DRAGON: Conversational Entity Retrieval from Knowledge Graphs

Accepted at WSDM 2026 | Paper

Authors: Mona Zamiri, Alexander Kotov (Wayne State University)

Neural ranking architecture for conversational entity retrieval from knowledge graphs. DRAGON aggregates fine-grained relevance signals using Graph Convolutional Networks and multi-head attention.

Model Configuration & Training Details

The following hyperparameters and architectural settings were used in all experiments reported in the paper:

Sub-graph Pruning

The pruning threshold λ for candidate entity sub-graph pruning was set to 0.5.

The coefficient λ′ in the feature weighting function was set to 3 × 10⁻⁴.

Optimization

All neural architectures were optimized using the AdamW optimizer.

Learning rate: 2.5 × 10⁻⁵

DRAGON Architecture

Hidden dimension of all convolutional layers: 32

Number of attention layers: 8

Number of GCN layers: 2

Number of fully connected layers: 6

The aggregated relevance vector for each candidate response entity is passed through six fully connected layers to produce the final ranking score.

---

Quick Start

Installation

pip install -r requirements.txt

Dependencies: PyTorch 2.4.1, DGL 2.4.0, Transformers 4.46.2

Training Pipeline (Step-by-Step)

Execute the following scripts in order:

Step 1: Convert Triples to Dictionary Format

python convert_triple_to_dict.py \
  --input_tsv /path/to/triples.tsv \
  --output_json ./data/dict_train.json

Step 2: Train Sub-graph Pruning Model (Optional)

python Pruning_train.py \
  --gpu 0 \
  --out_path ./models/pruning_model

Step 3: Calculate Fine-grained Features

python Calculate_features.py \
  --gpu 0 \
  --output_file ./data/features_train.json

Step 4: Train DRAGON Ranking Model

python GCN_train.py \
  --gpu 0 \
  --score_file_dir /path/to/scores/ \
  --triple_file_dir /path/to/triples/ \
  --model_path ./models/dragon_model \
  --epochs 100

Evaluation

python GCN_test.py \
  --gpu 0 \
  --model_path ./models/dragon_model.pt \
  --score_file_dir /path/to/scores/ \
  --triple_file_dir /path/to/triples/

---

Data Format

Input: Scores (JSON)

[{"q_id": {"candidate": [s1, s2, ..., s12, n1_s1, ...], ...}}]

Input: Triples (JSON)

{"q_id": {"p": "positive_entity", "n": ["neg1", "neg2", ...]}}

Output: Predictions (CSV)

question_id,candidate,score
q_001,entity_name,0.876

---

Citation

@inproceedings{zamiri2026dragon,
  author = {Zamiri, Mona and Kotov, Alexander},
  title = {Conversational Entity Retrieval from a Knowledge Graph using Aggregation of Fine-grained Relevance Signals with Graph Convolutions and Self-Attention},
  booktitle = {WSDM 2026},
  year = {2026}
}