Task and benchmark

Studied conditional generation of natural-language descriptions from DBpedia knowledge graphs using WebNLG. The project adapted decoder-only GPT-2 and compared text-linearized graph inputs with graph-embedding approaches. Evaluation distinguished seen DBpedia categories from categories absent during training.

My contributions

The report credits me with implementing the GPT-2 task from scratch using recent packages, implementing the first graph-embedding method, preprocessing input files, and co-writing the report and presentation. Samidha Verma implemented the adaptations to the supplied repository, the second embedding method, and the global graph construction; I describe those as team experiments.

GPT-2 adaptation and fine-tuning

The project introduced head/relation/tail markers, relation-label vocabulary entries, left padding, and an instruction to convert triples into English. Training used Adam at 3e-5 with linear learning-rate decay. Batch sizes of 2, 4, and 8 and beam widths of 1, 3, and 5 were selected by development BLEU. Effective changes from the custom implementation were integrated into the original codebase, with output trimming and redundant-fragment removal.

Graph encoder experiment I implemented

Converted linearized triples into NetworkX graphs and PyTorch Geometric objects, encoded nodes with a graph convolutional network, and mean-pooled node embeddings to condition GPT-2 alongside text. This experiment did not yield satisfactory generation, motivating further architecture and training work.

Broader team investigation

A second method built a heterogeneous global graph with entity, relation, category, and global nodes; GPT-2 embeddings initialized features. A graph transformer was trained with positive edges and negative samples. An MLP mapped concatenated triple embeddings toward GPT-2 input space with the GNN frozen. The report states comprehensive testing was incomplete and shows repetitive or nonsensical outputs; it does not demonstrate an improvement from graph embeddings.

Results and generalization

The final text-input pipeline reports BLEU 23.00 overall, 30.87 on seen categories, and 12.91 on unseen categories. Overall ROUGE-1/2/L F1 scores are 0.4268/0.2171/0.3227. Published BART/T5 baseline scores reproduced in the report are substantially higher; they are reference results, not models we trained. The seen/unseen gap and failed embedding runs expose limits in generalization and structural conditioning.