TempoQR: Temporal Question Reasoning over Knowledge Graphs

Authors: Costas Mavromatis, Prasanna Lakkur Subramanyam, Vassilis N. Ioannidis, Adesoji Adeshina, Phillip R Howard, Tetiana Grinberg, Nagib Hakim, George Karypis5825-5833

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that Tempo QR improves accuracy by 25 45 percentage points on complex temporal questions over state-of-the-art approaches and it generalizes better to unseen question types.
Researcher Affiliation Collaboration 1University of Minnesota 2University of Massachusetts Amherst 3Amazon Web Services 4Intel Labs
Pseudocode No The paper describes its methods using prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes For reproducibility, our code is available at: https://github.com/cmavro/Tempo QR.
Open Datasets Yes Cron Questions (Saxena, Chakrabarti, and Talukdar 2021) is a temporal QA benchmark based on the Wikidata TKG proposed in (Lacroix, Obozinski, and Usunier 2020).
Dataset Splits Yes Cron Questions consists of 410k unique question-answer pairs, 350k of which are for training and 30k for validation and for testing.
Hardware Specification No The paper does not provide specific hardware details such as CPU/GPU models or memory specifications used for running the experiments.
Software Dependencies No The paper mentions software like PyTorch, TComplEx, BERT, and RoBERTa, but does not specify their version numbers, which is required for reproducible software dependencies.
Experiment Setup Yes We learn TKG embeddings with the TCompl Ex method, where we set their dimensions D = 512. During, QA the pre-trained LM s parameters and the TKG embeddings are not updated. We set the number of transformer layers of the encoder f( ) to l = 6 with 8 heads per layer. We also observed the same performance when setting l = 3 with 4 heads per layer. The model s parameters are updated with Adam (Kingma and Ba 2014) with a learning rate of 0.0002. The model is trained for 20 maximum epochs and the final parameters are determined based on the best validation performance.