Subgraph KGQA

LLM Initial Training & Evaluation

The first part of the repository deals with training a pretrained LLM to obtain our n answer candidates for each question of Mintaka.

To train an LLM from scratch, run the following command:

python3 tmp.py

Alternatively, the beam search outputs from our trained LLM can be found on HuggingFace

Parsing Wikidata Clone

With our n answer candidates, our aim is to extract a subgraph for each <question, answer candidates> pair. In order to do so, we need to parse Wikidata into a local clone using iGraph.

To parse the Wikidata Dump, first download a version of choice via Wikidata. With the preffered downloaded dump, run the following command:

python3 tmp2.py

Subgraph Extraction with Wikidata Clone

With our iGraph Wikidata clone, we will construct our subgraphs with our n answer candidates and m question entities (golden entities provided by Mintaka).

To extract subgraphs, run the following command:

python3 tmp3.py

Subgraphs Reranking

With our subgraphs for each <question, answer candidate> pair, we want to rerank these subgraphs to boost our original Hits@1 obtained by the original LLM training (first part). We will rerank using the following approaches:

• Deterministic Sequences
• Graphormer

To train the Sequences Ranker from scratch, run the following command:

python3 tmp4.py

To train the Graphormer Ranker from scratch, run the following command:

python3 tmp5.py

To evaluate and rerank the trained ranker model, run the rerank.ipynb notebooks within sequences and graphormer folder respectively.

Alternatively, the trained weights for both approaches can be found on HuggingFace.