Commonsense Knowledge Augmentation for Low-Resource Languages via Adversarial Learning
Authors: Bosung Kim, Juae Kim, Youngjoong Ko, Jungyun Seo6393-6401
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We designed experiments to demonstrate that high-scoring triples obtained by the proposed model can be considered augmented knowledge. The experimental results show that our proposed method for a low-resource language, Korean, achieved 93.7% precision@1 on a manually labeled benchmark. Furthermore, to verify our model for other low-resource languages, we introduced new test sets for knowledge validation in 16 different languages. |
| Researcher Affiliation | Collaboration | Bosung Kim1 , Juae Kim2, 3 , Youngjoong Ko1, Jungyun Seo2 1 Department of Computer Science and Engineering, Sungkyunkwan University 2 Department of Computer Science and Engineering, Sogang University 3 AIRS Company, Hyundai Motor Group |
| Pseudocode | No | The paper describes the proposed method in text and with diagrams (Figure 1 and Figure 2), but does not provide any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | We will release the augmented Korean knowledge and test sets for 16 languages. The augmented knowledge in Korean, comprising 626,681 triples with a confidence of 93.7%, will be publicly released for further research. In addition, we will open automatically generated test sets in 16 languages. |
| Open Datasets | Yes | Labeled English Data: There are over 3.4M English triples in Concept Net. |
| Dataset Splits | Yes | Labeled English Data: There are over 3.4M English triples in Concept Net. We extracted 18,690 highly weighted triples, and divided them into 17,690, 500, and 500 for the training, development, and test sets, respectively. In order to train the triple classifier, which predicts whether the triples are valid or not, negative triples are needed. We generated negative triples by replacing the head or tail entity of the valid triple with the random entities. As a result, 106,140, 3,000, and 3,000 English triples were used for the training, development, and test sets, respectively. Labeled Korean Data: To evaluate the proposed method s performance on Korean, we manually constructed development and test sets. |
| Hardware Specification | No | The paper mentions training models and setting hyperparameters, but does not provide any specific details about the hardware used for these experiments (e.g., GPU models, CPU types, or memory). |
| Software Dependencies | No | The paper mentions using 'Adam W optimizer' and 'Bing and Google Translate translator APIs', and model architectures like 'Bi-LSTM', 'CNN', and 'M-BERT'. However, it does not provide specific version numbers for any software libraries, frameworks, or APIs used for the implementation. |
| Experiment Setup | Yes | When we trained the triple classifier (C), the negative triples are generated by replacing the true triple s head or tail entity with other random entities. Experimental results show that models perform best when the number of negative samples is five times the number of positive samples. Thus, we made five invalid triples according to each positive triple. In the experiments for the LSTM and CNN feature extractor models, we used the Adam W optimizer (Loshchilov and Hutter 2019) and set the batch size to 128, learning rate to 1e-4, and clip to (-0.01, 0.01) for language discriminator D. The size of the hidden layers of Bi-LSTM is 900, and the kernel sizes of CNN are 3, 4, and 5 with 400 feature maps. In training M-BERT, we set the batch size to 64, learning rate to 2e-5, and clipped gradient norm as 1.0 to avoid exploding gradients. λ in Equation 3 was chosen to be between {0.01, 0.1, 0.2, 0.4}, and k was chosen to be between {0, 3, 5, 10}. All hyper-parameters were tuned based on the development. |