TransOMCS: From Linguistic Graphs to Commonsense Knowledge

Authors: Hongming Zhang, Daniel Khashabi, Yangqiu Song, Dan Roth

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate the transferability of linguistic knowledge to commonsense knowledge and the effectiveness of the proposed approach in terms of quantity, novelty, and quality.
Researcher Affiliation Collaboration 1The Hong Kong University of Science and Technology 2Allen Institute for AI 3University of Pennsylvania
Pseudocode No The paper describes algorithms (e.g., Breadth-First Search for pattern extraction) but does not provide structured pseudocode blocks or figures labeled as 'Algorithm' or 'Pseudocode'.
Open Source Code Yes Trans OMCS is publicly available.1 1https://github.com/HKUST-KnowComp/Trans OMCS
Open Datasets Yes For the seed commonsense knowledge, we use the English subset of Concept Net 5.5 [Speer et al., 2017]. For the linguistic knowledge resource, we use the core subset of ASER [Zhang et al., 2020] with 37.9 million linguistic graphs3 to form G.
Dataset Splits No The paper mentions a train/test split for the annotated data ('We randomly select 10% of the dataset as the test set and the rest as the training set.'). It does not explicitly mention a separate validation split for either the classifier training or the main model evaluations.
Hardware Specification No The paper does not provide specific details about the hardware used, such as GPU models, CPU specifications, or memory.
Software Dependencies No The paper mentions using 'BERT [Devlin et al., 2019] and their pre-trained parameters (BERT-base) as the initialization' and general methods like 'cross-entropy as the loss function and stochastic gradient descent as the optimization method'. However, it does not provide specific version numbers for any software libraries, frameworks, or tools.
Experiment Setup Yes We randomly select 10% of the dataset as the test set and the rest as the training set.5 ... For the fair comparison, we keep all the model architecture and parameters the same across different trials. ... All the numbers are computed based on the average of four different random seeds rather than the best seed as reported in the original paper.