Neural Query Engines for Knowledge Hypergraphs: Complex Query Answering with Transformers

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "Neural Query Engines for Knowledge Hypergraphs: Complex Query
Answering with Transformers"

by

Mr. Hong Ting TSANG


Abstract:

Complex Query Answering (CQA) over knowledge graphs has been extensively 
studied, yet existing methods are predominantly designed for binary 
relations, where each edge connects exactly two entities. Real-world 
knowledge, however, is inherently n-ary, and forcing such facts into binary 
triplets fragments their semantics. While knowledge hypergraphs address this 
by modeling n-ary relations as hyperedges that connect arbitrarily many 
entities, complex query answering over such structures remains unexplored.

To bridge this gap, we introduce two new benchmark datasets, JF17k-HCQA and 
M-FB15k-HCQA, covering 14 query types with projection, negation, conjunction, 
and disjunction operations. We propose the Logical Knowledge Hypergraph 
Transformer (LKHGT), a two-stage transformer in which a Projection Encoder 
resolves atomic hyperedge predictions and a Logical Encoder composes them 
through conjunction, disjunction, and negation. To let the model tell the 
eight token types apart during self-attention, we equip both encoders with a 
Type Aware Bias (TAB). On our two benchmarks LKHGT is the strongest method 
overall, beating the HLMPNN, LSGT, and NQE baselines and still answering 
query types it never saw in training, including four-hop queries deeper and 
wider than any seen during training. Two ablation findings stand out: 
absolute positional encoding matters for ordered hyperedges, and a learned 
transformer can take over the logical operations normally left to fuzzy 
logic, sidestepping the matrix blow-up those methods suffer as relation arity 
grows.


Date:                   Friday, 7 August 2026

Time:                   11:00am - 1:00pm

Venue:                  Room 3494
                        Lift 25/26

Chairman:               Prof. Song GUO

Committee Members:      Dr. Yangqiu SONG (Supervisor)
                        Dr. Chaojian LI