Zhuoqun Wei

← Projects

Online truth discovery for knowledge graphs

M.S. thesis on keeping a knowledge graph accurate in real time when sources disagree about the same fact.

Stack
Python, LLMs, Knowledge graphs

Semantic-Augmented Online Truth Discovery for Real-Time Knowledge Graph Maintenance. M.S. thesis at San Jose State University, advised by Dr. Ching‑Seh Wu.

When several sources describe the same product, they often disagree. A knowledge graph that ingests them continuously has to decide which value to believe as new claims arrive, without paying for a large-model call on every conflict.

Illustrative example. Three sources describe the same product's weight as 250 grams, 0.25 kilograms, and 2.5 kilograms. The system has to decide which value is true as claims arrive, and the knowledge graph keeps weight equals 250 grams. Source Aweight: 250 gSource Bweight: 0.25 kgSource Cweight: 2.5 kgWhich value is true?decided as claims arriveKnowledge graphweight = 250 g
The problem. An illustrative conflict. Two sources agree in different units and one is off by a factor of ten; the graph has to settle on one value as the claims stream in.

Results

Evaluated on 4,177 product-data conflicts from the WDC Product Corpus and compared with brute-force LLM routing.

Accuracy

higher is better

This work
73.9%
Brute-force LLM routing
67.6%

Evaluation cost

lower is better

This work
$2.27
Brute-force LLM routing
$4.18

LLM calls

lower is better · baseline = 100

This work
54
Brute-force LLM routing
100
Better and cheaper. Accuracy, evaluation cost, and LLM calls compared with brute-force LLM routing in the tested setup. LLM calls are indexed so the baseline is 100.

The implementation is validated with 135 unit tests.