Papers › JointGT: Graph-Text Joint Representation Learning for Text Generation from Knowledge Graphs

JointGT: Graph-Text Joint Representation Learning for Text Generation from Knowledge Graphs

19 Jun 2021Findings (ACL) 2021 8arXiv:2106.10502archive 2025-07-28

Pei Ke, Haozhe Ji, Yu Ran, Xin Cui, LiWei Wang, Linfeng Song, Xiaoyan Zhu, Minlie Huang

Existing pre-trained models for knowledge-graph-to-text (KG-to-text) generation simply fine-tune text-to-text pre-trained models such as BART or T5 on KG-to-text datasets, which largely ignore the graph structure during encoding and lack elaborate pre-training tasks to explicitly model graph-text alignments. To tackle these problems, we propose a graph-text joint representation learning model called JointGT. During encoding, we devise a structure-aware semantic aggregation module which is plugged into each Transformer layer to preserve the graph structure. Furthermore, we propose three new pre-training tasks to explicitly enhance the graph-text alignment including respective text / graph reconstruction, and graph-text alignment in the embedding space via Optimal Transport. Experiments show that JointGT obtains new state-of-the-art performance on various KG-to-text datasets.

PaperPDFConference PDFCode

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

thu-coai/JointGT officialmentioned in paperpytorch report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Graph ReconstructionKG-to-Text GenerationKnowledge GraphsQuestion GenerationRepresentation LearningText Generation

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
KG-to-Text Generation PathQuestion JointGT (BART) BLEU 65.89 #1 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion JointGT (BART) METEOR 48.25 #1 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion JointGT (BART) ROUGE 78.87 #1 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion BART BLEU 63.74 #2 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion BART METEOR 47.23 #2 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion BART ROUGE 77.76 #2 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion JointGT (T5) BLEU 60.45 #4 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion JointGT (T5) METEOR 45.38 #4 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion JointGT (T5) ROUGE 77.59 #4 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion T5 BLEU 58.95 #5 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion T5 METEOR 44.72 #5 of 5 Archive leaderboard report
KG-to-Text Generation PathQuestion T5 ROUGE 76.58 #5 of 5 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) JointGT (T5) BLEU 61.01 #5 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) JointGT (T5) METEOR 46.32 #5 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) JointGT (T5) ROUGE 73.57 #5 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) T5 BLEU 58.66 #6 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) T5 METEOR 46.04 #6 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) T5 ROUGE 73.06 #6 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) JointGT (BART) BLEU 58.55 #7 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) JointGT (BART) METEOR 45.01 #7 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) JointGT (BART) ROUGE 72.31 #7 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) BART BLEU 56.65 #8 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) BART METEOR 44.51 #8 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Constrained) BART ROUGE 70.94 #8 of 9 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) JointGT (T5) BLEU 66.14 #2 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) JointGT (T5) METEOR 47.25 #2 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) JointGT (T5) ROUGE 75.91 #2 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) JointGT (BART) BLEU 65.92 #3 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) JointGT (BART) METEOR 47.15 #3 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) JointGT (BART) ROUGE 76.10 #3 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) BART BLEU 64.55 #6 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) BART METEOR 46.51 #6 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) BART ROUGE 75.13 #6 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) T5 BLEU 64.42 #7 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) T5 METEOR 46.58 #7 of 13 Archive leaderboard report
KG-to-Text Generation WebNLG 2.0 (Unconstrained) T5 ROUGE 74.77 #7 of 13 Archive leaderboard report
KG-to-Text Generation WebQuestions JointGT (BART) BLEU 30.02 #1 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions JointGT (BART) METEOR 32.05 #1 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions JointGT (BART) ROUGE 55.6 #1 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions BART BLEU 29.61 #2 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions BART METEOR 31.48 #2 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions BART ROUGE 55.42 #2 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions JointGT (T5) BLEU 28.95 #4 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions JointGT (T5) METEOR 31.29 #4 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions JointGT (T5) ROUGE 54.47 #4 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions T5 BLEU 28.78 #5 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions T5 METEOR 30.55 #5 of 5 Archive leaderboard report
KG-to-Text Generation WebQuestions T5 ROUGE 55.12 #5 of 5 Archive leaderboard report
Question Generation GrailQA-Compositional JointGT BLEU 31.46 #2 of 4 Archive leaderboard report
Question Generation GrailQA-Compositional JointGT FactSpotter 95.26 #2 of 4 Archive leaderboard report
Question Generation GrailQA-Compositional JointGT METEOR 36.08 #2 of 4 Archive leaderboard report
Question Generation GrailQA-IID JointGT BLEU 45.95 #2 of 4 Archive leaderboard report
Question Generation GrailQA-IID JointGT FactSpotter 98.62 #2 of 4 Archive leaderboard report
Question Generation GrailQA-IID JointGT METEOR 41.65 #2 of 4 Archive leaderboard report
Question Generation GrailQA-Zero-Shot JointGT FactSpotter 94.15 #1 of 4 Archive leaderboard report
Question Generation GrailQA-Zero-Shot JointGT METEOR 37.69 #1 of 4 Archive leaderboard report
Question Generation GrailQA-Zero-Shot JointGT bleu 32.94 #1 of 4 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdafactorAdamAttentionAttention DropoutBARTBPEDense ConnectionsDropoutGated Linear UnitInverse Square Root ScheduleLabel SmoothingLayer NormalizationLinear LayerMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSentencePieceSoftmaxT5Transformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections