ChenAI2015/A-Dataset-for-Complex-Reasoning-over-Textual-Knowledge-Graphs-in-Medicine

RiTeK: Medical Textual Knowledge Graph QA Benchmark

RiTeK is a benchmark for complex reasoning over medical Textual Knowledge Graphs (medical TKGs). It evaluates whether retrieval systems and Large Language Models (LLMs) can answer realistic medical questions by using both relational paths and textual entity descriptions.

Dataset Overview

The benchmark contains three medical graph QA subsets:

Dataset	Directory	Splits	KG file
ADint	Adint	train / dev / test	ADint_KG.txt
pharmKG	pharmKG	train / dev / test	Pharm_KG.txt
Prime	prime	train / test	External PrimeKG file

The local note in Dataset/pharmKG/readme.txt says that only the train and test datasets provide hop1_neighbors.

Textual KG Statistics

TKG Dataset	# Entity Types	# Relation Types	# Triples	# Coverage
STaRK-Amazon	4	4	9,443,802	--
STaRK-Mag	4	4	39,802,116	--
STaRK-Prime	10	18	8,100,498	15.29%
RiTeK-PharmKG	3	29	500,958	95.61%
RiTeK-ADint	102	15	1,017,284	36.73%

# Coverage refers to the proportion of nodes with textual descriptions. # Entity Types denotes the number of entity types, and # Relation Types indicates the number of relation types. The textual information of the provided nodes is difficult to quantify statistically for STaRK-Mag and STaRK-Amazon, so their coverage statistics are not included.

QA Dataset Statistics

QA Dataset	# Queries	# Topological Structures	# Instance Rate	Train / Val / Test
STaRK-Amazon	9,100	1	4	0.65 / 0.17 / 0.18
STaRK-Mag	13,323	4	1.25	0.60 / 0.20 / 0.20
STaRK-Prime	11,204	3	9.3	0.55 / 0.20 / 0.25
RiTeK-PharmKG	10,235	6	11.33	0.80 / 0.10 / 0.10
RiTeK-ADint	5,322	6	9.67	0.80 / 0.10 / 0.10

# Instance Rate refers to the average number of relational templates per topological structure.

Access Benchmark Data

All dataset files are under main. Each QA file is JSONL: one JSON object per line.

ADint

File	Rows	Description
ADint_train_with_topic_entity.json	3,859	Original train QA data with topic entity metadata
Adint_dev_with_topic_entity.json	483	Dev QA data
ADint_test_with_topic_entity.json	482	Original test QA data with topic entity metadata
ADint_train_QA_instruction.json	3,859	Train data in instruction format
ADint_dev_QA_instruction.json	483	Dev data in instruction format
ADint_test_QA_instruction.json	482	Test data in instruction format
ADint_KG.txt	1,385,264	Knowledge graph triples

pharmKG

File	Rows	Description
pharmKG_train_QA_with_topic_entity.json	8,188	Original train QA data with topic entity metadata
pharmKG_dev_with_topic_entity.json	1,024	Dev QA data with topic entity metadata
pharmKG_test_QA_with_topic_entity.json	1,023	Original test QA data with topic entity metadata
pharmKG_train_QA_instruction.json	8,188	Train data in instruction format
pharmKG_dev_QA_instruction.json	1,024	Dev data in instruction format
pharmKG_test_QA_instruction.json	1,023	Test data in instruction format
Pharm_KG.txt	500,360	Knowledge graph triples

Prime

File	Rows	Description
Prime_train_QA_with_topic_entity.json	5,052	Original train QA data with topic entity metadata
Prime_test_QA_with_topic_entity.json	2,801	Original test QA data with topic entity metadata
Prime_train_QA_instruction.json	5,052	Train data in instruction format
Prime_test_QA_instruction.json	2,801	Test data in instruction format

PrimeKG graph resources are referenced in Dataset/prime/readme.

Data Format

Original QA Format

The original QA files use this JSONL format:

{"question_id": "2", "question": "What is the name of the condition characterized by a complete interruption of the inferior vena cava, falling under congenital vena cava anomalies?", "answer": {"98851": "azygos continuation of the inferior vena cava", "98853": "inferior vena cava interruption"}, "topic_entity": [{"topic_name": "congenital systemic veins anomaly", "topic_id": "39483"}]}

Fields:

Field	Description
question_id	Unique question identifier within the dataset
question	Natural-language question
answer	Mapping from answer entity ID to answer surface name
topic_entity	Topic entity or entities used to anchor the question in the KG

Instruction QA Format

The instruction files are converted JSONL files for instruction tuning or direct QA prompting:

{"instruction": "You are an excellent linguist. The task is to answer the question", "context": "What drugs are used to treat epithelioid sarcoma and also affect the EZH2 gene product?", "response": "Tazemetostat\n", "category": "Prime_open_QA"}

Fields:

Field	Description
instruction	Fixed task instruction
context	The question text
response	Gold answer text. Multiple answers are joined with commas and end with \n
category	Dataset-specific task category, such as ADint_open_QA, pharmKG_open_QA, or Prime_open_QA

Knowledge Graph Format

Each KG file is a plain-text triple file:

137690    1    34589
1         3    144772

Each line represents:

head_entity_id    relation_id    tail_entity_id

To obtain entity names, entity descriptions, relation names, or node types, use the metadata files from the linked node-edge information resources:

• [edge_type_dict.pkl](under Graph_node_edge_info files)
• [node_info.pkl](under Graph_node_edge_info files)
• [node_type_dict.pkl](under Graph_node_edge_info files)

Example relation mapping:

{
    0: "ppi",
    1: "carrier",
    2: "enzyme",
    3: "target",
    4: "transporter",
    5: "contraindication",
    6: "indication",
    7: "off-label use",
    8: "synergistic interaction",
    9: "associated with",
    10: "parent-child",
    11: "phenotype absent",
    12: "phenotype present",
    13: "side effect",
    14: "interacts with",
    15: "linked to",
    16: "expression present",
    17: "expression absent"
}

Example node type mapping:

{
    0: "disease",
    1: "gene/protein",
    2: "molecular_function",
    3: "drug",
    4: "pathway",
    5: "anatomy",
    6: "effect/phenotype",
    7: "biological_process",
    8: "cellular_component",
    9: "exposure"
}

Baselines

We evaluate the following baselines on the benchmark datasets.

Reasoning and Retrieval Baselines

Method	Reference Paper	Description
GPT-4	GPT-4 Technical Report	Uses GPT-4 with a direct-answer instruction to generate answers.
Random Walk	Random Walks on Graphs: A Survey	Starts from the topic entity and applies random walk to explore paths in the textual triple graph up to maximum depth d.
MCTS	Monte-Carlo Tree Search	Starts from the topic entity and applies Monte Carlo tree search to explore paths in the textual triple graph up to maximum depth d. In this work, d = 3.
Chain-of-Thought (COT)	Chain-of-Thought Prompting Elicits Reasoning in Large Language Models	Guides GPT-4 to generate answers step by step, with output formatted as step-by-step reasoning: explanation, answer: medical terms.
Tree-of-Thought (TOT)	Tree of Thoughts: Deliberate Problem Solving with Large Language Models	Structures reasoning as a tree search where multiple intermediate reasoning paths are explored in parallel. GPT-4 evaluates and expands promising paths using voting or scoring.
Graph-of-Thought (GOT)	Graph of Thoughts: Solving Elaborate Problems with Large Language Models	Represents reasoning as a graph where nodes capture reasoning states and edges denote transitions. GPT-4 traverses the graph to aggregate information and synthesize the final answer.
Think-on-Graph (TOG)	Think-on-Graph: Deep and Responsible Reasoning of Large Language Model with Knowledge Graph	Enables LLMs to interactively perform beam search over knowledge graphs and evaluate promising reasoning paths without additional training.
G-retriever	G-retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering	Retrieves query-relevant subgraphs using the Prize-Collecting Steiner Tree (PCST) algorithm to enhance LLM reasoning over textual graphs.
KAR	Knowledge-Aware Query Expansion with Large Language Models for Textual and Relational Retrieval	Augments LLMs with structured document relations from a knowledge graph and uses relation-aware filtering to improve retrieval for semi-structured queries.

Supervised Retrieval Baselines

We also evaluate three representative retrieval models on our benchmark datasets and STaRK-Prime under supervised learning settings.

Method	Reference Paper	Description
G-retriever	G-retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering	Retrieves query-relevant subgraphs using the Prize-Collecting Steiner Tree (PCST) algorithm to enhance LLM understanding and reasoning over textual graphs.
GCR	Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models	Uses graph-constrained reasoning with document-based relational signals to improve retrieval for semi-structured queries.
GNN-RAG	GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning	Uses a GNN to retrieve relevant answers and extract shortest paths between the topic entity and answers, then verbalizes the paths for retrieval-augmented generation.

Evaluation

See Evaluation/README.md for the full evaluation workflow.

1. Prepare the Gold File

Use the test split of the dataset you want to evaluate, for example:

Adint/ADint_test_with_topic_entity.json
pharmKG/pharmKG_test_QA_with_topic_entity.json
prime/Prime_test_QA_with_topic_entity.json

2. Prepare the Prediction File

The recommended prediction format is:

{"id": "0", "answers": ["sebaceous adenoma"]}
{"id": "1", "answers": ["EDN1", "EDN2"]}

The evaluator also supports instruction files with response and legacy prediction files with predicted_answers.