heatmap centred on the target mention (offset 0; grey band). The x-axis shows relative token offset (negative: prev

Figure 2 Overview of the I-NOWJ architecture. The framework consists of three main phases: adaptive context selection, knowledge-infused representation, and trigger–argument detection. For image description, please refer to the figure legend and surrounding text.