sub:assertion {
<
https://doi.org/10.48550/arXiv.2404.14741>
dct:title "Generate-on-Graph: Treat LLM as both Agent and KG for Incomplete Knowledge Graph Question Answering" ;
<
http://purl.org/spar/cito/describes> <
https://neverblink.eu/ontologies/llm-kg/methods#GenerateOnGraph> ;
<
http://purl.org/spar/cito/discusses> <
https://neverblink.eu/ontologies/llm-kg/methods#ChainOfThoughtPrompting> , <
https://neverblink.eu/ontologies/llm-kg/methods#ChatKBQA> , <
https://neverblink.eu/ontologies/llm-kg/methods#DecomP> , <
https://neverblink.eu/ontologies/llm-kg/methods#GAG> , <
https://neverblink.eu/ontologies/llm-kg/methods#KBBinder> , <
https://neverblink.eu/ontologies/llm-kg/methods#ReAct> , <
https://neverblink.eu/ontologies/llm-kg/methods#RoG> , <
https://neverblink.eu/ontologies/llm-kg/methods#SelfConsistency> , <
https://neverblink.eu/ontologies/llm-kg/methods#StandardPrompting> , <
https://neverblink.eu/ontologies/llm-kg/methods#StructGPT> , <
https://neverblink.eu/ontologies/llm-kg/methods#ToG> ;
a prov:Entity .
<
https://neverblink.eu/ontologies/llm-kg/methods#ChainOfThoughtPrompting>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This method is used as a baseline for comparison. It aims to elicit complex reasoning from LLMs by encouraging them to generate intermediate reasoning steps, providing a stronger LLM-only baseline than standard prompting." ;
rdfs:label "Chain-of-Thought (CoT) prompting" .
<
https://neverblink.eu/ontologies/llm-kg/methods#ChatKBQA>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This is a semantic parsing (SP) method for KGQA, used as a baseline. Similar to KB-BINDER, it converts natural language questions into executable queries for KGs, and its limitations under incomplete KGs are highlighted." ;
rdfs:label "ChatKBQA" .
<
https://neverblink.eu/ontologies/llm-kg/methods#DecomP>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This method is cited as a motivation for GoG's approach to problem decomposition. De-comP solves complex tasks by breaking them down into simpler sub-tasks, contributing to the structured reasoning in GoG." ;
rdfs:label "De-comP" .
<
https://neverblink.eu/ontologies/llm-kg/methods#GAG>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This method is mentioned as being similar to GoG, particularly in its utilization of the knowledge modeling ability of Large Language Models. It serves as a point of reference for existing techniques that leverage LLMs for knowledge tasks." ;
rdfs:label "GAG" .
<
https://neverblink.eu/ontologies/llm-kg/methods#GenerateOnGraph>
dct:subject <
https://neverblink.eu/ontologies/llm-kg/categories#SynergizedReasoning> ;
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "Generate-on-Graph (GoG) is a training-free method designed for Incomplete Knowledge Graph Question Answering (IKGQA). It uses a Thinking-Searching-Generating framework where the LLM acts as both an agent to explore KGs (Searching) and a KG to generate new factual triples (Generating) when knowledge is missing. This process represents a synergistic reasoning approach, dynamically combining LLM's internal knowledge and external KG facts." ;
rdfs:label "Generate-on-Graph" ;
<
https://neverblink.eu/ontologies/llm-kg/hasTopCategory> <
https://neverblink.eu/ontologies/llm-kg/top-categories#SynergizedLLMKG> .
<
https://neverblink.eu/ontologies/llm-kg/methods#KBBinder>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This is a semantic parsing (SP) method for KGQA, used as a baseline. It transforms natural language questions into structured queries (e.g., SPARQL) which are then executed on the KG. Its performance is evaluated especially under the incomplete KG scenario." ;
rdfs:label "KB-BINDER" .
<
https://neverblink.eu/ontologies/llm-kg/methods#ReAct>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This method is cited as a motivation for GoG's \"Thinking\" component. ReAct treats LLMs as agents that interact with an environment and make decisions to retrieve information, inspiring GoG's interactive framework." ;
rdfs:label "ReAct" .
<
https://neverblink.eu/ontologies/llm-kg/methods#RoG>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This is a Retrieval Augmented (RA) method, specifically noted as the state-of-the-art among fine-tuned models for KGQA, and is used as a strong baseline for comparison. It generates relation paths as faithful plans to retrieve reasoning paths from KGs." ;
rdfs:label "RoG" .
<
https://neverblink.eu/ontologies/llm-kg/methods#SelfConsistency>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This method is used as a baseline for comparison. It is a technique for improving the robustness of LLM reasoning by sampling multiple reasoning paths and taking a majority vote to arrive at a more reliable answer." ;
rdfs:label "Self-Consistency (SC)" .
<
https://neverblink.eu/ontologies/llm-kg/methods#StandardPrompting>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This method is used as a baseline for comparison against the proposed Generate-on-Graph (GoG) method. It represents a simple, direct prompting approach to question answering without explicit KG integration beyond basic retrieval." ;
rdfs:label "Standard Prompting (IO prompt)" .
<
https://neverblink.eu/ontologies/llm-kg/methods#StructGPT>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This is a Retrieval Augmented (RA) method for KGQA, used as a baseline for comparison. It retrieves information from KGs to guide LLMs in generating answers, and its performance is assessed under both complete and incomplete KG settings." ;
rdfs:label "StructGPT" .
<
https://neverblink.eu/ontologies/llm-kg/methods#ToG>
a <
http://purl.org/spar/fabio/Workflow> ;
rdfs:comment "This is a Retrieval Augmented (RA) method used as a key baseline for comparison. ToG treats the LLM as an agent to interactively explore relation paths on KGs, and its limitations are thoroughly analyzed and contrasted with the proposed GoG method, particularly in incomplete KG scenarios." ;
rdfs:label "ToG" .
}